Best LLMs for coding in 2026

Last update: this post was originally created in March 2026 and was last updated on 4.27.26 to include recently released models such as GPT-5.5, Opus 4.7, Kimi K2.6 and DeepSeek V4 Pro.

TL;DR

The best LLM for coding in 2026 depends on your workload:

• Raw benchmark leader: GPT-5.5 (xhigh) tops the AA Coding Index at 59.1 and AA Intelligence Index at 60.2; Claude Opus 4.7 still leads SWE-Bench Verified at 87.6%.
• Best open-source coder at scale: Kimi K2.6 on Fireworks: 47.1 on the AA Coding Index, ~85 tok/s, $0.95/$4 per 1M, Modified MIT.
  • DeepSeek V4-Pro narrowly tops the composite (47.5) and leads Terminal-Bench Hard at $1.74/$3.48 per 1M, MIT, and is now live on Fireworks. AA per-host throughput on Fireworks for V4-Pro is not yet published; V4-Pro serves at ~38 tok/s on DeepSeek’s direct API per AA’s measurement.
• Lowest-cost open-weight option: gpt-oss-120B on Fireworks: $0.15/$0.60 per 1M, 131K context, Apache 2.0. Fireworks throughput on this model is 70 tok/s per AA’s standardized workload.
• Lowest per-token price with frontier context: DeepSeek V4-Flash direct from DeepSeek at $0.14/$0.28 per 1M, 1M context, 38.7 on the AA Coding Index.
• Longest context with open weights: DeepSeek V4-Pro and V4-Flash (1M, MIT): the only MIT-licensed open-weight models in this guide at 1M.
  • Qwen3.6 Plus matches on 1M but is closed-weight API-only.
• Most permissive license for commercial fine-tuning: GLM-5.1 or DeepSeek V4-Pro: both MIT, both LoRA- and RFT-ready on Fireworks. V4-Pro is now live on the Fireworks catalog; V4-Flash will follow.

The short answer to “which LLM is best for coding” depends on where you are in your product lifecycle. For experimentation and PMF, closed frontier models lead the benchmarks and handle edge cases better out-of-the-box. Once you’ve validated the workload, the math changes: open models on an inference platform can cost 6x to ~100x less per output token depending on which open and closed models you compare, and post-training on your own data closes the quality gap for your specific eval.

The table below compares all eleven models on the metrics that drive a production coding decision: parameter counts, context window, price, Artificial Analysis Coding Index, and license. Pricing shown for models hosted on Fireworks reflects Fireworks Serverless rates; pricing for closed models reflects each vendor’s public API; pricing for DeepSeek V4-Pro on Fireworks Serverless matches DeepSeek’s direct API rate; V4-Flash is not yet on Fireworks, so its pricing reflects DeepSeek’s direct API.

• –Benchmarks are from Artificial Analysis. Price sources are labeled inline: “(Fireworks)” reflects Fireworks Serverless rates as of 2026-04-27, verified against fireworks.ai/models. V4-Pro on Fireworks Serverless prices at $1.74 input / $0.14 cached input / $3.48 output per 1M tokens, matching DeepSeek’s direct API rate. “(DeepSeek API)” reflects DeepSeek’s own API at cache-miss input rates (cache-hit rates are $0.028/M for V4-Flash and $0.145/M for V4-Pro on DeepSeek’s API). Closed-source model rows show each vendor’s public API pricing. DeepSeek V4-Pro edges Kimi K2.6 by 0.4 points on the AA Coding Index (47.5 vs 47.1), making V4-Pro the top-scoring open-source entry in this guide by that composite; see the per-model section for the full per-benchmark breakdown between the two.

Three benchmarks from Artificial Analysis discriminate coding capability meaningfully across all eleven models:

• Artificial Analysis Coding Index (a composite of Terminal-Bench Hard and SciCode)
• Terminal-Bench Hard (end-to-end shell-based agentic tasks)
• SciCode (scientific code generation)

The Artificial Analysis Intelligence Index is included as a general-purpose baseline. SWE-Bench Verified scores are the industry-standard coding number and are shown in a separate column below; these are vendor-reported or community-reported, not from Artificial Analysis.

• –Claude Opus 4.7 (Adaptive Reasoning, Max Effort) is the current frontier reference for the bolding baseline; scores that beat Opus 4.7 on a given benchmark are bolded. GPT-5.5 (xhigh) is the new AA Coding Index and AA Intelligence Index leader as of April 23, 2026. SWE-Bench scores are vendor-reported unless otherwise noted; GPT-5.5 and GPT-5.4 have not published SWE-Bench Verified scores, so SWE-Bench Pro is substituted with the variant named inline. DeepSeek V4-Pro’s AA composites (AA Coding Index 47.5, Terminal-Bench Hard 0.462, AA Intelligence Index 52) became available on Artificial Analysis on April 24, 2026 and are included here; the vendor-reported SWE-Bench Verified, LiveCodeBench, and Terminal-Bench 2.0 scores appear in the V4-Pro section below. gpt-oss-120B’s ~30% is community-reported. Bolded scores indicate the model beats Claude Opus 4.7 on that benchmark. Actual performance varies based on prompt engineering, quantization, and inference settings.

Three takeaways from the table:

• GPT-5.5 resets the frontier on coding. It leads the AA Coding Index by 1.8 points over GPT-5.4 (59.1 vs 57.3), by 6.6 points over Opus 4.7 (59.1 vs 52.5), and is the first model in this list to clear Opus 4.7 on all four AA columns. Opus 4.7 still holds SWE-Bench Verified at 87.6%; OpenAI reports SWE-Bench Pro only for GPT-5.5 (58.6%).
• The top open-source spot on the AA Coding Index is now a near-tie between DeepSeek V4-Pro (47.5) and Kimi K2.6 (47.1), both live on Fireworks Serverless. V4-Pro edges the composite and leads on Terminal-Bench Hard (0.462 vs 0.439); Kimi K2.6 leads on AA Intelligence (53.9 vs 52) and SciCode (0.535 vs 0.500). Kimi K2.6 is roughly ~2x faster on AA throughput (~85 tok/s on Fireworks vs V4-Pro’s ~38 tok/s on DeepSeek’s direct API; AA per-host throughput on Fireworks for V4-Pro is not yet published) and roughly 20% cheaper on a blended 3:1 basis ($1.71/M vs $2.18/M). Both remain 5-6 points behind Opus 4.7 on the Coding Index composite (52.5), with lower cost the open-source trade.
• gpt-oss-120B scores lowest on the composite (28.6) because it lags on agentic Terminal-Bench Hard (0.235), but its LiveCodeBench score (0.878) remains competitive. DeepSeek V4-Flash now sits above it on every AA benchmark at near-identical input pricing ($0.14 vs $0.15 per 1M) and roughly half the output price ($0.28 vs $0.60), and on AA’s standardized workload V4-Flash on DeepSeek’s API (~84 tok/s) also serves slightly faster than gpt-oss-120B on Fireworks (70 tok/s). gpt-oss-120B’s edges today are Fireworks hosting (V4-Flash isn’t on Fireworks yet), the Apache 2.0 license, and FireOptimizer-ready fine-tuning.

• Parameters: Undisclosed
• Context window: 1M tokens (API); 400K in Codex
• Input $/1M: $5.00
• Output $/1M: $30.00
• License: Proprietary
• Release date: April 23, 2026 (ChatGPT and Codex); API live April 24, 2026
• SWE-Bench Pro: 58.6% (OpenAI has not published a SWE-Bench Verified score)
• AA Coding Index: 59.1 (highest in this list)
• Terminal-Bench Hard: 0.606
• SciCode: 0.561
• Tool calling: Strong (OpenAI native; Codex-integrated for long-horizon coding work)
• Prompt caching: Yes, 90% read discount (unchanged from GPT-5.4)

GPT-5.5 is OpenAI’s frontier reasoning model released April 23, 2026. At xhigh reasoning effort it leads the AA Coding Index (59.1), AA Intelligence Index (60.2), and Terminal-Bench Hard (0.606), and is the first model in this guide to clear Opus 4.7 on every AA-tracked column. OpenAI’s own announcement frames the step up as “state-of-the-art intelligence at half the cost of competitive frontier coding models” and emphasizes matching GPT-5.4’s per-token latency while solving harder tasks with fewer tokens.

