Benchmark Almanac

ProgramBench

Claude Opus 5.5 18.5% leads since Sep 2026 In use

200 tasks: rebuild a full program from only its compiled binary and docs. Hidden behavioral tests (248k total) must match the original.

Tasks range from jq to FFmpeg and SQLite. The Opus 5.5 system card reports a 100-agent swarm result.

In the authors’ words

Turning ideas into full software projects from scratch has become a popular use case for language models. Agents are being deployed to seed, maintain, and grow codebases over extended periods with minimal human oversight. Such settings require models to make high-level software architecture decisions. However, existing benchmarks measure focused, limited tasks such as fixing a single bug or developing a single, specified feature. We therefore introduce ProgramBench to measure the ability of software engineering agents to develop software holisitically. In ProgramBench, given only a program and its documentation, agents must architect and implement a codebase that matches the reference executable's behavior. End-to-end behavioral tests are generated via agent-driven fuzzing, enabling evaluation without prescribing implementation structure. Our 200 tasks range from compact CLI tools to widely used software such as FFmpeg, SQLite, and the PHP interpreter. We evaluate 9 LMs and find that none fully resolve any task, with the best model passing 95% of tests on only 3% of tasks. Models favor monolithic, single-file implementations that diverge sharply from human-written code.

Abstract of ProgramBench: Can Language Models Rebuild Programs From Scratch? (arXiv:2605.03546, CC0)
Made by Meta Superintelligence Labs , Stanford University and Harvard University Released May 2026 Building whole projects All tasks public Last in a launch post Oct 2026 Last model added Oct 2026 Scores refreshed 2 Nov 2026
Level
Research frontier Researchers, on open problems or tasks that take days
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Code A program or patch run against tests.
facebookresearch/ProgramBench 978 stars Paper Dataset Leaderboard programbench.com
25
Models scored on its own leaderboard
13
Launch posts list it, from 8 labs
14
System cards report it, from 7 labs

Claude Opus 5.5 leads with 18.5%

Best score per model from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

25 models. Scroll sideways for the rest.

  1. 18.5 Claude Opus 5.5
  2. 9.00 Claude Fable 5.1
  3. 8.50 Muse Wonder Full
  4. 5.50 Muse Wonder
  5. 5.00 GPT-6 Astra
  6. 4.50 Claude Opus 5
  7. 2.50 Muse Spark 1.3
  8. 1.00 GPT-5.6 Sol
  9. 0.50 GPT-5.5
  10. 0.50 Gemini 3.6 Flash
  11. 0.0 GPT-5 mini
  12. 0.0 Claude Haiku 4.5
  13. 0.0 Gemini 3 Flash
  14. 0.0 Claude Opus 4.6
  15. 0.0 Claude Sonnet 4.6
  16. 0.0 Gemini 3.1 Pro
  17. 0.0 GPT-5.4
  18. 0.0 GPT-5.4 Mini
  19. 0.0 Claude Opus 4.7
  20. 0.0 Gemini 3.5 Flash
  21. 0.0 Claude Opus 4.8
  22. 0.0 GLM-5.2
  23. 0.0 Muse Spark 1.1
  24. 0.0 Muse Spark 1.2
  25. 0.0 Gemini 3.7 Flash
  1. 1 Claude Opus 5.5 xhigh 18.5%
  2. 2 Claude Fable 5.1 xhigh 9.00%
  3. 3 Muse Wonder Full 8.50%
  4. 4 Muse Wonder xhigh 5.50%
  5. 5 GPT-6 Astra max 5.00%
  6. 6 Claude Opus 5 xhigh 4.50%
  7. 7 Muse Spark 1.3 max 2.50%
  8. 8 GPT-5.6 Sol xhigh 1.00%
  9. 9 GPT-5.5 xhigh 0.50%
  10. 10 Gemini 3.6 Flash 0.50%
  11. 11 GPT-5 mini 0%
  12. 12 Claude Haiku 4.5 0%
  13. 13 Gemini 3 Flash 0%
  14. 14 Claude Opus 4.6 0%
  15. 15 Claude Sonnet 4.6 0%
  16. 16 Gemini 3.1 Pro 0%
  17. 17 GPT-5.4 0%
  18. 18 GPT-5.4 Mini 0%
  19. 19 Claude Opus 4.7 xhigh 0%
  20. 20 Gemini 3.5 Flash 0%
  21. 21 Claude Opus 4.8 xhigh 0%
  22. 22 GLM-5.2 0%
  23. 23 Muse Spark 1.1 xhigh 0%
  24. 24 Muse Spark 1.2 xhigh 0%
  25. 25 Gemini 3.7 Flash 0%
#ModelScoreSettingDateSource
1 Claude Opus 5.5 18.5% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
2 Claude Fable 5.1 9.00% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
3 Muse Wonder Full 8.50% mini-SWE-agent Oct 2026 Official leaderboard
4 Muse Wonder 5.50% xhigh, mini-SWE-agent Oct 2026 Official leaderboard
5 GPT-6 Astra 5.00% max, mini-SWE-agent Sep 2026 Official leaderboard
6 Claude Opus 5 4.50% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
7 Muse Spark 1.3 2.50% max, mini-SWE-agent Sep 2026 Official leaderboard
8 GPT-5.6 Sol 1.00% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
9 GPT-5.5 0.50% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
10 Gemini 3.6 Flash 0.50% mini-SWE-agent Jul 2026 Official leaderboard
11 GPT-5 mini 0% mini-SWE-agent Aug 2025 Official leaderboard
12 Claude Haiku 4.5 0% mini-SWE-agent Oct 2025 Official leaderboard
13 Gemini 3 Flash 0% mini-SWE-agent Dec 2025 Official leaderboard
14 Claude Opus 4.6 0% mini-SWE-agent Feb 2026 Official leaderboard
15 Claude Sonnet 4.6 0% mini-SWE-agent Feb 2026 Official leaderboard
16 Gemini 3.1 Pro 0% mini-SWE-agent Feb 2026 Official leaderboard
17 GPT-5.4 0% mini-SWE-agent Mar 2026 Official leaderboard
18 GPT-5.4 Mini 0% mini-SWE-agent Mar 2026 Official leaderboard
19 Claude Opus 4.7 0% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
20 Gemini 3.5 Flash 0% mini-SWE-agent May 2026 Official leaderboard
21 Claude Opus 4.8 0% xhigh, mini-SWE-agent May 2026 Official leaderboard
22 GLM-5.2 0% mini-SWE-agent Jun 2026 Official leaderboard
23 Muse Spark 1.1 0% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
24 Muse Spark 1.2 0% xhigh, mini-SWE-agent Aug 2026 Official leaderboard
25 Gemini 3.7 Flash 0% mini-SWE-agent Aug 2026 Official leaderboard

