Benchmark Almanac

DeepRSI

Claude Opus 5.5 1,629 leads since October 2026
Not in model cards

15 open-ended ML tasks with a 48-hour budget each. Private verifiers score gains over baselines. Models ranked by Bradley-Terry Elo.

Tasks come from proprietary research. Repo has only 5 sample tasks. Agents see a public verifier, but only the private one counts.

Made by
Hillclimb
Released
October 2026
Kind
AI research tasks
Level
Research frontier
Scored
Checked automatically
Answers
Agent task
Tasks
Kept private
Strongest model tested
GPT-6 Astra #3 on the capability index
Last model added
October 2026
Scores refreshed
November 12th, 2026
8
Models scored on its own leaderboard

Claude Opus 5.5 leads with 1,629

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

Bars start at 600.

  1. 1,629 Claude Opus 5.5
  2. 1,548 Claude Fable 5.1
  3. 1,232 GPT-6 Astra
  4. 965 Gemini 4 Argon
  5. 955 Muse Wonder
  6. 820 Kimi K3
  7. 685 Grok 4.7
  8. 643 Gemini 3.8 Flash

Bars start at 600.

  1. 1 Claude Opus 5.5 max 1,629
  2. 2 Claude Fable 5.1 max 1,548
  3. 3 GPT-6 Astra max 1,232
  4. 4 Gemini 4 Argon high 965
  5. 5 Muse Wonder xhigh 955
  6. 6 Kimi K3 max 820
  7. 7 Grok 4.7 xhigh 685
  8. 8 Gemini 3.8 Flash high 643

Source: Official DeepRSI leaderboard

#ModelScoreSettingDateSource
1 Claude Opus 5.5 1,629 max Sep 2026 Official leaderboard
2 Claude Fable 5.1 1,548 max Sep 2026 Official leaderboard
3 GPT-6 Astra 1,232 max Sep 2026 Official leaderboard
4 Gemini 4 Argon 965 high Sep 2026 Official leaderboard
5 Muse Wonder 955 xhigh Oct 2026 Official leaderboard
6 Kimi K3 820 max Jul 2026 Official leaderboard
7 Grok 4.7 685 xhigh Sep 2026 Official leaderboard
8 Gemini 3.8 Flash 643 high Sep 2026 Official leaderboard

Elo against average tokens per attempt

Each point is one model at one reasoning effort. Tokens count every model token in an attempt, input included. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

0 500 1,000 1,500 2,000 0255075100125 Average tokens per attempt (million), lower is better Elo Kimi K3 (max): Elo 820, average tokens per attempt 35.6 million Gemini 3.8 Flash (high): Elo 643, average tokens per attempt 111 million Grok 4.7 (xhigh): Elo 685, average tokens per attempt 21.6 million Gemini 4 Argon (high): Elo 965, average tokens per attempt 119 million Muse Wonder (xhigh): Elo 955, average tokens per attempt 11.2 million Claude Fable 5.1 (max): Elo 1,548, average tokens per attempt 46.5 million GPT-6 Astra (max): Elo 1,232, average tokens per attempt 7.84 million Claude Opus 5.5 (max): Elo 1,629, average tokens per attempt 85.4 million Claude Opus 5.5 (max)Claude Fable 5.1 (max)GPT-6 Astra (max) 0 500 1,000 1,500 2,000 0255075100125 Average tokens per attempt (million), lower is better Elo Kimi K3 (max): Elo 820, average tokens per attempt 35.6 million Gemini 3.8 Flash (high): Elo 643, average tokens per attempt 111 million Grok 4.7 (xhigh): Elo 685, average tokens per attempt 21.6 million Gemini 4 Argon (high): Elo 965, average tokens per attempt 119 million Muse Wonder (xhigh): Elo 955, average tokens per attempt 11.2 million Claude Fable 5.1 (max): Elo 1,548, average tokens per attempt 46.5 million GPT-6 Astra (max): Elo 1,232, average tokens per attempt 7.84 million Claude Opus 5.5 (max): Elo 1,629, average tokens per attempt 85.4 million Claude Opus 5.5 (max)Claude Fable 5.1 (max)GPT-6 Astra (max)
ModelSettingEloAverage tokens per attempt
Claude Opus 5.5 max 1,629 85.45 million
Claude Fable 5.1 max 1,548 46.53 million
GPT-6 Astra max 1,232 7.84 million
Gemini 4 Argon high 964.86 119.26 million
Muse Wonder xhigh 954.89 11.22 million
Kimi K3 max 820.13 35.58 million
Grok 4.7 xhigh 685.24 21.59 million
Gemini 3.8 Flash high 642.62 110.51 million

The best score, 1,629, costs $494 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

OpenAI Anthropic Google DeepMind Meta Cost frontier One model at each effort level
0 500 1,000 1,500 2,000 $100 $200 $500 $1,000 Cost per task, log scale Claude Opus 5.5, max GPT-6 Astra, max Gemini 4 Argon, high Muse Wonder, xhigh 0 500 1,000 1,500 2,000 $100 $1,000 Cost per task, log scale Claude Opus 5.5, max GPT-6 Astra, max

Source: Official DeepRSI leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 5.5 max 1,629 $494
Claude Fable 5.1 max 1,548 $628
GPT-6 Astra max 1,232 $404
Gemini 4 Argon high 965 $278
Muse Wonder xhigh 955 $107
Kimi K3 max 820 $159
Grok 4.7 xhigh 685 $282
Gemini 3.8 Flash high 643 $129

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Own leaderboard 1 leader
Claude Opus 5.5 1,629

Source: Official DeepRSI leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard
750 1,000 1,250 1,500 1,750 750 1,000 1,250 1,500 1,750

Source: Official DeepRSI leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in AI research

All 104 in AI research
  • PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics. Qwen 3.8 Max 93%
  • MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
  • PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
  • MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6 Astra 93.8%
  • KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup. MiniMax-M3 28.8%
  • PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Inkling 2 Small 494.3