Two deployment notes at launch. First, GPT-5.5 shipped through ChatGPT and Codex on April 23, 2026, followed by the Responses and Chat Completions APIs one day later on April 24, 2026 at $5/$30 per 1M input/output tokens. Second, a separate tier called GPT-5.5 Pro is available in ChatGPT at $30/M input and $180/M output. Pro gains most on FrontierMath Tier 4 (+4.2 pts over base) and BrowseComp (+5.7 pts), and is reserved for high-stakes analytical work where the price delta is justified.

See the Fireworks model catalog for cost-comparable open-source alternatives.

Teams shipping agentic coding tools where the binding constraint is long-horizon reasoning quality and where the workload can absorb the $5/$30 pricing. GPT-5.5 is the natural fit for Codex-style IDE agents, autonomous coding bots that hold context across large systems, and long-running refactor or migration work. It is also the right call when your eval depends on frontier performance on Terminal-Bench-class agentic tasks, where the 0.091-point gap over Opus 4.7 on Terminal-Bench Hard (0.606 vs 0.515) is the largest lead on that benchmark in the AA leaderboard.

• The API rollout was staged. Codex and ChatGPT got GPT-5.5 on April 23, 2026; the Responses and Chat Completions APIs went live one day later on April 24, 2026. Teams that integrated against ChatGPT or Codex first can migrate to the API endpoints now that they are live. Source: OpenAI GPT-5.5 release notes.
• Pricing doubles relative to GPT-5.4. Input moves from $2.50 to $5.00 per 1M and output moves from $15.00 to $30.00 per 1M, and the model uses fewer tokens per task in Codex but not uniformly across all workloads. Re-run your own cost-per-successful-task math before switching a high-volume pipeline. Batch and Flex pricing are available at half the standard API rate; Priority processing runs at 2.5x.
• GPT-5.5 Pro (the $30/$180 tier) is a separate product, not an effort level. Pro’s advantages are concentrated on FrontierMath Tier 4, BrowseComp, and Investment Banking Modeling, with parity or near-parity on coding benchmarks. Do not conflate “GPT-5.5 xhigh” with “GPT-5.5 Pro”; the latter has not yet landed in the API.
• OpenAI has tightened cyber safeguards on GPT-5.5 and warns that “some users may find them annoying initially” before tuning settles. Security-tool developers, red-team tooling, and CTF-adjacent workflows should expect more refusals at launch and apply for Trusted Access for Cyber if legitimate access is denied. Source: OpenAI GPT-5.5 release notes.
• Reasoning tokens continue to bill as output tokens and occupy context without appearing in the response, consistent with GPT-5.4. Budget headroom in max_tokens to avoid truncated long-horizon runs.

Q: How does GPT-5.5 compare to Claude Opus 4.7 on coding?

GPT-5.5 leads the AA Coding Index by 6.6 points (59.1 vs 52.5), Terminal-Bench Hard by 0.091 (0.606 vs 0.515), and AA Intelligence by 2.9 points (60.2 vs 57.3). Opus 4.7 retains SWE-Bench Verified at 87.6% (OpenAI reports only SWE-Bench Pro at 58.6% for GPT-5.5). On literal-instruction interpretation and 1M context recall under Anthropic’s tokenizer, Opus 4.7 still differs operationally from GPT-5.5; the choice is workload-dependent.

Q: Can GPT-5.5 run on Fireworks?

No. GPT-5.5 is closed-source and will only be available through OpenAI’s API once they complete the API rollout. For cost-comparable open-source alternatives on Fireworks, Kimi K2.6 sits at 47.1 on the AA Coding Index at $0.95/$4.00 per million tokens.

Q: Is GPT-5.5 available in the API?

Yes, as of April 24, 2026. GPT-5.5 launched in ChatGPT and Codex on April 23, 2026 and in the Responses and Chat Completions APIs one day later. API pricing is $5 per 1M input tokens and $30 per 1M output tokens.

• Parameters: Undisclosed
• Context window: 1M tokens (pricing doubles above 270K input tokens)
• Input $/1M: $2.50
• Output $/1M: $15.00
• License: Proprietary
• SWE-Bench Pro: 57.7% (OpenAI has not published a SWE-Bench Verified score)
• AA Coding Index: 57.3
• Terminal-Bench Hard: 0.576
• SciCode: 0.566
• Tool calling: Strong (OpenAI native; see aggressive parallel-batch caveat below)
• Prompt caching: Yes, 90% read discount

GPT-5.4 is OpenAI’s previous-generation frontier reasoning model, released March 5, 2026 and superseded by GPT-5.5 on April 23, 2026. It remains in this guide because the 2x pricing step to GPT-5.5 ($5/$30 vs $2.50/$15 per 1M) is not justified for every workload, and because the API for GPT-5.5 was still rolling out at time of publication. At the xhigh reasoning effort level, GPT-5.4 scores 57.3 on the AA Coding Index and ships native computer use, letting the model drive a browser or OS-level agent loop directly. On SWE-Bench Verified, Claude Opus 4.7 remains ahead at 87.6%.

The catch is latency and cost. Median time to first token at xhigh is approximately 205 seconds per Artificial Analysis, so interactive use requires a lower effort level. Reasoning tokens bill as output tokens and consume context window space without appearing in the response, so OpenAI’s own guidance recommends 25,000 tokens of headroom to avoid truncation on long agentic runs.

See the Fireworks model catalog for cost-comparable open-source alternatives.

Teams shipping agentic coding tools where quality is the binding constraint and where computer-use primitives matter. v0, Cursor-style IDE agents, and long-horizon code-refactor bots that benefit from multi-step reasoning are the natural fit. It is also the right call when your eval depends on strict benchmark-leader performance and the workload volume is low enough to absorb $15/M output pricing.

• At xhigh reasoning, GPT-5.4 spawns aggressive parallel tool-call batches and sometimes runs dependent shell commands concurrently (for example, git add alongside git commit), which breaks shell scripts that assume sequential completion. Source: openai/codex#14485 (open).
• Above 270K input tokens, input pricing doubles to $5.00 and output rises 50% to $22.50. OpenAI’s own evals indicate perfect 1M recall is not reliable. OpenAI’s guidance recommends keeping critical context near the front of the window. Source: OpenAI latest-model guide.
• Codex users report burning 20% of a token allowance in about two hours of routine prompting at High reasoning effort, without rate-limit adjustments to match. Source: openai/codex#14593 (open).
• Reasoning tokens are invisible to the response but billed as output tokens and occupy context. Budget at least 25,000 tokens of headroom to avoid truncated responses.

Q: Why choose GPT-5.4 over GPT-5.5?

Cost and API availability. GPT-5.4 costs half as much on both input and output ($2.50/$15 vs $5/$30 per 1M) and is already available in the Responses and Chat Completions APIs. GPT-5.5 surpasses GPT-5.4 on every AA-tracked coding benchmark, but high-volume workloads that are cost-bound or that depend on API access today can justify staying on GPT-5.4 until the pricing math or the GPT-5.5 API catches up.

Q: How does GPT-5.4 compare to Claude Opus 4.7 on coding?