Labs report scores for 20 models in their cards. We plot every source in the charts below.

The best score, 18.5%, costs $36 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 Meta Anthropic Cost frontier One model at each effort level
0% 5% 10% 15% 20% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Claude Opus 5.5, xhigh GPT-6 Astra, max Muse Spark 1.3, max Muse Spark 1.3, xhigh GPT-5 mini 0% 5% 10% 15% 20% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Claude Opus 5.5, xhigh GPT-6 Astra, max
ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 5.5 xhigh mini-SWE-agent 18.5% $36 434,511
Claude Fable 5.1 xhigh mini-SWE-agent 9.00% $57 445,587
Muse Wonder Full mini-SWE-agent 8.50% 352,279
Muse Wonder xhigh mini-SWE-agent 5.50% $37 345,971
GPT-6 Astra max mini-SWE-agent 5.00% $11 91,595
Claude Opus 5 xhigh mini-SWE-agent 4.50% $51 357,785
Muse Spark 1.3 max mini-SWE-agent 2.50% $6.46 387,723
GPT-5.6 Sol xhigh mini-SWE-agent 1.00% $6.08 56,179
Muse Spark 1.3 xhigh mini-SWE-agent 1.00% $2.02 216,872
GPT-5.5 xhigh mini-SWE-agent 0.50% $8.85 51,624
GPT-5.5 high mini-SWE-agent 0.50% $3.65 27,379
GPT-5.6 Sol mini-SWE-agent 0.50% $1.00 13,084
Gemini 3.6 Flash mini-SWE-agent 0.50% $4.83 107,792
GPT-5 mini mini-SWE-agent 0% $0.03 7,146
Claude Haiku 4.5 mini-SWE-agent 0% $0.80 47,506
Gemini 3 Flash mini-SWE-agent 0% $0.30 22,561
Claude Opus 4.6 mini-SWE-agent 0% $11 89,147
Claude Sonnet 4.6 mini-SWE-agent 0% $27 190,227
Gemini 3.1 Pro mini-SWE-agent 0% $1.51 18,071
GPT-5.4 mini-SWE-agent 0% $0.33 7,408
GPT-5.4 Mini mini-SWE-agent 0% $0.04 3,147
Claude Opus 4.7 xhigh mini-SWE-agent 0% $11 97,205
Claude Opus 4.7 mini-SWE-agent 0% $3.81 41,669
GPT-5.5 mini-SWE-agent 0% $1.21 11,928
Gemini 3.5 Flash mini-SWE-agent 0% $5.60 132,679
Claude Opus 4.8 xhigh mini-SWE-agent 0% $21 270,298
GLM-5.2 mini-SWE-agent 0% $25 237,624
Muse Spark 1.1 xhigh mini-SWE-agent 0% $0.73 357,115
Muse Spark 1.2 xhigh mini-SWE-agent 0% $1.48 329,742
Gemini 3.7 Flash mini-SWE-agent 0% $2.04 74,885

5 models have held number one since its release in May 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.

