Benchmark Almanac

ProgramBench almost solved

Claude Opus 5.5 70.5% leads since September 2026
Claude Opus 5.5 70.5% leads since September 2026
In use

Agents rebuild 200 programs from only their compiled binaries and docs. The score is the share of tasks whose rebuilt program passes at least 95% of the hidden behavioral tests.

The board shows this as its Almost column and ranks by fully resolved tasks, which the ProgramBench record holds. Many labs report this number.

Released
May 2026
Kind
Building whole projects
Level
Research frontier
Scored
Checked automatically
Answers
Code
Tasks
All public
Strongest model tested
GPT-6 Astra #3 on the capability index
Last in a launch post
November 2026
Last model added
November 2026
Scores refreshed
November 12th, 2026

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)
25
Models scored on its own leaderboard
6
Launch posts list it, from 6 labs
3
System cards report it, from 2 labs

Claude Opus 5.5 leads with 70.5%

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.

Claude Opus 5.5 leads with 70.5%, claims included

Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.

25 models. Scroll sideways for the rest.

  1. 70.5 Claude Opus 5.5
  2. 60.0 Muse Wonder Full
  3. 55.5 Claude Fable 5.1
  4. 52.5 GPT-6 Astra
  5. 50.5 Muse Wonder
  6. 37.0 Claude Opus 5
  7. 25.0 Muse Spark 1.3
  8. 16.5 Claude Opus 4.8
  9. 15.5 GPT-5.6 Sol
  10. 13.5 GPT-5.5
  11. 8.50 GLM-5.2
  12. 5.50 Gemini 3.7 Flash
  13. 4.50 Claude Opus 4.7
  14. 4.50 Muse Spark 1.2
  15. 4.00 Muse Spark 1.1
  16. 4.00 Gemini 3.6 Flash
  17. 3.00 Gemini 3.5 Flash
  18. 2.50 Claude Opus 4.6
  19. 1.00 Claude Sonnet 4.6
  20. 0.0 GPT-5 mini
  21. 0.0 Claude Haiku 4.5
  22. 0.0 Gemini 3 Flash
  23. 0.0 Gemini 3.1 Pro
  24. 0.0 GPT-5.4
  25. 0.0 GPT-5.4 Mini
  1. 1 Claude Opus 5.5 xhigh 70.5%
  2. 2 Muse Wonder Full 60%
  3. 3 Claude Fable 5.1 xhigh 55.5%
  4. 4 GPT-6 Astra max 52.5%
  5. 5 Muse Wonder xhigh 50.5%
  6. 6 Claude Opus 5 xhigh 37%
  7. 7 Muse Spark 1.3 max 25%
  8. 8 Claude Opus 4.8 xhigh 16.5%
  9. 9 GPT-5.6 Sol xhigh 15.5%
  10. 10 GPT-5.5 xhigh 13.5%
  11. 11 GLM-5.2 8.50%
  12. 12 Gemini 3.7 Flash 5.50%
  13. 13 Claude Opus 4.7 xhigh 4.50%
  14. 14 Muse Spark 1.2 xhigh 4.50%
  15. 15 Muse Spark 1.1 xhigh 4.00%
  16. 16 Gemini 3.6 Flash 4.00%
  17. 17 Gemini 3.5 Flash 3.00%
  18. 18 Claude Opus 4.6 2.50%
  19. 19 Claude Sonnet 4.6 1.00%
  20. 20 GPT-5 mini 0%
  21. 21 Claude Haiku 4.5 0%
  22. 22 Gemini 3 Flash 0%
  23. 23 Gemini 3.1 Pro 0%
  24. 24 GPT-5.4 0%
  25. 25 GPT-5.4 Mini 0%

Source: Official ProgramBench almost solved leaderboard

31 models. Scroll sideways for the rest.

