Benchmark Almanac

FrontierSWE

GPT-6 Astra 65.5% leads since Sep 2026 In use

34 ultra-long-horizon engineering and research tasks. 5 trials of up to 20 hours each. Score is mean task score 0-100.

This record is FrontierSWE v2 (2026-09-02), the version Epoch AI tracks. It added 21 tasks to reach 34 and runs every model in the proximus harness. v1 (April 2026, 17 tasks) has its own record.

Made by Proximal Released Sep 2026 Building whole projects All tasks public Last in a launch post Oct 2026 Last model added Oct 2026 Scores refreshed 31 Oct 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.
Proximal-Labs/frontier-swe-v2 51 stars Paper Leaderboard
21
Models scored in Epoch AI data
3
Launch posts list it, from 2 labs
7
System cards report it, from 4 labs

GPT-6 Astra leads with 65.5%

Best score per model in Epoch AI’s data. Where a model ran with several settings or harnesses, we keep its best.

21 models. Scroll sideways for the rest.

  1. 65.5 GPT-6 Astra
  2. 62.3 Claude Opus 5.5
  3. 61.9 Claude Sonnet 5.5
  4. 56.3 Claude Fable 5.1
  5. 55.0 Gemini 4 Argon
  6. 52.0 Claude Opus 5
  7. 47.0 Claude Fable 5
  8. 43.8 Claude Haiku 5.5
  9. 32.2 GPT-5.6 Sol
  10. 30.2 GLM-5.3
  11. 29.5 Grok 4.7
  12. 25.9 Kimi K3
  13. 25.8 Muse Spark 1.3
  14. 25.3 Grok 4.6
  15. 20.3 Gemini 3.7 Flash
  16. 19.6 Gemini 3.8 Flash
  17. 18.1 GLM-5.3-Flash
  18. 17.8 Qwen3.8 Max (0902)
  19. 15.8 Qwen 3.8 Max
  20. 12.0 Muse Spark 1.2
  21. 4.12 Inkling
  1. 1 GPT-6 Astra max 65.5%
  2. 2 Claude Opus 5.5 max 62.3%
  3. 3 Claude Sonnet 5.5 max 61.9%
  4. 4 Claude Fable 5.1 max 56.3%
  5. 5 Gemini 4 Argon high 55%
  6. 6 Claude Opus 5 max 52%
  7. 7 Claude Fable 5 max 47%
  8. 8 Claude Haiku 5.5 max 43.8%
  9. 9 GPT-5.6 Sol max 32.2%
  10. 10 GLM-5.3 max 30.2%
  11. 11 Grok 4.7 xhigh 29.5%
  12. 12 Kimi K3 max 25.9%
  13. 13 Muse Spark 1.3 max 25.8%
  14. 14 Grok 4.6 xhigh 25.3%
  15. 15 Gemini 3.7 Flash high 20.3%
  16. 16 Gemini 3.8 Flash high 19.6%
  17. 17 GLM-5.3-Flash max 18.1%
  18. 18 Qwen3.8 Max (0902) xhigh 17.8%
  19. 19 Qwen 3.8 Max xhigh 15.8%
  20. 20 Muse Spark 1.2 xhigh 12%
  21. 21 Inkling xhigh 4.12%
#ModelScoreSettingDateSource
1 GPT-6 Astra 65.5% max, proximus Sep 2026 Epoch AI
2 Claude Opus 5.5 62.3% max, proximus Sep 2026 Epoch AI
3 Claude Sonnet 5.5 61.9% max, proximus Sep 2026 Epoch AI
4 Claude Fable 5.1 56.3% max, proximus Sep 2026 Epoch AI
5 Gemini 4 Argon 55% high, proximus Sep 2026 Epoch AI
6 Claude Opus 5 52% max, proximus Jul 2026 Epoch AI
7 Claude Fable 5 47% max, proximus Jun 2026 Epoch AI
8 Claude Haiku 5.5 43.8% max, proximus Oct 2026 Epoch AI
9 GPT-5.6 Sol 32.2% max, proximus Jul 2026 Epoch AI
10 GLM-5.3 30.2% max, proximus Aug 2026 Epoch AI
11 Grok 4.7 29.5% xhigh, proximus Sep 2026 Epoch AI
12 Kimi K3 25.9% max, proximus Jul 2026 Epoch AI
13 Muse Spark 1.3 25.8% max, proximus Sep 2026 Epoch AI
14 Grok 4.6 25.3% xhigh, proximus Aug 2026 Epoch AI
15 Gemini 3.7 Flash 20.3% high, proximus Aug 2026 Epoch AI
16 Gemini 3.8 Flash 19.6% high, proximus Sep 2026 Epoch AI
17 GLM-5.3-Flash 18.1% max, proximus Aug 2026 Epoch AI
18 Qwen3.8 Max (0902) 17.8% xhigh, proximus Sep 2026 Epoch AI
19 Qwen 3.8 Max 15.8% xhigh, proximus Aug 2026 Epoch AI
20 Muse Spark 1.2 12% xhigh, proximus Aug 2026 Epoch AI
21 Inkling 4.12% xhigh, proximus Jul 2026 Epoch AI

Labs report scores for 9 models in their cards. We plot every source in the charts below.

2 models have held number one since its release in Sep 2026

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

Anthropic OpenAI Ant Group
Epoch AI runs 2 leaders
GPT-6 Astra 65.5%
Lab claims 3 leaders
Ling-3.1-flash 75.2%

Scores over time

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

Run or collected by Epoch AI Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Ling-3.1-flash GPT-6 Astra Claude Fable 5.1 Claude Opus 5 Claude Fable 5 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Ling-3.1-flash GPT-6 Astra

2 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.

Google DeepMind
Ant Group
DateLaunch postScoreAs the post writes it
28 Oct 2026 Gemini 4 Carbon 61.2 FrontierSWE 2
13 Oct 2026 Ling-3.1-flash 75.2 FrontierSWE
13 Oct 2026 Gemini 4 Argon 55 FrontierSWE v2

7 system cards report it, from 4 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
13 Oct 2026 Gemini 4 Argon model card 55 FrontierSWE v2 (results table, p.5) Capabilities
7 Oct 2026 Claude Haiku 5.5 system card 43.8 FrontierSWE v2, max effort, Proximal harness, five trials per task (8.6) Capabilities
28 Sep 2026 Claude Sonnet 5.5 system card 61.9 FrontierSWE v2, mean of five trials, Proximal harness at max effort (8.7) Capabilities
22 Sep 2026 Claude Opus 5.5 system card 62.3 FrontierSWE v2, mean of five trials per task, Proximal harness at max effort (8.7) Capabilities
21 Sep 2026 Grok 4.7 model card 29 FrontierSWE V2, mean@5 partial-credit score, xhigh, run by Proximal Labs (2.4) Capabilities
10 Sep 2026 DeepSeek V4.1 Flash technical report 32.9 FrontierSWE v2, no-GPU public subset, Mean@5, multi-agent Agent Team mode at 20h deadline (Figure 10 text). Single-agent 28.2 Capabilities
1 Sep 2026 Claude Fable 5.1 system card 57 FrontierSWE v2, mean score 0.57 on a 0 to 1 scale, max effort, Proximal harness, 5 trials per task (8.5) Capabilities

Versions

Replaces FrontierSWE v1

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