OpenAI Anthropic
Own leaderboard 5 leaders
Claude Opus 5.5 18.5%
Lab claims Passed 85% in Jul 2026, by Claude Opus 5
Claude Opus 5 93%

The best score passed 85% in Jul 2026

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 Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Claude Opus 5 Claude Opus 4.8 GPT-5.5 GPT-5 mini 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Claude Opus 5 Claude Opus 4.8

8 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Moonshot AI
Z.ai
ByteDance Seed
Tencent
DeepSeek
StepFun
Xiaomi Corp
Meta
DateLaunch postScoreAs the post writes it
24 Oct 2026 Kimi K3.1 Pro 31.5 Program Bench (binary score, >=95% counts as solved)
17 Oct 2026 DeepSeek V4.1 Flash 1017 29.1 ProgramBench
12 Oct 2026 Muse Wonder 50 ProgramBench
12 Oct 2026 Muse Wonder Full 59 ProgramBench
22 Sep 2026 MiMo-V2.6 Pro 26.5 ProgramBench
20 Sep 2026 Step 5 Preview 80.5 ProgramBench (Pass Rate)
10 Sep 2026 DeepSeek V4.1 Flash 20.3 ProgramBench (Almost@1)
28 Aug 2026 Tencent Hy4 preview 17.5 ProgramBench
14 Aug 2026 GLM-5.3 19 ProgramBench (Almost Solved)
16 Jul 2026 Kimi K3 77.8 ProgramBench
23 Jun 2026 Seed2.1 Pro 50.3 ProgramBench (resolved / almost resolved / average pass rate) 0/1/50.3
16 Jun 2026 GLM-5.2 63.7 ProgramBench
12 Jun 2026 Kimi K2.7 Code 53.6 Program Bench

14 system cards report it, from 7 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
12 Oct 2026 Muse Wonder model card 50 ProgramBench, rollouts passing at least 95% of tests, mini-swe-agent, Muse Wonder (max) (capabilities table image, page 3) Capabilities
12 Oct 2026 Muse Wonder Full model card 59 ProgramBench, rollouts passing at least 95% of tests, mini-swe-agent, Muse Wonder Full (max) (capabilities table image, page 3) Capabilities
7 Oct 2026 Claude Haiku 5.5 system card 82 ProgramBench, hidden test pass rate on 166 of 200 tasks, mini-swe-agent without time limit (8.7.1) Capabilities
28 Sep 2026 Claude Sonnet 5.5 system card 79.7 ProgramBench hidden test pass rate on 166 tasks, single agent, mini-swe-agent harness (8.10.1) Capabilities
22 Sep 2026 MiMo-V2.6 Pro technical report 26.5 ProgramBench (Table 3) Capabilities
22 Sep 2026 Claude Opus 5.5 system card 91.2 ProgramBench hidden test pass rate on 166 of 200 tasks, single agent, mini-swe-agent harness, no time limit (8.10.1) Capabilities
10 Sep 2026 DeepSeek V4.1 Flash technical report 20.3 ProgramBench (Almost@1) (Table 3). Figure 10 text: 30.04 multi-agent and 20.39 single-agent at 8h on a 172-task high-confidence subset Capabilities
1 Sep 2026 Claude Fable 5.1 system card 87.6 ProgramBench hidden test pass rate on 166 tasks, single agent, mini-swe-agent harness, no time limit (8.11.1) Capabilities
27 Jul 2026 Kimi K3 technical report 77.8 ProgramBench (Table 2) Capabilities
24 Jul 2026 Claude Opus 5 system card 93 ProgramBench hidden test pass rate on 166 of 200 tasks after five 1M-token episodes, single agent (83% after the first episode) (8.9.1) Capabilities
14 Jul 2026 Grok 4.5 model card 57.2 ProgramBench, tests-passed rate, high (2.8) Capabilities
30 Jun 2026 Claude Sonnet 5 system card ProgramBench, hidden test pass rate on 166 of 200 tasks across 5 episodes, given only as a range 76 to 86%, no single value (8.8) Capabilities
23 Jun 2026 Seed2.1 Pro model card 50.3 ProgramBench, resolved 0 / almost resolved 1 / average pass rate 50.3 (Table 5) Capabilities
28 May 2026 Claude Opus 4.8 system card 79 ProgramBench hidden test pass rate on 166 of 200 tasks, one episode of up to 1M tokens, near-final snapshot. Text gives 79 to 88% across 1 to 5 episodes (8.5, Figure 8.5.A) Capabilities

More in Coding

All 243 in Coding
  • SWE-bench Verified In use 500 human-validated GitHub issues from 12 Python repos. Agent writes a patch that must pass hidden tests. Claude Opus 4.7 83.5%
  • SWE-bench Multilingual In use 300 GitHub-issue tasks across nine programming languages in SWE-bench format. Patch graded by repository tests. Claude Opus 5.5 93.9%
  • LiveCodeBench v6 In use Competitive programming problems collected after model cutoffs from LeetCode, AtCoder, Codeforces. Scored pass@1 on hidden tests. DeepSeek-V4-Pro 93.5%
  • DeepSWE In use 113 original long-horizon engineering tasks across 91 repos and five languages, written from scratch. Hand-written verifiers accept alternative solutions. GPT-6 Astra 74.1%
  • SciCode In use Scientist-curated research coding problems split into subproblems across natural sciences. Code checked against test cases. Seed2.1 Pro 59.8%
  • Codeforces rating In use Model solves recent Codeforces contest problems. An Elo-style rating is computed from simulated contest performance. DeepSeek V4.1 Flash 3,471