GPT-5.4 leads the AA Coding Index by 4.8 points (57.3 vs 52.5) and Terminal-Bench Hard by 0.061. Opus 4.7 leads on SWE-Bench Verified (87.6% vs 57.7% SWE-Bench Pro for GPT-5.4) and edges the AA Intelligence Index by 0.5 points (57.3 vs 56.8).

Q: Can GPT-5.4 run on Fireworks?

No. GPT-5.4 is closed-source and only available through OpenAI’s API. For cost-comparable open-source alternatives on Fireworks, Kimi K2.6 sits at 47.1 on the coding index at $0.95/$4.00 per million tokens.

Q: Does computer use require a different API?

Yes. Computer use is a separate tool-type in the Responses API and requires specific harness scaffolding outside standard chat completions.

• Parameters: Undisclosed
• Context window: 1M tokens (65K max output)
• Input $/1M: $2.00
• Output $/1M: $12.00
  • Note: Pricing doubles to $4.00 / $18.00 above 200K input tokens
• License: Proprietary (Preview tier)
• Training cutoff: January 2025 (official)
• SWE-Bench Verified: 80.6% (vendor-reported)
• AA Coding Index: 55.5
• Terminal-Bench Hard: 0.54
• SciCode: 0.589
• Tool calling: Workable (requires echoing thought_signature across turns; see tradeoffs)
• Prompt caching: Yes, 90% read discount; hourly storage fee applies

Gemini 3.1 Pro Preview is Google’s flagship coding and reasoning model released February 19, 2026. It scores the highest SciCode in the list (0.589) and ties for top general intelligence (57.2 on AA Intelligence Index, within 0.1 of Opus 4.7). Google reports 77.1% on ARC-AGI-2, a discriminative score on a benchmark designed to be contamination-resistant.

Preview status matters: pay-as-you-go traffic runs on Vertex AI’s Dynamic Shared Quota, which shares a single global capacity pool across all users. Individual requests can hit 429 rate-limit errors based on worldwide load rather than your own usage, and Provisioned Throughput reservations are GA-only. Production workloads should wait for GA before deployment.

See the Fireworks model catalog for open-source alternatives with comparable long-context capability.

Teams running long-context codebase analysis where 1M tokens of input beats chunking or RAG. SciCode is a natural domain fit given the 0.589 score. Teams already committed to Google Cloud and Vertex AI will get the cleanest integration path.

• Multi-turn function calls require echoing back an opaque thought_signature on every subsequent turn. Missing signatures return HTTP 400, breaking OpenAI-compatible wrappers that strip unknown fields. Source: Gemini thought signatures docs.
• Parallel function calls require grouping all functionCall parts before functionResponse parts. Interleaving FC1, FR1, FC2, FR2 returns HTTP 400. Only the first functionCall carries the signature. Source: same docs.
• Pay-as-you-go preview traffic runs on Vertex AI’s Dynamic Shared Quota: a single global capacity pool dynamically allocated by real-time demand with no predefined per-project limit, where Provisioned Throughput is not available. Production workloads should wait for GA and a Provisioned Throughput reservation. Source: Vertex AI throughput-quota docs.
• On Windows, gemini-3.1-pro-preview hits 100% error rate in gemini-cli v0.32.1: requests hang in the thinking state without streaming or error. P1 priority per maintainers. Source: google-gemini/gemini-cli#21937 (open).
• Third-party API routers strip the content array from assistant messages on turn 2+, returning “no assistant messages” and breaking multi-turn agent loops. Source: RooCodeInc/Roo-Code#11629 (open).

Q: How does Gemini 3.1 Pro compare to GPT-5.4 on coding?

GPT-5.4 leads on the composite Coding Index by 1.8 points (57.3 vs 55.5). Gemini leads on SciCode (0.589 vs 0.566) and on the AA Intelligence Index (57.2 vs 56.8).

Q: Is the 1M context window usable in practice?

Vendor docs don’t caveat recall as heavily as OpenAI does for GPT-5.4. Community benchmarks show Gemini 3 Pro’s long-context performance as among the strongest in the class, though 1M remains workload-dependent.

Q: Can Gemini 3.1 Pro run on Fireworks?

No. Gemini is proprietary and only available through Google Cloud and Vertex AI.

• Parameters: Undisclosed
• Context window: 1M tokens (128K max output)
• Input $/1M: $5.00
• Output $/1M: $25.00
• License: Proprietary
• SWE-Bench Verified: 87.6% (highest in this list)
• AA Coding Index: 52.5
• Terminal-Bench Hard: 0.515
• SciCode: 0.545
• Tool calling: Strong (Anthropic native; see literal-interpretation caveat)
• Prompt caching: Yes, 90% read discount (tied with GPT-5.4 for deepest in the list)

Claude Opus 4.7 is Anthropic’s flagship model released April 16, 2026. At max effort with adaptive reasoning, it ties Gemini 3.1 Pro on the AA Intelligence Index (57.3 vs 57.2) and leads every model in the list on context coherence when paired with the 1M window at standard pricing (no long-context premium). At $5/$25 per million tokens, it is also the most expensive model in this list.

Opus 4.7 represents the sharpest operational break from its predecessor of any model here. Anthropic’s own migration guide calls out literal instruction interpretation: prompts that previously relied on the model inferring unstated intent now underperform. The new tokenizer also produces 1.0-1.35x as many tokens for the same text, changing cost and context-budget math.

See the Fireworks model catalog for open alternatives. Kimi K2.6 is the closest match in coding capability at 6x lower output cost.

Teams whose existing harnesses are tuned to Anthropic’s API and who need the strongest available model for literal, plan-and-execute tasks with maximum reliability. Long-horizon agentic coding runs where the 1M context matters and where per-task cost is not the binding constraint.

• Opus 4.7 interprets instructions more literally than 4.6. Prompts that relied on inference or generalization need rewriting to be explicit. Source: Anthropic Opus 4.7 migration docs.
• The new tokenizer uses 1.0-1.35x as many tokens for the same input (up to 35% more). Update max_tokens headroom and compaction triggers. Source: same docs.
• Setting temperature, top_p, or top_k returns HTTP 400. All three sampling knobs are removed. Any pipeline that used temperature=0 for determinism must be rewritten. Source: same docs.
• Claude Code silently switches sessions to Opus 4.7 [1M] without UI indication, and because the 1M variant burns tokens at higher rates, one reported case shows ~4x burn rate. Source: anthropics/claude-code#49541 (open).
• Adaptive thinking is off by default. Requests without a thinking field skip thinking entirely. Thinking content is omitted from responses unless you opt in with display: summarized.

Q: How does Opus 4.7 compare to Opus 4.6?

On coding specifically, Opus 4.7 scores 52.5 on the AA Coding Index vs Opus 4.6’s 48.1 (+4.4 points). It also gains 0.053 on Terminal-Bench Hard (0.515 vs 0.462). Migration cost is nontrivial: tokenizer, sampling, and literal-interpretation changes combine.

Q: Can Claude Opus 4.7 run on Fireworks?

No. Claude is proprietary and only available through Anthropic and their cloud partners. Kimi K2.6 on Fireworks is the closest open-source approach to Opus 4.7’s coding capability at 6.25x lower output cost.

Q: Is the 1M context worth the token multiplier?

For long-horizon agentic coding with large codebases in-context, yes. For single-file review or chat-style use, the 200K variant is the more economical baseline.

• Parameters: Undisclosed
• Context window: 1M tokens (64K max output, at standard pricing, no premium)
• Input $/1M: $3.00
• Output $/1M: $15.00
• License: Proprietary
• SWE-Bench Verified: 79.6%
• AA Coding Index: 50.9
• Terminal-Bench Hard: 0.530
• SciCode: 0.468
• Tool calling: Strong (Anthropic native; scoped-instruction caveats in Claude Code)
• Prompt caching: Yes, 90% read discount

Claude Sonnet 4.6 at max effort sits between Opus 4.7 and the open-source leaders. At 50.9 on the Coding Index it closes most of the gap to Opus 4.7 (52.5), and it matches the 1M context window at standard pricing without the $25/M output cost. For teams already running Sonnet-class pipelines, 4.6 is the direct upgrade path.

