Coding agents at a fraction of frontier cost
Fireworks Nexus orchestrates open and closed models inside Claude Code, Codex, Cursor and OpenCode, so you cut cost per task without changing your workflow.
Fireworks Nexus orchestrates open and closed models inside Claude Code, Codex, Cursor and OpenCode, so you cut cost per task without changing your workflow.
How inference, training and orchestration on Fireworks apply to coding agents.
Prototype on serverless, then move production agents to on-demand with the agentic-coding preset; Cursor Fast Apply runs at ~1000 tokens/sec on Fireworks.
When an open model misses your bar, Reinforcement Fine-Tuning (RFT) goes from evaluator to deployed model in hours; Vercel v0 reached 93% error-free generation with it.
As spend grows, Nexus auto sends easy work to open models and hard work to your closed provider; Factory runs Kimi K2.7 at 20% of frontier cost.
Why use Fireworks for coding agents?
Composer 2 costs Cursor 6–10x less to serve than comparable frontier models.
Keep Claude Code or Codex. Nexus adds open models next to the closed models you already use.
Sourcegraph cut latency 30% and lifted acceptance from 15% to 40%.
Open models run with Zero Data Retention by default. See how below.
Run the installer from nexus.fireworks.ai. The Nexus CLI is open source and works with Claude Code, Codex, Cursor, VS Code, Copilot, OpenCode and more.
Run nexus login, then sign in with your browser or paste a Fireworks API key.
Run nexus claude. Claude Code now runs on Nexus auto, which picks an open or closed model for each turn.
Keep Nexus auto, or pin an open model such as GLM 5.2 Fast.
Move production agents to an on-demand deployment or Reserved Throughput. The FAQ below compares pricing and SLAs.
Coding agents send your files as context. The Fireworks security docs set out what happens to that data.
Fireworks does not log or store prompt or generation data for any open model. Logging starts only when you explicitly opt in.
With storage on, the Response API keeps data for 30 days. It is the one documented exception to Zero Data Retention.
These are the open models the Fireworks docs currently recommend for code, plus a code-tuned Kimi variant listed for the Nexus CLI.
Token price alone hides the real number. Track what one finished task costs, including retries and the repository context your agent resends each turn. Compare that figure across models before you move traffic, because a cheaper model that needs extra retries can cost more.
Many agent turns, like reading a file or making a small edit, don’t need a frontier model. Pin an open model for that work, or let Nexus auto make the call on each request. If most of your tasks need the frontier model, orchestration saves less.
Coding agents resend the same system prompt and files on every turn. Prompt caching is on by default and bills cached tokens at 50% off. Keep stable content, such as instructions and a repository map, at the front of the prompt and put the changing request last.
Yes. The Nexus CLI adds Fireworks open models and Nexus auto to Claude Code or Codex. The steps above show the setup.
Serverless bills per token and has no SLA. On-demand bills per GPU-hour, and Reserved Throughput is SLA-backed.
Not by default. Open models run with Zero Data Retention unless you opt in. The Response API’s storage setting is the one exception; see the Security section above.
Prompt caching is on by default and cuts time to first token by as much as 80%. On dedicated deployments, speculative decoding lowers generation latency by having the model verify proposed tokens in parallel.
No. Your harness keeps its commands and workflow, and you can reverse the Nexus CLI setup.
Keep Claude Code, Codex, Cursor or OpenCode. Nexus orchestrates open and closed models behind them to cut your cost per task.
