FrontierMemory

Fifteen coding and knowledge-work tasks adapted from DeepSWE, BigCodeBench, AutomationBench, APEX-Agents, EnterpriseOps-Gym, and SkillsBench so that each needs something only earlier runs show: a task’s unstated contract, a user’s preferences, or where the environment keeps what the agent needs. The model under test reads 750 past agent trajectories, labeled success or failure and most of them unrelated, and writes memory in Letta Dream for up to six hours with no internet. A Letta Code agent then attempts each task with that memory, and the score is the share it fully passes with the better of two task models at max effort, averaged over three independent runs of FrontierMemory v1.1.

Last updated · Source leaderboard checked daily

Main Insights

Results

FrontierMemory results, ranked by score (downstream solve rate), highest first. Last updated .
RankModelProviderScore (downstream solve rate)Cost (USD / task)
1Claude Opus 5.5Anthropic73.3%$3.42
2Claude Fable 5.2Anthropic73.3%$3.71
3DeepSeek V4.1 FlashDeepSeek71.1%$3.86
4Gemini 4 CarbonGoogle71.1%$3.99
5Kimi K3Moonshot AI71.1%$4.16
6Claude Haiku 5.5Anthropic68.9%$3.98
7GLM 5.3Z.ai68.9%$4.08
8Gemini 4 ArgonGoogle64.4%$1.05
9Gemini 4 HydrogenGoogle64.4%$4.02
10Inkling 2 SmallThinking Machines60.0%$4.18
11Kimi K3.1 ProMoonshot AI55.6%$3.61
12GPT-6.1 SolOpenAI51.1%$4.00
13Muse Wonder 1.0Meta48.9%$3.71
14GPT-6.1 TerraOpenAI46.7%$3.97
15Muse EmberMeta42.2%$3.60
16Grok 4.7xAI37.8%$3.70
17GPT-6 LunaOpenAI28.9%$4.05