  1. 70.5 Claude Opus 5.5
  2. 60.0 Muse Wonder Full
  3. 55.5 Claude Fable 5.1
  4. 52.5 GPT-6 Astra
  5. 50.5 Muse Wonder
  6. 37.0 Claude Opus 5
  7. 31.5 Kimi K3.1 Pro
  8. 30.0 Tencent Hy4
  9. 28.0 Qwen3.8 Max (0902)
  10. 25.5 LongCat-2.5
  11. 25.0 Muse Spark 1.3
  12. 20.3 DeepSeek V4.1 Flash
  13. 19.0 GLM-5.3
  14. 16.5 Claude Opus 4.8
  15. 15.5 GPT-5.6 Sol
  16. 13.5 GPT-5.5
  17. 8.50 GLM-5.2
  18. 5.50 Gemini 3.7 Flash
  19. 4.50 Claude Opus 4.7
  20. 4.50 Muse Spark 1.2
  21. 4.00 Muse Spark 1.1
  22. 4.00 Gemini 3.6 Flash
  23. 3.00 Gemini 3.5 Flash
  24. 2.50 Claude Opus 4.6
  25. 1.00 Claude Sonnet 4.6
  26. 0.0 GPT-5 mini
  27. 0.0 Claude Haiku 4.5
  28. 0.0 Gemini 3 Flash
  29. 0.0 Gemini 3.1 Pro
  30. 0.0 GPT-5.4
  31. 0.0 GPT-5.4 Mini
  1. 1 Claude Opus 5.5 xhigh 70.5%
  2. 2 Muse Wonder Full 60%
  3. 3 Claude Fable 5.1 xhigh 55.5%
  4. 4 GPT-6 Astra max 52.5%
  5. 5 Muse Wonder xhigh 50.5%
  6. 6 Claude Opus 5 xhigh 37%
  7. 7 Kimi K3.1 Pro 31.5%
  8. 8 Tencent Hy4 30%
  9. 9 Qwen3.8 Max (0902) 28%
  10. 10 LongCat-2.5 25.5%
  11. 11 Muse Spark 1.3 max 25%
  12. 12 DeepSeek V4.1 Flash 20.3%
  13. 13 GLM-5.3 19%
  14. 14 Claude Opus 4.8 xhigh 16.5%
  15. 15 GPT-5.6 Sol xhigh 15.5%
  16. 16 GPT-5.5 xhigh 13.5%
  17. 17 GLM-5.2 8.50%
  18. 18 Gemini 3.7 Flash 5.50%
  19. 19 Claude Opus 4.7 xhigh 4.50%
  20. 20 Muse Spark 1.2 xhigh 4.50%
  21. 21 Muse Spark 1.1 xhigh 4.00%
  22. 22 Gemini 3.6 Flash 4.00%
  23. 23 Gemini 3.5 Flash 3.00%
  24. 24 Claude Opus 4.6 2.50%
  25. 25 Claude Sonnet 4.6 1.00%
  26. 26 GPT-5 mini 0%
  27. 27 Claude Haiku 4.5 0%
  28. 28 Gemini 3 Flash 0%
  29. 29 Gemini 3.1 Pro 0%
  30. 30 GPT-5.4 0%
  31. 31 GPT-5.4 Mini 0%

