Gid Bench: AI model benchmarks in one score
Gid Bench puts 13 public benchmarks on one scale and averages them in 4 groups: coding, agents and long tasks, math and reasoning, knowledge. Most results come from independent test runs; where a model maker's own launch numbers are used, they are marked "as reported". Updated daily, last on Nov 17, 2026.
Ranking
19 models with enough tests · group scores out of 100
| # | Model | Gid Bench | Coding | Agents and long tasks | Math and reasoning | Knowledge | Tests | Released |
|---|---|---|---|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 73.9 | 61.93 tests | 56.73 tests | 90.25 tests | 86.92 tests | 13 of 13 | 2026-09-03 |
| 2= | Gemini 4 ArgonGoogle DeepMind | 72.4 | 53.33 tests | 71.62 tests | 77.15 tests | 87.62 tests | 12 of 13 | 2026-10-13 |
| 2= | Claude Opus 5.5Anthropic | 72.3 | 71.13 tests | 51.83 tests | 84.95 tests | 81.62 tests | 13 of 13 | 2026-09-22 |
| 4= | GPT-6.1 SolOpenAI | 71.2 | 58.72 tests | 51.83 tests | 88.65 tests | 85.82 tests | 12 of 13 | 2026-09-29 |
| 4= | GPT-6.1 TerraOpenAI | 70.3 | 60.82 tests | 53.63 tests | 89.94 tests | 76.82 tests | 11 of 13 | 2026-10-26 |
| 6= | Claude Sonnet 5.5Anthropic | 66.2 | 66.93 tests | 47.03 tests | 78.45 tests | 72.32 tests | 13 of 13 | 2026-09-28 |
| 6= | Claude Opus 5Anthropic | 65.7 | 58.63 tests | 54.92 tests | 71.45 tests | 77.82 tests | 12 of 13 | 2026-07-24 |
| 8 | Claude Fable 5Anthropic | 59.4 | 49.33 tests | 36.02 tests | 74.65 tests | 77.62 tests | 12 of 13 | 2026-06-09 |
| 9 | GPT-6 SolOpenAI | 57.7 | 39.82 tests | 34.63 tests | 78.05 tests | 78.52 tests | 12 of 13 | 2026-09-22 |
| 10= | Gemini 3.7 FlashGoogle DeepMind | 52.1 | 30.72 tests | 33.92 tests | 60.85 tests | 83.12 tests | 11 of 13 | 2026-08-13 |
| 10= | Claude Haiku 5.5Anthropic | 51.8 | 49.83 tests | 40.62 tests | 60.05 tests | 56.82 tests | 12 of 13 | 2026-10-07 |
| 12= | Gemini 3.8 FlashGoogle DeepMind | 50.0 | 24.93 tests | 33.32 tests | 58.15 tests | 83.72 tests | 12 of 13 | 2026-09-02 |
| 12= | Grok 4.6xAI | 49.2 | 33.22 tests | 39.82 tests | 51.45 tests | 72.62 tests | 11 of 13 | 2026-08-12 |
| 14= | Kimi K3Moonshot | 46.2 | 32.53 tests | 26.92 tests | 52.95 tests | 72.62 tests | 12 of 13 | 2026-07-16 |
| 14= | DeepSeek V4.1 FlashDeepSeek | 45.9 | 38.72 tests | 22.72 tests | 55.75 tests | 66.52 tests | 11 of 13 | 2026-09-09 |
| 14= | Grok 4.7xAI | 45.8 | 34.93 tests | 27.32 tests | 46.15 tests | 75.02 tests | 12 of 13 | 2026-09-21 |
| 17 | GPT-5.6 LunaOpenAI | 43.1 | 15.72 tests | 30.42 tests | 59.75 tests | 66.82 tests | 11 of 13 | 2026-07-09 |
| 18 | GPT-6 LunaOpenAI | 40.5 | 20.02 tests | 21.53 tests | 54.25 tests | 66.22 tests | 12 of 13 | 2026-09-22 |
| 19 | Mistral Large 4 PreviewMistral AI | 37.9 | 26.82 tests | 25.82 tests | 45.05 tests | 53.82 tests | 11 of 13 | 2026-10-06 |
Head to head
The top models and the newest release, test by test. Best result in each row is highlighted.
