Benchmark Almanac

Retrobench v0.4

Gemini 4 Carbon 99.9% leads since October 2026
Not in model cards

Agents reconstruct C from five Nintendo 64 assembly targets. The organizer scores the average percentage of matching bytes under a budget of 25 compilations per target.

The current leaderboard ranks 15 models. The maker reports achievable ceilings below 100 percent on targets affected by alignment padding.

Released
August 2026
Kind
Function-level problems
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
Kept private
Measures
Capability
Language
English
In and out
Text, Code → Code, Actions
Interaction
Agentic
Environment
Code sandbox
Items from
Real work and use
Run by
A host
In views
Agents
Strongest model tested
Claude Fable 5.2 #2 on the capability index
Last model added
November 2026
Scores refreshed
November 20th, 2026
15
Models scored on its own leaderboard

2 models share the top score of 99.9%

Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best.

  1. 1 Gemini 4 Carbon 99.9%
  2. 2 Claude Fable 5.2 99.9%
  3. 3 Claude Opus 5.5 98.4%
  4. 4 GPT-6.1 Sol 94.9%
  5. 5 GPT-6 Astra 94.3%
  6. 6 Gemini 4 Argon 84.7%
  7. 7 Claude Sonnet 5.5 84.2%
  8. 8 GLM-5.3 80.2%
  9. 9 Kimi K3 1M 79.5%
  10. 10 DeepSeek V4.1 Flash 1017 75.1%
  11. 11 Gemini 3.8 Flash 73.7%
  12. 12 MiMo-V2.6 Pro 72.1%
  13. 13 GLM-5.3-Flash 66.7%
  14. 14 GPT-6 Luna 60.4%
  15. 15 Claude Haiku 5.5 56.9%

Source: Official Retrobench v0.4 leaderboard

#ModelScoreSettingDateSource
1 Gemini 4 Carbon 99.9% Oct 2026 Official leaderboard
2 Claude Fable 5.2 99.9% Nov 2026 Official leaderboard
3 Claude Opus 5.5 98.4% Sep 2026 Official leaderboard
4 GPT-6.1 Sol 94.9% Sep 2026 Official leaderboard
5 GPT-6 Astra 94.3% Sep 2026 Official leaderboard
6 Gemini 4 Argon 84.7% Sep 2026 Official leaderboard
7 Claude Sonnet 5.5 84.2% Sep 2026 Official leaderboard
8 GLM-5.3 80.2% Aug 2026 Official leaderboard
9 Kimi K3 79.5% 1M Jul 2026 Official leaderboard
10 DeepSeek V4.1 Flash 1017 75.1% Oct 2026 Official leaderboard
11 Gemini 3.8 Flash 73.7% Sep 2026 Official leaderboard
12 MiMo-V2.6 Pro 72.1% Sep 2026 Official leaderboard
13 GLM-5.3-Flash 66.7% Aug 2026 Official leaderboard
14 GPT-6 Luna 60.4% Sep 2026 Official leaderboard
15 Claude Haiku 5.5 56.9% Oct 2026 Official leaderboard

Rankings by part

We read 5 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as PEBBLE, ATLAS and FORGE. Kimi K3 leads 3 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

PEBBLE

15 models we track have a score here.

  1. 1 Kimi K3 ↑8 100%
  2. 2 GLM-5.3 ↑6 100%
  3. 3 GLM-5.3-Flash ↑10 100%
  4. 4 Gemini 3.8 Flash ↑7 100%
  5. 5 GPT-6 Astra 100%
  6. 6 MiMo-V2.6 Pro ↑6 100%
  7. 7 Claude Opus 5.5 ↓4 100%
  8. 8 GPT-6 Luna ↑6 100%
  9. 9 Claude Sonnet 5.5 ↓2 100%
  10. 10 GPT-6.1 Sol ↓6 100%

5 models have held number one on the official leaderboard since its release in August 2026

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

Moonshot AI Z.ai OpenAI Anthropic Google DeepMind
Own leaderboard Passed 85% in Sep 2026, by GPT-6 Astra

Source: Official Retrobench v0.4 leaderboard

GPT-6 Astra on the official leaderboard passed 85% in September 2026

That was 1 month after its release. By the same age, 20% of the benchmarks from 2026 had passed 85%. All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2027 Near the ceiling (85%) Gemini 4 Carbon GPT-6 Astra GLM-5.3 0% 20% 40% 60% 80% 100% 2027 Near the ceiling (85%) Gemini 4 Carbon GLM-5.3

Source: Official Retrobench v0.4 leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Coding

All 523 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

Data audit

All 10 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access The tasks are private.
  • Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (15 entries).
  • Pass Models scored 15 models we track.
  • Pass Newest score The newest is from November 2026, for Claude Fable 5.2.
  • Pass Frontier models 7 of the top 10 have a score, led by Claude Fable 5.2 (#2).
  • Note Headroom Saturated. The best is 99.9%, past 85% since September 2026.
Leaderboard import 3 pass
  • Pass Importer We read the board on retrobench.ai: 15 rows for models we track at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check