Sources: Official ProgramBench almost solved leaderboard and labs’ model cards

#ModelScoreSettingDateSource
1 Claude Opus 5.5 70.5% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
2 Muse Wonder Full 60% mini-SWE-agent Oct 2026 Official leaderboard
3 Claude Fable 5.1 55.5% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
4 GPT-6 Astra 52.5% max, mini-SWE-agent Sep 2026 Official leaderboard
5 Muse Wonder 50.5% xhigh, mini-SWE-agent Oct 2026 Official leaderboard
6 Claude Opus 5 37% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
7 Muse Spark 1.3 25% max, mini-SWE-agent Sep 2026 Official leaderboard
8 Claude Opus 4.8 16.5% xhigh, mini-SWE-agent May 2026 Official leaderboard
9 GPT-5.6 Sol 15.5% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
10 GPT-5.5 13.5% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
11 GLM-5.2 8.50% mini-SWE-agent Jun 2026 Official leaderboard
12 Gemini 3.7 Flash 5.50% mini-SWE-agent Aug 2026 Official leaderboard
13 Claude Opus 4.7 4.50% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
14 Muse Spark 1.2 4.50% xhigh, mini-SWE-agent Aug 2026 Official leaderboard
15 Muse Spark 1.1 4.00% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
16 Gemini 3.6 Flash 4.00% mini-SWE-agent Jul 2026 Official leaderboard
17 Gemini 3.5 Flash 3.00% mini-SWE-agent May 2026 Official leaderboard
18 Claude Opus 4.6 2.50% mini-SWE-agent Feb 2026 Official leaderboard
19 Claude Sonnet 4.6 1.00% mini-SWE-agent Feb 2026 Official leaderboard
20 GPT-5 mini 0% mini-SWE-agent Aug 2025 Official leaderboard
21 Claude Haiku 4.5 0% mini-SWE-agent Oct 2025 Official leaderboard
22 Gemini 3 Flash 0% mini-SWE-agent Dec 2025 Official leaderboard
23 Gemini 3.1 Pro 0% mini-SWE-agent Feb 2026 Official leaderboard
24 GPT-5.4 0% mini-SWE-agent Mar 2026 Official leaderboard
25 GPT-5.4 Mini 0% mini-SWE-agent Mar 2026 Official leaderboard
#ModelScoreSettingDateSource
1 Claude Opus 5.5 70.5% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
2 Muse Wonder Full 60% mini-SWE-agent Oct 2026 Official leaderboard
3 Claude Fable 5.1 55.5% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
4 GPT-6 Astra 52.5% max, mini-SWE-agent Sep 2026 Official leaderboard
5 Muse Wonder 50.5% xhigh, mini-SWE-agent Oct 2026 Official leaderboard
6 Claude Opus 5 37% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
7 Kimi K3.1 Pro 31.5% Oct 2026 Lab model card
8 Tencent Hy4 30% Nov 2026 Lab model card
9 Qwen3.8 Max (0902) 28% Sep 2026 Lab model card
10 LongCat-2.5 25.5% Nov 2026 Lab model card
11 Muse Spark 1.3 25% max, mini-SWE-agent Sep 2026 Official leaderboard
12 DeepSeek V4.1 Flash 20.3% Sep 2026 Lab model card
13 GLM-5.3 19% Aug 2026 Lab model card
14 Claude Opus 4.8 16.5% xhigh, mini-SWE-agent May 2026 Official leaderboard
15 GPT-5.6 Sol 15.5% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
16 GPT-5.5 13.5% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
17 GLM-5.2 8.50% mini-SWE-agent Jun 2026 Official leaderboard
18 Gemini 3.7 Flash 5.50% mini-SWE-agent Aug 2026 Official leaderboard
19 Claude Opus 4.7 4.50% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
20 Muse Spark 1.2 4.50% xhigh, mini-SWE-agent Aug 2026 Official leaderboard
21 Muse Spark 1.1 4.00% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
22 Gemini 3.6 Flash 4.00% mini-SWE-agent Jul 2026 Official leaderboard
23 Gemini 3.5 Flash 3.00% mini-SWE-agent May 2026 Official leaderboard
24 Claude Opus 4.6 2.50% mini-SWE-agent Feb 2026 Official leaderboard
25 Claude Sonnet 4.6 1.00% mini-SWE-agent Feb 2026 Official leaderboard
26 GPT-5 mini 0% mini-SWE-agent Aug 2025 Official leaderboard
27 Claude Haiku 4.5 0% mini-SWE-agent Oct 2025 Official leaderboard
28 Gemini 3 Flash 0% mini-SWE-agent Dec 2025 Official leaderboard
29 Gemini 3.1 Pro 0% mini-SWE-agent Feb 2026 Official leaderboard
30 GPT-5.4 0% mini-SWE-agent Mar 2026 Official leaderboard
31 GPT-5.4 Mini 0% mini-SWE-agent Mar 2026 Official leaderboard

Labs report scores for 8 models in their cards. Switch the scores above to rank them together.