Adaptive thinking is the main operational difference from Opus 4.7. Sonnet 4.6 supports adaptive thinking (recommended) or manual thinking: {type: enabled, budget_tokens: N} budgets, which are deprecated but still functional. Agent harnesses currently hard-coding task-specific budgets should plan the migration to adaptive thinking.

See the Fireworks model catalog for high-volume open alternatives at the Sonnet price point.

Teams running high-volume coding workloads where Opus-class quality is overkill and where the $3/$15 price point hits the cost ceiling. Well-suited to code review (with the hedging caveat below), bug triage, and refactoring of known-good code. The 1M context at standard pricing is a clear win over Sonnet 4.5.

• In Claude Code, Sonnet 4.6 can misinterpret scoped instructions: it searches the entire repo when told to work on specific files and treats “remove from docs” as “remove from script signatures.” Source: anthropics/claude-code#41707 (open).
• In code review, Sonnet-class models trend toward hedging language (“might”, “could”, “possibly”) and frame fixes as exploratory suggestions rather than assertive calls. CodeRabbit’s benchmark on the prior Sonnet 4.5 measured ~34% hedged comments with a documented gap vs Opus on concurrency bugs, memory leaks, and cross-module issues; comparable Sonnet 4.6 data has not been published. Source: CodeRabbit analysis (Sonnet 4.5).
• Some requests for standard Sonnet 4.6 are incorrectly classified as 1M-context requests, returning “Extra usage is required for long context requests” errors with no 1M affordance in the UI. Source: anthropics/claude-code#48274 (open).

Q: How does Sonnet 4.6 compare to Opus 4.7?

Opus 4.7 leads by 1.6 points on the Coding Index (52.5 vs 50.9) and 5.6 points on AA Intelligence (57.3 vs 51.7). Sonnet 4.6’s output price is 60% of Opus 4.7’s ($15 vs $25 per 1M, or 40% less), and it uses the prior tokenizer.

Q: Can Claude Sonnet 4.6 run on Fireworks?

No. Sonnet is proprietary.

Q: Is Sonnet 4.6 the right code-review model?

It is a strong general-purpose code reviewer. CodeRabbit’s Sonnet 4.5 benchmark measured ~34% hedged comments, and Sonnet-class harnesses should pair the model with guardrails that force assertive verdicts where appropriate until comparable 4.6 data is published.

• Parameters: 1.6T total, 49B active (MoE)
• Context window: 1M tokens (384K max output)
• Input $/1M: $1.74 (Fireworks Serverless, cache miss) / $0.14 (cache hit)
• Output $/1M: $3.48 (Fireworks Serverless)
• License: MIT (open weights)
• Release date: April 24, 2026
• SWE-Bench Verified: 80.6% (vendor-reported); 55.4 on SWE-Bench Pro
• AA Coding Index: 47.5 (top open-source score in this guide, edging Kimi K2.6’s 47.1 by 0.4 points)
• Terminal-Bench Hard: 0.462
• SciCode: 0.500
• Tool calling: Workable (dual-mode OpenAI + Anthropic APIs, Thinking / Non-Thinking)
• Prompt caching: Automatic on Fireworks Serverless ($0.14 cached input vs $1.74 cache miss); also automatic on DeepSeek’s direct API
• Fine-tunable on Fireworks: Yes (LoRA, SFT, RFT via FireOptimizer)

DeepSeek V4-Pro is DeepSeek’s April 24, 2026 frontier release, shipped as an “open-source SOTA in agentic coding benchmarks” per DeepSeek’s own positioning. The architecture introduces token-wise compression plus DeepSeek Sparse Attention (DSA), which is what enables the 1M-token context as the default across all official DeepSeek services without blowing up compute and memory costs. Both V4-Pro and V4-Flash expose OpenAI-compatible and Anthropic-compatible endpoints and support dual Thinking / Non-Thinking modes.

On the vendor tech report, V4-Pro leads the featured comparison group (Opus-4.6-Max, GPT-5.4-xHigh, Gemini-3.1-Pro-High) on LiveCodeBench (93.5) and Codeforces (3206), beats both Opus-4.6 and GPT-5.4 on SimpleQA-Verified (57.9 vs 46.2 and 45.3) but trails Gemini-3.1-Pro (75.6) on that metric, and matches Opus-4.6 on SWE-Bench Verified at 80.6% (within 0.2 pts). It trails GPT-5.4 on Terminal-Bench 2.0 (67.9% vs 75.1%) and IMOAnswerBench (89.8 vs 91.4). Worth flagging that DeepSeek’s comparison baseline is Opus-4.6, not Opus-4.7.

Artificial Analysis has now published the AA composite indexes for V4-Pro: AA Coding Index 47.5 (narrowly the top open-source score in this guide, +0.4 over Kimi K2.6 at 47.1), Terminal-Bench Hard 0.462 (above Kimi K2.6’s 0.439), AA Intelligence Index 52 (below Kimi K2.6’s 53.9), and SciCode 0.500 (below Kimi K2.6’s 0.535). The pattern: V4-Pro edges Kimi K2.6 on agentic coding composite and Terminal-Bench Hard; Kimi K2.6 edges V4-Pro on general intelligence and scientific-code generation. The other notable AA-measured fact is throughput: V4-Pro serves at ~38 tokens/sec on DeepSeek’s direct API per AA’s measurement, and AA’s model page calls out the V4-Pro Max Effort variant as “notably slow” relative to its open-weight peer group. AA per-host throughput on Fireworks for V4-Pro is not yet published; teams should run their own benchmarks against the Fireworks endpoint until AA reports it.

🚀 V4-Pro is now live on Fireworks Serverless. View the DeepSeek V4-Pro API on Fireworks.

Teams building agentic coding pipelines that want the strongest open-weight AA Coding Index score (47.5), a MIT license, and 1M context, and that can tolerate slower throughput than V4-Flash or gpt-oss-120B. V4-Pro is also the natural fit for teams already running DeepSeek V3.2 or V3.1 in production: the migration is an endpoint and model-parameter change, not a harness rewrite. For teams that want to fine-tune against their own repo, the open weights remove license friction entirely.

• Output speed is slow for the price tier. AA measures V4-Pro Max Effort at ~38 tokens/sec on DeepSeek’s direct API: roughly half V4-Flash’s ~84 tok/s and roughly half gpt-oss-120B’s ~70 tok/s on Fireworks. AA per-host throughput on Fireworks for V4-Pro is not yet published. Long agentic runs that loop through many tool calls feel the difference; budget for longer wall-clock times. Source: Artificial Analysis model page for V4-Pro.
• V4-Pro’s AA Intelligence Index (52) and SciCode (0.500) both trail Kimi K2.6 (53.9 and 0.535), even though V4-Pro edges Kimi on the AA Coding Index composite (47.5 vs 47.1). Workloads that depend on scientific-code generation or broad general intelligence should prefer Kimi K2.6; V4-Pro is the better fit where agentic coding composite and Terminal-Bench Hard dominate the eval. Source: Artificial Analysis model page for V4-Pro.
• V4-Pro’s SimpleQA-Verified score is 57.9%, well behind Gemini-3.1-Pro’s 75.6% in the same vendor comparison. Factual-recall-heavy coding workflows (CLI-arg lookup, API method recall, version-specific syntax) should expect more retrieval gaps than the flagship closed models. Source: DeepSeek V4 tech report.
• The legacy deepseek-chat and deepseek-reasoner endpoints on DeepSeek’s direct API retire on July 24, 2026 and currently route to deepseek-v4-flash. Production traffic on those legacy endpoints should migrate explicitly to deepseek-v4-pro (or to V4-Pro on Fireworks Serverless) before that date. Source: DeepSeek V4 release notes.
• The vendor’s “open-source SOTA in agentic coding” framing rests on internal evaluations and DSA-specific optimizations. Community reproductions are still early; independent agentic-coding benchmarks should be run on your own workload before committing. Source: DeepSeek V4 release notes.
• V4’s token-wise compression layered on top of DSA is new. Custom serving stacks that worked for V3.2’s DSA implementation will need updated parsers and KV-cache handling to run V4-Pro efficiently. The vendor reports V4-Pro requires only 10% of DeepSeek-V3.2’s KV cache at 1M-token context, a ~90% reduction, which depends on the compression path being wired correctly.

