Benchmark Almanac

ProjectTest

Claude-3.5-Sonnet (snapshot unspecified) 84.3% tops its leaderboard Gemini 3 Pro 82% leads the models we track
Not in model cards

Models write unit tests for existing Python, Java, and JavaScript projects. The authors run the tests and measure code coverage.

Nine models attempted test generation on 20 projects per language.

Released
February 2025
Kind
Fixing real repositories
Level
University
Scored
Checked automatically
Answers
Code
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code → Code
Interaction
Single turn
Environment
Code sandbox
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
Strongest model tested
Gemini 3 Pro #56 on the capability index
Last model added
November 2025
Scores refreshed
November 20th, 2026
11
Entries in its paper, 6 models we track

Claude-3.5-Sonnet (snapshot unspecified) tops the leaderboard with 84.3% Gemini 3 Pro leads with 82% among models released by November 2025

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The paper reports no overall score, so this is the plain mean of correctness rates for Python, Java and JavaScript. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Gemini 3 Pro, at 82%. This view shows only models we track, each at its best.

The newest model with a score here came out in November 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 Claude-3.5-Sonnet (snapshot unspecified) Manual fixing 84.3%
  2. 2 Gemini 3 Pro Manual fixing 82%
  3. 3 o1 Manual fixing 81.3%
  4. 4 GPT-5 mini Manual fixing 81.3%
  5. 5 GPT-4-Turbo (snapshot unspecified) Manual fixing 74%
  6. 6 Gemini 2.0 Flash Feb 2025, Manual fixing 64.3%
  7. 7 GPT-3.5-Turbo (snapshot unspecified) Manual fixing 64%
  8. 8 DeepSeek Coder 6.7B Manual fixing 57.3%
  9. 9 CodeGemma Manual fixing 50.3%
  10. 10 CodeQwen1.5-7B Manual fixing 46%
  11. 11 CodeLlama Manual fixing 43%

Source: The ProjectTest paper

#EntryScoreSettingDateSource
1 Claude-3.5-Sonnet (snapshot unspecified) 84.3% Manual fixing Official leaderboard
2 Gemini 3 Pro 82% Manual fixing Nov 2025 Official leaderboard
3 o1 81.3% Manual fixing Dec 2024 Official leaderboard
4 GPT-5 mini 81.3% Manual fixing Aug 2025 Official leaderboard
5 GPT-4-Turbo (snapshot unspecified) 74% Manual fixing Official leaderboard
6 Gemini 2.0 Flash Feb 2025 64.3% Manual fixing Feb 2025 Official leaderboard
7 GPT-3.5-Turbo (snapshot unspecified) 64% Manual fixing Official leaderboard
8 DeepSeek Coder 6.7B 57.3% Manual fixing Nov 2023 Official leaderboard
9 CodeGemma 50.3% Manual fixing Official leaderboard
10 CodeQwen1.5-7B 46% Manual fixing Apr 2024 Official leaderboard
11 CodeLlama 43% Manual fixing Official leaderboard
#ModelScoreSettingDateSource
1 Gemini 3 Pro 82% Manual fixing Nov 2025 Official leaderboard
2 o1 81.3% Manual fixing Dec 2024 Official leaderboard
3 GPT-5 mini 81.3% Manual fixing Aug 2025 Official leaderboard
4 Gemini 2.0 Flash Feb 2025 64.3% Manual fixing Feb 2025 Official leaderboard
5 DeepSeek Coder 6.7B 57.3% Manual fixing Nov 2023 Official leaderboard
6 CodeQwen1.5-7B 46% Manual fixing Apr 2024 Official leaderboard

Rankings by part

We read 3 parts from the benchmark’s paper: Python, Java and JavaScript. Each ranks models on one part of the test. Gemini 3 Pro leads 1 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.

Python

6 models we track have a score here.