The best score, 70.5%, 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% 20% 40% 60% 80% 100% $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 Muse Spark 1.2, xhigh GPT-5 mini 0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Claude Opus 5.5, xhigh GPT-6 Astra, max

Source: Official ProgramBench almost solved leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 5.5 xhigh mini-SWE-agent 70.5% $36 434,511
Muse Wonder Full mini-SWE-agent 60% 352,279
Claude Fable 5.1 xhigh mini-SWE-agent 55.5% $57 445,587
GPT-6 Astra max mini-SWE-agent 52.5% $11 91,595
Muse Wonder xhigh mini-SWE-agent 50.5% $37 345,971
Claude Opus 5 xhigh mini-SWE-agent 37% $51 357,785
Muse Spark 1.3 max mini-SWE-agent 25% $6.46 387,723
Claude Opus 4.8 xhigh mini-SWE-agent 16.5% $21 270,298
Muse Spark 1.3 xhigh mini-SWE-agent 16.5% $2.02 216,872
GPT-5.6 Sol xhigh mini-SWE-agent 15.5% $6.08 56,179
GPT-5.5 xhigh mini-SWE-agent 13.5% $8.85 51,624
GLM-5.2 mini-SWE-agent 8.50% $25 237,624
Gemini 3.7 Flash mini-SWE-agent 5.50% $2.04 74,885
GPT-5.5 high mini-SWE-agent 5.00% $3.65 27,379
Claude Opus 4.7 xhigh mini-SWE-agent 4.50% $11 97,205
Muse Spark 1.2 xhigh mini-SWE-agent 4.50% $1.48 329,742
Muse Spark 1.1 xhigh mini-SWE-agent 4.00% $0.73 357,115
Gemini 3.6 Flash mini-SWE-agent 4.00% $4.83 107,792
Claude Opus 4.7 mini-SWE-agent 3.00% $3.81 41,669
Gemini 3.5 Flash mini-SWE-agent 3.00% $5.60 132,679
Claude Opus 4.6 mini-SWE-agent 2.50% $11 89,147
GPT-5.6 Sol mini-SWE-agent 2.50% $1.00 13,084
GPT-5.5 mini-SWE-agent 1.50% $1.21 11,928
Claude Sonnet 4.6 mini-SWE-agent 1.00% $27 190,227
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
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

5 models have held number one on the official leaderboard since its release in May 2026 5 models have held number one with claims included 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 Z.ai Qwen Meta
Own leaderboard 5 leaders
Claude Opus 5.5 70.5%
Lab claims 3 leaders
Muse Wonder Full 59%

Sources: Official ProgramBench almost solved leaderboard and labs’ model cards

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 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 5.5 Claude Fable 5.1 Claude Opus 5 Claude Opus 4.8 Claude Opus 4.7 GPT-5 mini 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Claude Opus 5.5 Claude Fable 5.1

Sources: Official ProgramBench almost solved leaderboard and labs’ model cards

6 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.

Z.ai
Qwen
DeepSeek
Moonshot AI
Meituan
Tencent
DateLaunch postScoreAs the post writes it
Nov 4, 2026 Tencent Hy4 30 ProgramBench (almost solved) *
Nov 2, 2026 LongCat-2.5 25.5 Program Bench (Almost Resolved)
Oct 24, 2026 Kimi K3.1 Pro 31.5 Program Bench (binary score, >=95% counts as solved)
Sep 10, 2026 DeepSeek V4.1 Flash 20.3 ProgramBench (Almost@1)
Sep 2, 2026 Qwen3.8 Max (0902) 28 Black-box software replication, ProgramBench (Almost Solved)
Aug 14, 2026 GLM-5.3 19 ProgramBench (Almost Solved)

3 system cards report it, from 2 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
Oct 12, 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
Oct 12, 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
Sep 10, 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

The lab's number comes out ahead in 0 of 2 pairs that differ

Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. Every benchmark.

Model Lab says Board says Gap
Muse Wonder Full 59% ProgramBench, rollouts passing at least 95% of tests, mini-swe-agent, Muse Wonder Full (max) (capabilities table image, page 3) 60% Leaderboard, mini-SWE-agent −1.0 pts
Muse Wonder 50% ProgramBench, rollouts passing at least 95% of tests, mini-swe-agent, Muse Wonder (max) (capabilities table image, page 3) 50.5% Leaderboard, xhigh, mini-SWE-agent −0.5 pts

More in Coding

All 282 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%
  • 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%
  • SWE-bench Multilingual In use 300 GitHub-issue tasks across nine programming languages in SWE-bench format. Patch graded by repository tests. Claude Fable 5.2 96.6%
  • 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