| Test | Gemini 4 Carbonnew | GPT-6 Astra | Gemini 4 Argon | Claude Opus 5.5 | GPT-6.1 Sol | GPT-6.1 Terra |
|---|---|---|---|---|---|---|
| Gid Bench | waiting for 1 group | 73.9 | 72.4 | 72.3 | 71.2 | 70.3 |
| Coding | ||||||
| MirrorCodeEpoch AI | — | 60.0%max effort | 58.9%high effort | 95.6%max effort | 61.1%max effort | 76.7%max effort |
| FrontierSWE v2Proximal | 61.2%high effort | 65.5%max effort | 55.0%high effort | 62.3%max effort | — | — |
| WeirdML v3Håvard Tveit Ihle | — | 54.0%xhigh effort | 41.4%high effort | 49.8%xhigh effort | 50.6%xhigh effort | 40.5%xhigh effort |
| Agents and long tasks | ||||||
| Furniture AssemblyEpoch AI | 77.5%high effort | 80.0%max effort | 72.5%high effort | 83.3%max effort | 80.0%max effort | 79.2%max effort |
| SpireClimbEpoch AI | — | 34.0%max effort | — | 5.8%max effort | 23.8%max effort | 22.3%max effort |
| APEX-AgentsMercor | 84.2% | 64.7% | 82.4% | 73.5%max effort | 60.0%max effort | 68.3%max effort |
| Math and reasoning | ||||||
| FrontierMath, tiers 1-3Epoch AI | 90.5%high effort | 93.7%max effort | 86.0%high effort | 91.2%max effort | 93.7%max effort | 90.9%max effort |
| FrontierMath, tier 4Epoch AI | 95.1%high effort | 97.6%high effort | 78.0%high effort | 95.0%max effort | 100.0%max effort | 92.7%max effort |
| OTIS Mock AIMEEpoch AI | 100.0%high effort | 100.0%max effort | 100.0%high effort | 100.0%max effort | 100.0%max effort | 100.0%max effort |
| Chess PuzzlesEpoch AI | 78.0%high effort | 74.0%max effort | 59.0%high effort | 67.0%max effort | 68.0%max effort | — |
| Mystery Game PuzzlesEpoch AI | 73.0%high effort | 84.0%max effort | 64.0%high effort | 71.0%max effort | 80.0%max effort | 74.0%max effort |
| Knowledge | ||||||
| GPQA DiamondEpoch AI | 96.5%high effort | 95.8%max effort | 96.4%high effort | 90.6%max effort | 95.4%max effort | 94.3%max effort |
| SimpleQA VerifiedEpoch AI | 77.7%high effort | 75.6%max effort | 76.0%high effort | 72.2%max effort | 73.9%max effort | 57.3%max effort |
† As reported by the model's maker on its launch page (Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2); used only until an independent run exists. Each cell is the model's best effort level on that test.
New models waiting for tests
Released recently, not enough tests yet for a Gid Bench score.
- Claude Fable 5.2 · Anthropic · released 2026-11-09 · needs more tests in coding
- DeepSeek V4.1 Flash 1017 · DeepSeek · released 2026-10-17 · needs more tests in coding, agents and long tasks
- Doubao Seed Evolving · ByteDance · released 2026-11-02 · needs more tests in coding, agents and long tasks
- Gemini 4 Carbon · Google DeepMind · released 2026-11-10 · needs more tests in coding
- Kimi K3.1 Pro · Moonshot · released 2026-10-23 · needs more tests in coding, agents and long tasks
- Mistral Large 4 · Mistral AI · released 2026-10-29 · needs more tests in coding, agents and long tasks
- Muse Spark Open · Meta AI · released 2026-10-15 · needs more tests in coding, agents and long tasks
- Muse Wonder · Meta AI · released 2026-10-12 · needs more tests in knowledge
Effort level against cost
Terminal-Bench 4.0: tasks solved and what the whole run cost, one point per effort level (low to max). Hover a point for the numbers.
Source: Terminal-Bench (tbench.ai), agent plus model runs with the cost of each run as published. All Terminal-Bench results →
How Gid Bench is built
- Each test is put on a 0 to 100 scale between its random-guess baseline and its ceiling, as Epoch AI defines them, so a hard test and an easy one weigh the same.
- A model's best effort level counts. A group score is the average of its tests; Gid Bench is the average of the 4 groups.
- No score until a model has at least 2 tests in every group, so a model is not ranked on one lucky test.
- Scores less than 1 point apart share a place (=): one test more or less moves a score by about that much.
- A maker's own number counts only on a test an independent source also runs, and only after its numbers for the other models in its table match the independent runs within 1 point (same test version). An independent run replaces it. Scores that use one are marked "prelim.".
- Only tests whose terms allow reuse are included; some well-known ones are left out for that reason.
Maker numbers checked: FrontierSWE v2: 5 of 5 other models match, used.
Sources
- Epoch AI: Epoch AI, ‘Capabilities & benchmarking’, epoch.ai, CC BY 4.0
- FrontierSWE v2 by Proximal, via Epoch AI's hub
- WeirdML v3 by Håvard Tveit Ihle, via Epoch AI's hub
- APEX-Agents by Mercor, via Epoch AI's hub