Benchmark Almanac

ARCS

GPT-5 57.9% leads since October 2026
Not in model cards

Models identify ambiguities in a database question and write SQL for the intended interpretation. The scorer executes the query against the database.

Megagon Labs lists 28 configurations of 23 models. Its 311 test cases cover 101 questions across six databases.

Released
October 2026
Kind
Data science and SQL
Level
University
Scored
Checked automatically
Answers
Code
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code → Text, Code
Interaction
Single turn
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
Strongest model tested
GPT-6 Sol #13 on the capability index
Last model added
September 2026
Scores refreshed
November 20th, 2026
23
Models scored on its own leaderboard

GPT-5 leads with 57.9% among models released by September 2026

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. Every run is in the next chart.

The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-5 medium 57.9%
  2. 2 GPT-6 Sol medium 51.8%
  3. 3 GPT-6 Luna medium 46.3%
  4. 4 GPT-5 mini medium 44.4%
  5. 5 o4-mini high 44%
  6. 6 Gemini 3 Pro high 44%
  7. 7 Claude Sonnet 4.5 high 43.7%
  8. 8 Claude Opus 4.5 high 38.6%
  9. 9 Gemini 2.5 Pro Jun 2025 30.9%
  10. 10 GPT-4.1 29.9%
  11. 11 GPT-5 nano medium 29.3%
  12. 12 Claude Haiku 4.5 high 28.6%
  13. 13 Gemini 2.5 Flash Sep 2025 27.6%
  14. 14 gpt-oss-120b 26.1%
  15. 15 DeepSeek-V3.1 23.5%
  16. 16 DeepSeek-R1 May 2025 23.1%
  17. 17 GPT-4.1 mini 21.9%
  18. 18 Qwen3-235B-A22B-Instruct Jul 2025 20.3%
  19. 19 Kimi K2 Thinking 19.3%
  20. 20 Qwen3-Coder-480B-A35B 15.8%
  21. 21 gpt-oss-20b 7.40%
  22. 22 GPT-4.1 nano 7.07%
  23. 23 Qwen3-8B 5.79%

Source: Official ARCS leaderboard

#ModelScoreSettingDateSource
1 GPT-5 57.9% medium Aug 2025 Official leaderboard
2 GPT-6 Sol 51.8% medium Sep 2026 Official leaderboard
3 GPT-6 Luna 46.3% medium Sep 2026 Official leaderboard
4 GPT-5 mini 44.4% medium Aug 2025 Official leaderboard
5 o4-mini 44% high Apr 2025 Official leaderboard
6 Gemini 3 Pro 44% high Nov 2025 Official leaderboard
7 Claude Sonnet 4.5 43.7% high Sep 2025 Official leaderboard
8 Claude Opus 4.5 38.6% high Nov 2025 Official leaderboard
9 Gemini 2.5 Pro Jun 2025 30.9% Jun 2025 Official leaderboard
10 GPT-4.1 29.9% Apr 2025 Official leaderboard
11 GPT-5 nano 29.3% medium Aug 2025 Official leaderboard
12 Claude Haiku 4.5 28.6% high Oct 2025 Official leaderboard
13 Gemini 2.5 Flash Sep 2025 27.6% Sep 2025 Official leaderboard
14 gpt-oss-120b 26.1% Aug 2025 Official leaderboard
15 DeepSeek-V3.1 23.5% Aug 2025 Official leaderboard
16 DeepSeek-R1 May 2025 23.1% May 2025 Official leaderboard
17 GPT-4.1 mini 21.9% Apr 2025 Official leaderboard
18 Qwen3-235B-A22B-Instruct Jul 2025 20.3% Jul 2025 Official leaderboard
19 Kimi K2 Thinking 19.3% Nov 2025 Official leaderboard
20 Qwen3-Coder-480B-A35B 15.8% Jul 2025 Official leaderboard
21 gpt-oss-20b 7.40% Aug 2025 Official leaderboard
22 GPT-4.1 nano 7.07% Apr 2025 Official leaderboard
23 Qwen3-8B 5.79% Apr 2025 Official leaderboard

Rankings by part

We read 3 parts from the benchmark’s own leaderboard: Full Recall, Perfect and EX disamb. Each ranks models on one part of the test. GPT-5 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.

Full Recall

23 models we track have a score here.

  1. 1 GPT-5 61.1%
  2. 2 GPT-5 mini ↑2 42.1%
  3. 3 GPT-6 Sol ↓1 40.8%
  4. 4 GPT-6 Luna ↓1 31.2%
  5. 5 o4-mini 30.6%
  6. 6 Gemini 3 Pro 26.4%
  7. 7 Claude Sonnet 4.5 23.5%
  8. 8 Claude Opus 4.5 22.5%
  9. 9 Gemini 2.5 Pro Jun 2025 20.9%
  10. 10 Claude Haiku 4.5 ↑2 17.7%

The best score, 57.9%, costs $0.09 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 Cost frontier One model at each effort level
0% 20% 40% 60% $0.0001 $0.001 $0.01 $0.10 $1.00 Cost per task, log scale GPT-5, medium GPT-6 Sol, medium GPT-6 Luna, medium GPT-5 nano, medium gpt-oss-120b gpt-oss-20b 0% 20% 40% 60% $0.0001 $0.001 $0.01 $0.10 $1.00 Cost per task, log scale GPT-5, medium GPT-6 Sol, medium

Source: Official ARCS leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5 medium 57.9% $0.09
GPT-5 high 57.6% $0.16
GPT-6 Sol medium 51.8% $0.04
GPT-5 low 48.5% $0.04
GPT-6 Luna medium 46.3% $0.006
GPT-5 mini medium 44.4% $0.01
o4-mini high 44% $0.06
Gemini 3 Pro high 44% $0.14
Claude Sonnet 4.5 high 43.7% $0.09
o4-mini medium 42.4% $0.03
Claude Opus 4.5 high 38.6% $0.15
GPT-5 minimal 37.6% $0.02
o4-mini low 36% $0.02
Gemini 2.5 Pro Jun 2025 30.9% $0.06
GPT-4.1 29.9% $0.03
GPT-5 nano medium 29.3% $0.005
Claude Haiku 4.5 high 28.6% $0.03
Gemini 2.5 Flash Sep 2025 27.6% $0.01
gpt-oss-120b 26.1% $0.003
DeepSeek-V3.1 23.5% $0.01
DeepSeek-R1 May 2025 23.1% $0.04
GPT-4.1 mini 21.9% $0.007
Qwen3-235B-A22B-Instruct Jul 2025 20.3% $0.02
Kimi K2 Thinking 19.3% $0.03
Qwen3-Coder-480B-A35B 15.8% $0.03
gpt-oss-20b 7.40% $0.0003
GPT-4.1 nano 7.07% $0.006
Qwen3-8B 5.79% $0.009

Number one over time

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

OpenAI
Own leaderboard Not tested since Oct 2026
GPT-5 57.9%

Source: Official ARCS leaderboard

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
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) GPT-5 o4-mini GPT-4.1 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) GPT-5 o4-mini

Source: Official ARCS 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 All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (23 entries).
  • Pass Models scored 23 models we track.
  • Pass Newest score The newest is from September 2026, for GPT-6 Luna.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-6 Sol (#13).
  • Pass Headroom The best is 57.9%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on megagonlabs.github.io: 28 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