Q: Is V4-Pro on Fireworks?

Yes. V4-Pro is now live on Fireworks Serverless at $1.74/$3.48 per 1M tokens (matching DeepSeek’s direct API rate), with automatic prompt caching at $0.14 cached input. View the DeepSeek V4-Pro API on Fireworks.

Q: How does V4-Pro compare to GPT-5.5 on coding?

GPT-5.5 leads on every AA-tracked composite: AA Coding Index 59.1 vs 47.5 (+11.6 points), Terminal-Bench Hard 0.606 vs 0.462 (+0.144), AA Intelligence 60.2 vs 52 (+8.2 points), SciCode 0.561 vs 0.500 (+0.061). On vendor-reported numbers, V4-Pro leads LiveCodeBench (93.5) and has a higher Codeforces rating (3206 vs 3168 for GPT-5.4; GPT-5.5 not yet reported), and trails on Terminal-Bench 2.0 (67.9% vs 82.7% for GPT-5.5). V4-Pro’s killer argument is cost: $1.74/$3.48 per 1M is roughly a third of GPT-5.5’s input price and about one-ninth of its output price, with open weights.

Q: Can V4-Pro drive Claude Code or OpenClaw directly?

Yes. DeepSeek explicitly calls out Claude Code, OpenClaw, and OpenCode integration in the release notes, and the API exposes an Anthropic-compatible base URL at https://api.deepseek.com/anthropic.

• Parameters: 1T total, 32B active (MoE)
• Context window: 256K tokens
• Input $/1M: $0.95 (Fireworks)
• Output $/1M: $4.00 (Fireworks)
• License: Modified MIT (commercial use allowed)
• SWE-Bench Verified: 80.2%
• AA Coding Index: 47.1
• Terminal-Bench Hard: 0.439
• SciCode: 0.535
• Tool calling: Workable (thinking-mode restricts tool_choice to auto/none; full reasoning_content must persist across turns)
• Prompt caching: Automatic on Fireworks Serverless
• Fine-tunable on Fireworks: Yes (LoRA, SFT via FireOptimizer)

Kimi K2.6 is Moonshot AI’s flagship open-source model released April 20, 2026, scoring 47.1 on the AA Coding Index, second only to DeepSeek V4-Pro’s 47.5 among open-source models in this list. Architecturally it is an MoE with 384 experts plus one always-on shared expert, 61 layers, Multi-head Latent Attention (MLA), SwiGLU activation, and a 160K vocabulary. Moonshot pitches it as a long-horizon agentic coding model capable of orchestrating 300 sub-agents and 4,000 coordinated steps on hard software engineering tasks.

Kimi K2.6 on Fireworks serves at 6.25x lower output cost than Claude Opus 4.7. It trails Opus 4.7 by 5.4 points on the composite Coding Index (47.1 vs 52.5), ties on SciCode within 0.010 (0.535 vs 0.545), and comes within 3.4 points on AA Intelligence (53.9 vs 57.3), the closest open-source approach to the current frontier.

🚀 See the Kimi K2.6 API details on Fireworks.

Teams building coding agents that need a 256K working memory, sub-second TTFT, and the ability to fine-tune on domain data. If you have a coding workload where Claude Opus 4.7 hits your quality bar and you are paying $5/$25 per 1M tokens for it, Kimi K2.6 on Fireworks closes most of the cost gap at a 6.25x drop in per-output-token cost ($0.95/$4.00 per 1M), and narrows the quality gap to 5.4 points on the AA Coding Index (47.1 vs 52.5). Fine-tuning with FireOptimizer closes the remaining quality gap for your specific eval.

• K2.6 ships with thinking mode enabled by default and requires both -tool-call-parser kimi_k2 and -reasoning-parser kimi_k2 flags on vLLM/SGLang. Without both, tool calls and thinking content leak into the wrong output fields. Source: Kimi K2.6 deploy guidance.
• When thinking mode is active, tool_choice is restricted to auto or none; any other value errors. Multi-step tool loops require keeping the full reasoning_content from every prior assistant message in context, which inflates prompt cost. Source: Kimi platform docs.
• The built-in $web_search tool is incompatible with thinking mode. Retrieval-heavy coding workflows must disable thinking for search steps, which defeats K2.6’s interleaved-tool-use design.
• For long-horizon browsing evaluations, K2.6 uses a discard-all context management strategy: once accumulated input exceeds the 256K window, prior tool outputs are hidden and only the most recent round of tool-related messages is retained. Long-running agents can silently lose tool history mid-task. Source: Kimi K2.6 model card.
• Native INT4 quantization applies only to MoE components, not dense layers. Benchmark-quality tool use requires self-hosted deployment or a production platform that runs the full config; kimi.com‘s chat mode uses a reduced tool subset that does not reproduce published scores.

Q: How does Kimi K2.6 compare to Claude Opus 4.7 on coding?

On the AA Coding Index, Opus 4.7 leads by 5.4 points (52.5 vs 47.1). Kimi K2.6 closes the SciCode gap to within 0.010 (0.535 vs 0.545) and the AA Intelligence gap to 3.4 points (53.9 vs 57.3). Price-per-output-token is 6.25x lower on Fireworks.

Q: Can Kimi K2.6 handle agentic coding with 100+ tool calls?

Moonshot’s own claims describe 300-sub-agent and 4,000-step orchestration. Operationally, plan for the discard-all context strategy when accumulated input clears 256K, and budget for the inflated prompt cost of retaining reasoning_content across turns.

Q: Is fine-tuning Kimi K2.6 on Fireworks practical?

Yes. FireOptimizer supports LoRA and SFT on Fireworks-hosted MoE models. Fireworks serves hundreds of LoRA adapters per base model, which keeps per-tenant adaptation costs linear rather than quadratic.

🚀 View Kimi K2.6 on Fireworks →

• Parameters: 754B total, 40B active (MoE)
  • Note: Artificial Analysis reports 744B / 40B for this model
• Context window: 200K tokens (128K max output)
• Input $/1M: $1.40 (Fireworks)
• Output $/1M: $4.40 (Fireworks)
• License: MIT
• SWE-Bench Pro: 58.4 (vendor-reported); SWE-Bench Verified not published for GLM-5.1 (the 77.8% on the linked Z AI page refers to GLM-5, the predecessor)
• AA Coding Index: 43.4
• Terminal-Bench Hard: 0.432
• SciCode: 0.438
• Tool calling: Issues (multiple open bugs: content-stripping in multi-turn, malformed JSON, English-to-Chinese drift)
• Prompt caching: Automatic on Fireworks Serverless
• Fine-tunable on Fireworks: Yes (LoRA, SFT via FireOptimizer)

GLM-5.1 is Z AI’s flagship reasoning model released April 7, 2026. It uses a Dynamic Sparse Attention MoE (GLM_MOE_DSA) with 256 experts and 8 activated per token. It ships under an MIT license with open weights, and supports deployment on vLLM and SGLang.