  1. 1 Gemini 3 Pro 91%
  2. 2 o1 89%
  3. 3 GPT-5 mini 83%
  4. 4 Gemini 2.0 Flash Feb 2025 61%
  5. 5 DeepSeek Coder 6.7B 53%
  6. 6 CodeQwen1.5-7B 46%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Gemini 3 Pro
Vanilla Self-fixing
82%
o1
Vanilla Manual fixi…
81.3%
GPT-5 mini
Vanilla Self-fixing Manual fixi…
81.3%
Gemini 2.0 Flash
Vanilla Self-fixing Manual fixi…
64.3%
DeepSeek Coder 6.7B
Self-fixing Vanilla Manual fixi…
57.3%
CodeQwen1.5-7B
Vanilla Self-fixing Manual fixi…
46%

Source: The ProjectTest paper

ModelEffortHarnessScoreCost per taskOutput tokens
Gemini 3 Pro Manual fixing 82%
o1 Manual fixing 81.3%
GPT-5 mini Manual fixing 81.3%
Gemini 3 Pro Self-fixing 79%
GPT-5 mini Self-fixing 70.7%
Gemini 3 Pro Vanilla 69%
Gemini 2.0 Flash Feb 2025 Manual fixing 64.3%
o1 Self-fixing 63%
o1 Vanilla 62.7%
GPT-5 mini Vanilla 60.3%
DeepSeek Coder 6.7B Manual fixing 57.3%
Gemini 2.0 Flash Feb 2025 Self-fixing 51%
CodeQwen1.5-7B Manual fixing 46%
Gemini 2.0 Flash Feb 2025 Vanilla 41.3%
DeepSeek Coder 6.7B Vanilla 35.7%
CodeQwen1.5-7B Self-fixing 27.3%
CodeQwen1.5-7B Vanilla 15.7%
DeepSeek Coder 6.7B Self-fixing 15%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Gemini 3 Pro
Vanilla Self-fixing
82%
o1
Vanilla Manual fixi…
81.3%
GPT-5 mini
Vanilla Self-fixing Manual fixi…
81.3%
Gemini 2.0 Flash
Vanilla Self-fixing Manual fixi…
64.3%
DeepSeek Coder 6.7B
Self-fixing Vanilla Manual fixi…
57.3%
CodeQwen1.5-7B
Vanilla Self-fixing Manual fixi…
46%

Source: The ProjectTest paper

ModelEffortHarnessScoreCost per taskOutput tokens
Gemini 3 Pro Manual fixing 82%
o1 Manual fixing 81.3%
GPT-5 mini Manual fixing 81.3%
Gemini 3 Pro Self-fixing 79%
GPT-5 mini Self-fixing 70.7%
Gemini 3 Pro Vanilla 69%
Gemini 2.0 Flash Feb 2025 Manual fixing 64.3%
o1 Self-fixing 63%
o1 Vanilla 62.7%
GPT-5 mini Vanilla 60.3%
DeepSeek Coder 6.7B Manual fixing 57.3%
Gemini 2.0 Flash Feb 2025 Self-fixing 51%
CodeQwen1.5-7B Manual fixing 46%
Gemini 2.0 Flash Feb 2025 Vanilla 41.3%
DeepSeek Coder 6.7B Vanilla 35.7%
CodeQwen1.5-7B Self-fixing 27.3%
CodeQwen1.5-7B Vanilla 15.7%
DeepSeek Coder 6.7B Self-fixing 15%

2 models have held number one in the benchmark’s paper since its release in February 2025

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

OpenAI Google DeepMind
Benchmark paper Not tested since Nov 2025

Source: The ProjectTest paper

Scores over time

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

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2024 2025 2026 Near the ceiling (85%) Gemini 3 Pro o1 DeepSeek Coder 6.7B 0% 20% 40% 60% 80% 100% 2024 2025 2026 Near the ceiling (85%) Gemini 3 Pro DeepSeek Coder 6.7B

Source: Official ProjectTest 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

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 12 months ago, in November 2025, for Gemini 3 Pro. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the benchmark’s paper (11 entries).
  • Pass Models scored 6 models we track, plus 5 other entries on its leaderboard.
  • Warning Newest score The newest came 12 months ago, in November 2025, for Gemini 3 Pro.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 3 Pro (#56).
  • Pass Headroom The best is 82%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 18 rows for models we track and 5 other entries 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