On AA’s composite Coding Index, GLM-5.1 sits at 43.4. On Z AI’s own SWE-Bench Pro evaluation (not in AA’s tracked benchmarks), Z AI reports 58.4, tipping past GPT-5.4 (self-reported 57.7) and Claude Opus 4.6 (57.3). The benchmark divergence between AA-tracked and vendor-reported scores is real; both are worth consulting.

🚀 See the GLM-5.1 API details on Fireworks.

Teams who need the most permissive open-source license in the list for commercial deployment (MIT) and who can absorb the tool-calling fragility described below. Strong fit for structured reasoning, agent engineering, and teams that want to fine-tune without license friction.

• On vLLM with -tool-call-parser glm47 and -chat-template-content-format auto, GLM-5.1-FP8 ignores tool-response content in multi-turn conversations. The actual prompt sent to the model replaces tool results with an empty <tools>\\\\n</tools> XML wrapper. Workaround: force -chat-template-content-format string. Source: vllm-project/vllm#39614 (open).
• GLM-5.1 spontaneously switches to Chinese output mid-conversation in English-only contexts. Unsafe for English-only coding assistants without post-filtering. Source: zai-org/GLM-5#54 (open).
• GLM-5-FP8 served via vLLM exhibits tool-call failures in Claude Code. Tool calls degrade or silently fail when routing Claude Code traffic to GLM-5 FP8. No fix merged. Source: zai-org/GLM-5#45 (open).
• Official Zhipu API does not expose an OpenAI-compatible /responses endpoint. Codex-style clients that depend on responses-API semantics must proxy through vLLM or build translation layers. Source: zai-org/GLM-5#39 (open).
• NVIDIA NIM deployments intermittently return malformed JSON arguments (missing/truncated braces) during tool calls, triggering infinite retry loops. Source: zai-org/GLM-5#15 (open).

Q: Why does GLM-5.1 score below Kimi K2.6 on AA but claim to beat Claude Opus 4.6 on SWE-Bench Pro?

The benchmarks measure different things. AA’s Coding Index is a composite of Terminal-Bench Hard and SciCode, where Kimi K2.6 leads GLM-5.1. Z AI’s reported 58.4 is specifically SWE-Bench Pro, a separate evaluation not in the AA Coding Index. Scores are vendor-reported. Always run your own eval against your own code.

Q: Can GLM-5.1 be fine-tuned?

Yes. The MIT license is permissive. FireOptimizer supports LoRA and SFT on GLM-5.1 on Fireworks.

Q: Does the training hardware path affect deployment?

Not for inference. GLM-5.1 serves on standard Nvidia GPUs through vLLM, SGLang, and Fireworks.

🚀 View GLM-5.1 on Fireworks →

• Parameters: 396B total (MoE; Fireworks model page)
• Context window: 1M tokens (65K max output)
• Input $/1M: $0.50 (Fireworks)
• Output $/1M: $3.00 (Fireworks)
• License: Proprietary (DashScope API-only)
• SWE-Bench Verified: 78.8% (vendor-reported)
• AA Coding Index: 42.9
• Terminal-Bench Hard: 0.439
• SciCode: 0.407
• Tool calling: Workable (open-weight Qwen3 family shows XML-vs-JSON format quirks under SGLang; Plus inherits architecturally)
• Prompt caching: Automatic on Fireworks Serverless
• Fine-tunable on Fireworks: No for Plus (closed-weight); lower-tier Qwen3.6 variants are fine-tunable

Qwen3.6 Plus is Alibaba’s flagship Qwen3.6 model released April 2, 2026. It uses a hybrid architecture combining linear attention with sparse mixture-of-experts routing, which is what makes the 1M context window feasible without computational costs exploding. Always-on chain-of-thought reasoning is the default behavior.

Important buying note: Qwen3.6 Plus is closed-weight. Unlike the lower-tier Qwen3 releases (Qwen3.6-35B-A3B, coder variants), Plus is accessible only through Alibaba Cloud’s DashScope / Bailian platform and through partner inference providers including Fireworks. Self-hosting, air-gapped deployment, and direct fine-tuning are not available for Plus. On Fireworks, Plus is served through the same Serverless API as the open-weight Qwen models, which removes the DashScope-specific operational overhead.

🚀 See the Qwen3.6 Plus API details on Fireworks.

Teams that need long-context coding analysis (1M tokens for whole-repo or whole-codebase in-context) at $0.50/M input pricing. Vision-capable workloads where the coding task includes UI mockups or visual artifacts. Teams willing to accept closed-weight constraints in exchange for the larger context window and vision capability.

• Closed-weight, API-only. Self-hosting, on-prem deployment, air-gapped workloads, and direct fine-tuning are not supported on the Plus variant. Lower-tier Qwen3 releases remain open-weight. Source: Alibaba Cloud DashScope docs.
• Direct DashScope access requires a DashScope-specific API key (sk-prefixed), not an Alibaba Cloud AccessKey pair, and each model must be individually activated in the Model Studio console. Fireworks-hosted access avoids this auth surface entirely. Source: DashScope first API call docs.
• The open-weight Qwen3 family (architectural sibling, proxy for Plus behavior) has documented chat-template issues: empty historical <think> blocks emitted into prompts cause drift and prompt-cache invalidation every turn. Source: QwenLM/Qwen3.6#131 (open).
• Open-weight Qwen3.5 9B / 35B-A3B produce XML-style tool calls instead of JSON under SGLang’s tool-call parser. Downstream agent harnesses built for JSON will fail. Source: QwenLM/Qwen3.6#125 (open).
• Qwen3.5-27B (same-family open weight) generates gibberish and repetitive loops on long context under vLLM 0.17.0. Long-context stability work is in progress across the Qwen3.x family. Source: QwenLM/Qwen3.6#115 (open).

Q: Can I fine-tune Qwen3.6 Plus?

Not directly. Plus is closed-weight and API-only. For fine-tunable Qwen3.6 variants, the 35B-A3B open-weight model is available on Fireworks and supports LoRA and SFT via FireOptimizer.

Q: How does Qwen3.6 Plus compare to Kimi K2.6?

Kimi K2.6 leads on the AA Coding Index by 4.2 points (47.1 vs 42.9) and has a permissive Modified MIT license with open weights. Qwen3.6 Plus has a nearly 4x larger context window (1M vs 256K) and is vision-capable.

Q: Is there a SWE-bench score?

Alibaba reports strong SWE-bench numbers in their own coverage. AA does not track Plus on SWE-bench. Run a domain-matched eval against your code before committing.

🚀 View Qwen3.6 Plus on Fireworks →

• Parameters: 284B total, 13B active (MoE)
• Context window: 1M tokens (384K max output)
• Input $/1M: $0.14 (DeepSeek API, cache miss) / $0.028 (cache hit)
• Output $/1M: $0.28 (DeepSeek API)
• License: MIT (open weights)
• Release date: April 24, 2026
• SWE-Bench Verified: 79.0% (vendor-reported)
• AA Coding Index: 38.7
• Terminal-Bench Hard: 0.356
• SciCode: 0.449
• Tool calling: Workable (dual-mode OpenAI + Anthropic APIs; same interface as V4-Pro)
• Prompt caching: Automatic on DeepSeek’s API (cache hit at 20% of cache-miss input price)
• Fine-tunable on Fireworks: Not yet; the open weights support external fine-tuning

DeepSeek V4-Flash is the lower-tier variant in the V4 family, released alongside V4-Pro on April 24, 2026. It trades parameter count for price: 284B total / 13B active MoE at $0.14 / $0.28 per 1M tokens. The AA Coding Index puts it at 38.7, above gpt-oss-120B (28.6) and within striking distance of Qwen3.6 Plus (42.9), and the vendor tech report shows its agentic and coding scores closely tracking V4-Pro at a fraction of the price (SWE-Bench Verified 79.0% vs V4-Pro’s 80.6%, LiveCodeBench 91.6 vs 93.5).

The important tradeoff vs gpt-oss-120B: V4-Flash wins on every AA quality metric and offers 1M context and MIT weights, and on AA’s standardized workload V4-Flash on the DeepSeek API (~84 tok/s) serves slightly faster than gpt-oss-120B on Fireworks (70 tok/s). gpt-oss-120B’s only edges over V4-Flash today are Fireworks hosting (V4-Flash isn’t yet), the Apache 2.0 license, and the option to fine-tune via FireOptimizer.

V4-Flash was not yet on Fireworks at time of publication.

Teams running high-volume coding agents where per-token cost is the binding constraint and where ~84 tok/s is fast enough for the loop. Teams that need 1M context at the lowest price in this guide. Teams migrating off deepseek-chat or deepseek-reasoner before the July 24, 2026 retirement, since both legacy endpoints currently route to V4-Flash. Teams running noncommercial research or eval pipelines where the MIT-style open weights eliminate license friction.

• At ~84 tok/s on the DeepSeek direct API (per Artificial Analysis), V4-Flash is slightly faster than gpt-oss-120B on Fireworks (70 tok/s per AA’s standardized workload). High-throughput autocomplete use cases should run both models on the same eval before choosing. Source: Artificial Analysis V4-Flash page.
• V4-Flash’s 13B active parameters leave it well behind V4-Pro on agentic tasks: Terminal-Bench 2.0 56.9% vs 67.9% (11-point gap) and SimpleQA-Verified 34.1% vs 57.9%. Workloads that involve multi-step tool use or factual recall should validate V4-Flash specifically, not extrapolate from V4-Pro. Source: DeepSeek V4 tech report.
• Not on Fireworks at time of publication.
• deepseek-chat and deepseek-reasoner currently route to V4-Flash (non-thinking and thinking modes respectively) and retire July 24, 2026. Production callers on the legacy model names should update to deepseek-v4-flash explicitly to avoid surprise behavior differences during and after the transition. Source: DeepSeek V4 release notes.
• DSA and dual-mode behavior are new. Self-hosters attempting vLLM or SGLang deployments should expect parser work on the thinking/non-thinking switch and cache-hit accounting that differs from V3.x.

Q: Is V4-Flash a replacement for gpt-oss-120B on high-throughput autocomplete?

Only at the quality tier. V4-Flash leads gpt-oss-120B on every AA quality metric. Input pricing is near-identical ($0.14 vs $0.15 per 1M), and V4-Flash’s output is less than half of gpt-oss-120B’s ($0.28 vs $0.60). On AA’s standardized workload, V4-Flash on the DeepSeek API (~84 tok/s) runs slightly faster than gpt-oss-120B on Fireworks (70 tok/s). For a Fireworks-native deployment today (V4-Flash isn’t yet), gpt-oss-120B is the available choice at the lowest output price; for quality-weighted reasoning at near-identical input price, V4-Flash on the DeepSeek API is better.

Q: Can V4-Flash handle 1M context reliably?

Long-context behavior is vendor-claimed; independent long-context benchmarks are still early. The vendor-reported MRCR 1M score is 78.7% and CorpusQA 1M is 60.5%, both below V4-Pro. Workloads that push the 1M window should plan independent evals.

Q: Will V4-Flash land on Fireworks?

Probable but not confirmed at time of publication. V4-Pro landed on Fireworks shortly after its April 24, 2026 release, and Fireworks’ day-zero pattern has held for Kimi K2.6 (April 20), DeepSeek V3.x, and other MoE releases. Check fireworks.ai/models for current availability.

• Parameters: 117B total, 5.1B active (MoE)
• Context window: 128K tokens
• Input $/1M: $0.15 (Fireworks)
• Output $/1M: $0.60 (Fireworks)
• License: Apache 2.0
• Release date: August 5, 2025 (oldest model in this list)
• SWE-Bench Verified: ~30% (community-reported)
• AA Coding Index: 28.6
• Terminal-Bench Hard: 0.235
• SciCode: 0.389
• LiveCodeBench: 0.878
• Tool calling: self-hosted has issues that have been resolved on Fireworks (e.g. parallel calling broken in current HF weights; vLLM Chat Completions parsers don’t reliably parse Harmony; works cleanly only via /v1/responses)
• Prompt caching: Automatic on Fireworks Serverless
• Fine-tunable on Fireworks: Yes (Apache 2.0; LoRA, SFT via FireOptimizer)

gpt-oss-120B is OpenAI’s first modern open-weights release, shipped August 5, 2025 under Apache 2.0. At 117B total / 5.1B active, it is the smallest MoE in this list by active parameters. AA’s per-provider data puts Fireworks throughput at 70 tokens/sec on the standardized 10K-input workload, with the all-provider median around 190 tokens/sec for this model and Cerebras and SambaNova above 600 tok/s; gpt-oss-120B is therefore not the throughput choice on Fireworks, but it is the lowest-cost model in the list, and at $0.15/$0.60 per million tokens it costs roughly 42x less than Claude Opus 4.7 on output, though at roughly half Opus 4.7’s AA Coding Index (28.6 vs 52.5) and best suited to function-level work.

The benchmark pattern is the sharpest divergence in the list: lowest Coding Index (28.6, vs Opus 4.7’s 52.5), lowest Terminal-Bench Hard (0.235, vs Opus 4.7’s 0.515), and LiveCodeBench 0.878, the only model in this guide with AA-tracked LiveCodeBench coverage. DeepSeek V4-Pro and V4-Flash report vendor LiveCodeBench scores of 93.5 and 91.6 that AA has not yet tracked. Function-level code generation is strong; long multi-step agentic coding runs are not.

🚀 See the gpt-oss-120B API details on Fireworks.

High-throughput function-level coding workloads: autocomplete, test generation, one-shot refactors, code-explanation endpoints, where speed and cost dominate the decision and where the coding task is scoped to short-horizon code generation. The Apache 2.0 license makes it the most permissive option in the list for commercial deployment and modification. Teams doing retrieval-heavy coding agents where gpt-oss serves as the fast inner-loop model and a larger model handles orchestration.

• Parallel tool calling is broken in the current open weights. The fix that prevents post-tool-call hallucination (commit 8b193b0) removed <|call|> as a stop token, so the model now returns one tool call at a time even when reasoning correctly identifies multiple are needed. Source: openai/gpt-oss-120b discussion #151.
• Streaming tool calls fail mid-stream on vLLM. A portion of the tool-call payload gets written into the reasoning channel instead of tool_calls, breaking OpenAI-format parsers. Source: vllm-project/vllm#27641 (open).
• On Blackwell DGX Spark (SM121) with MXFP4 weights, vLLM’s Marlin kernel generates a wrong first Harmony token. The content and reasoning fields arrive null for the first chunk, corrupting downstream parsers that expect Harmony’s channel markers. Source: vllm-project/vllm#37030 (open).
• Tool calling works reliably only via vLLM’s /v1/responses (Harmony) endpoint. The OpenAI Chat Completions /v1/chat/completions endpoint with any -tool-call-parser option fails: the hermes parser crashes startup; mistral and llama3_json parsers produce parameter errors. Source: vLLM gpt-oss recipe docs.
• The shipped Jinja chat template has errors compared to the Harmony spec: missing <|constrain|> token, incorrect CoT/analysis misattribution on pre-tool-call content, and wrong turn-end token (<|end|> instead of <|return|>). Raw transformers usage produces malformed outputs until third-party template fixes are applied. Source: openai/gpt-oss-120b discussion #69.

Q: Why does gpt-oss-120B score so low on Coding Index but so high on LiveCodeBench?

The composite Coding Index weights agentic and long-horizon code tasks heavily, which is where gpt-oss struggles (Terminal-Bench Hard 0.235). LiveCodeBench is function-level code generation, where gpt-oss scores 0.878, the only AA-tracked LiveCodeBench score in this guide. DeepSeek’s own tech report puts V4-Pro at 93.5 and V4-Flash at 91.6 on LiveCodeBench; AA has not yet published those numbers. If your workload is autocomplete-shaped, gpt-oss is a strong and very cheap fit and ships today.

Q: Is the Harmony format a barrier?

Yes for raw-weight or raw-vLLM usage, no on Fireworks. Fireworks handles the Harmony format at the platform level, presenting a standard OpenAI-compatible API to the developer.

Q: Can gpt-oss-120B replace GPT-4-class models for coding?

For short-horizon function generation, yes, at roughly 42x lower output cost than Opus 4.7. gpt-oss is a strong low-cost fit. For multi-step agentic coding with heavy tool use, no: the Coding Index and Terminal-Bench Hard gaps are large. Use it as the fast inner-loop model paired with a stronger orchestrator.

🚀 View gpt-oss-120B on Fireworks →

Five open-source models in this guide are live on Fireworks today: Kimi K2.6, GLM-5.1, Qwen3.6 Plus, gpt-oss-120B, and DeepSeek V4-Pro. All five serve at a fraction of the closed frontier price, with DeepSeek V4-Flash expected to follow. For open-source coding models in production, Fireworks is the scale layer.

FireAttention is Fireworks’ proprietary inference engine. Per the FireAttention V2 benchmarks:

• 3-12x lower latency versus vLLM.
• Up to ~5.6x higher throughput single-host, up to ~8x in multi-host FP8 mode.
• Adaptive speculative decoding reaches roughly 1,000 tokens/sec on workloads like Cursor’s fine-tuned Llama-3-70b.

Live Fireworks traffic now serves more than 10 trillion tokens per day.

Day-zero support is a standing commitment. Kimi K2.6 released April 20, 2026 and was live on Fireworks the same day. New weights across DeepSeek, Qwen, Kimi, GLM, Llama, and GPT-OSS families typically ship to Fireworks within 24 hours of public release.

FireOptimizer is the Fireworks post-training stack: SFT, LoRA adapters, RFT, RL pipelines, and evaluation infrastructure. Three shipped results on Fireworks:

• v0 (Vercel): error-free generation reached 93% with RL post-training on an open model. Sonnet 3.5 compiled at 62% on the same eval. Vercel CTO Malte Ubl: “a fine-tuned reinforcement learning model with Fireworks, we perform substantially better than SOTA.”
• Cursor: ~9x speedup over its previous GPT-4 speculative-edits deployment at roughly 1,000 tokens/sec after switching speculative decoding to Fireworks.
• Cresta: thousands of LoRA adapters for per-domain fine-tunes on Fireworks, with a documented 100x cost reduction versus GPT-4.

Prompt caching is automatic on Fireworks Serverless. Coding workloads repeat large system prompts (agent scaffolding, tool definitions, codebase stubs) on every turn, and caching cuts the input portion of those repeats without a code change. The open-source models here get it for free.

For comparison on cached input discounts:

• Anthropic: 90% read discount.
• Google: up to 90% on Gemini 2.5+.
• OpenAI: 50-90%, depending on the model family.

FireOptimizer supports LoRA, supervised fine-tuning (SFT), and reinforcement fine-tuning (RFT) on every Fireworks-hosted open-weight model in this guide (Kimi K2.6, GLM-5.1, gpt-oss-120B, DeepSeek V4-Pro).

DeepSeek V4-Flash is open-weight and will inherit FireOptimizer support once Fireworks adds it to the catalog; until then it can be fine-tuned through alternative open-weight pipelines. The closed-weight Qwen3.6 Plus is not fine-tunable on any platform; its lower-tier open-weight siblings are.

Rough rules of thumb from Fireworks’ own guidance:

• More than 1,000 labeled samples: start with SFT or LoRA. Fireworks bills fine-tuning per million training tokens. Training token count equals dataset tokens times number of epochs.
• Fewer than 1,000 samples: consider RFT if the task is verifiable (pass/fail, compile/no-compile, test suite). v0’s error-free generation rate lift came from this path.
• Multi-tenant adaptation: Fireworks serves hundreds of LoRA adapters per base-model deployment at no extra inference cost. This keeps per-tenant adaptation costs linear rather than quadratic.
• Serving: fine-tuned models run at the same per-token price as the base model on Fireworks.

The full fine-tuning walkthrough is in the Fireworks docs under docs.fireworks.ai/fine-tuning/.

Three tiers cover the workload spectrum:

• Serverless: pay-per-token, no cold boots, Fireworks-curated hardware. The default entry point. Your first API call runs within minutes at fireworks.ai. Shared rate limits, industry-leading speed on a Fireworks-tuned stack.
• On-Demand: private GPUs, pay-per-GPU-second, scales to zero when idle. Fireworks software on dedicated GPUs delivers significantly faster per-GPU throughput than vLLM. The right tier when production volume justifies dedicated capacity and you need deterministic performance without sharing.
• Enterprise Reserved: dedicated reserved GPUs, SLAs, priority support, bring-your-own-cloud. For regulated industries, compliance constraints, or workloads requiring capacity guarantees.

OpenAI-compatible API. Point your existing client library at https://api.fireworks.ai/inference/v1 and ship. One-line migration from any OpenAI-compatible code path.

FireOptimizer covers the full post-training stack: SFT, LoRA adapters (serve hundreds concurrently on a single base), RFT, RL pipelines, and evaluation infrastructure. Adaptive speculative decoding with automatic draft-model training is on by default.

Hundreds of open-source models across text, audio, image, embedding, and vision. Day-zero support on frontier releases. Multi-cloud and multi-hardware backends so your workload runs on the right silicon without you configuring it.

There is no single best LLM for coding in 2026. The AA Coding Index puts GPT-5.5, GPT-5.4, Gemini 3.1 Pro Preview, and Claude Opus 4.7 at the top of the raw benchmarks, the right call for experimentation and proving a coding product works.

GPT-5.5 (April 23, 2026) is the new frontier on AA coding and intelligence composites, though Claude Opus 4.7 still leads SWE-Bench Verified and GPT-5.5’s API is still rolling out. Once the workload is real and the volume is scaling, the binding constraints shift to cost per output token, throughput at peak load, and fine-tuning headroom.

That is where the open-source pack becomes the strategic choice, led by DeepSeek V4-Pro and Kimi K2.6 (effectively tied on the AA Coding Index at 47.5 and 47.1, with Kimi K2.6 ~2x faster and ~20% cheaper blended; V4-Pro a better fit for Terminal-Bench-class agentic coding), with V4-Flash at $0.14/$0.28 per 1M as the cost floor and GLM-5.1, Qwen3.6 Plus, and gpt-oss-120B filling specific workload niches (most permissive MIT license + Fireworks hosting, largest context at Fireworks pricing, lowest output price on Fireworks respectively).

Post-training is what flips the rankings on a specific workload. The v0 and Cursor results above show the pattern: an open model tuned on production data, running on Fireworks, outperforms closed frontier models on the eval that actually matters, at a fraction of the per-token cost.

For developers running Kimi in a personal coding harness, Fire Pass is an Early Access flat-rate pass at $7 per week. It covers unlimited requests to Kimi K2.5 Turbo, a faster private-preview variant of Kimi K2.5, in Claude Code, OpenCode, Cline, Kilo Code, OpenClaw, and LangChain Deep Agents. Per-token charges do not apply to that model. No invite code is required. Fire Pass covers K2.5 Turbo specifically, not the Kimi K2.6 featured above.

🚀 Ready to run these models in production?

Get started at fireworks.ai. No sales process required.

The information provided in this article is accurate at the time of publication. Model capabilities, benchmarks, and availability change frequently. Always verify current specifications on official model repositories.