The AI models leading today, and the ones we build on.
Ranked by the Epoch Capabilities Index, from Epoch’s benchmark data dated 11 November 2026. Below the table: which model we would pick for which job, and the numbers behind each pick.
The top ten
Top 10 by the Epoch Capabilities Index.
| Model | Developer | Capabilities Index | GPQA Diamond | SimpleQA Verified | We build on |
|---|---|---|---|---|---|
| Claude Fable 5.2Anthropic | Anthropic | 173.44 | 93.9% | 94.0% | Amtec builds on |
| Gemini 4 CarbonGoogle DeepMind | Google DeepMind | 169.76 | 96.5% | 77.7% | |
| GPT-6 AstraOpenAI | OpenAI | 169.00 | 95.8% | 75.6% | Amtec builds on |
| Claude Opus 5.5Anthropic | Anthropic | 168.54 | 90.6% | 72.2% | Amtec builds on |
| GPT-6.1 SolOpenAI | OpenAI | 168.48 | 95.4% | 73.9% | Amtec builds on |
| GPT-6.1 TerraOpenAI | OpenAI | 167.67 | 94.3% | 57.3% | Amtec builds on |
| Claude Fable 5.1Anthropic | Anthropic | 166.99 | Not tested | 70.8% | Amtec builds on |
| Claude Sonnet 5.5Anthropic | Anthropic | 166.56 | 95.6% | 46.5% | Amtec builds on |
| Gemini 4 ArgonGoogle DeepMind | Google DeepMind | 166.48 | 96.4% | 76.0% | |
| Claude Opus 5Anthropic | Anthropic | 164.68 | 93.9% | 59.9% | Amtec builds on |
Capabilities Index: overall performance across several tests. GPQA Diamond: expert-level science questions. SimpleQA Verified: factual accuracy.
Source: Epoch AI, Benchmarking data (CC BY 4.0). Data as of 11 November 2026.
Which model for which job
The index ranks how capable a model is. It does not say what a model costs to do your work, or which work it does best. Epoch tracks that too, so here is how we read it.
Which models: the ten in the table above, plus Claude Haiku 5.5, the smallest model we build on. Every figure comes from Epoch’s benchmark data dated 11 November 2026. Each is an average over one test’s tasks, not a promise for any one job.
Professional work is close at the top
APEX-Agents gives an agent the multi-step tasks that banking, consulting and law teams hand out, and checks the finished work. Each bar is a model’s best setting. Bars less than about 12 points apart can be inside the margin of error.
Coding: what each extra point costs
CursorBench sets coding tasks in real codebases and records what each attempt costs. Each line is one model, from its lowest effort setting to its highest. Up and to the left is better: more solved for less. It measures coding only, so it does not rank models for other work.
No CursorBench result yet for GPT-6 Astra or GPT-6.1 Sol or GPT-6.1 Terra, so they are not on the chart.
What we would pick
Coding in an existing codebase
Claude Opus 5.5, high effort
On CursorBench it solves 56% of tasks at $3.97 a task. Every setting that solves more costs at least 2.1 times as much per task: the cheapest is Gemini 4 Carbon at medium effort, 60.1% for $8.49. Max effort adds 1.8 points for 3.4 times the cost.
Routine work at volume
Claude Sonnet 5.5, high effort
85% of Opus 5.5's high-effort CursorBench score for 30% of the cost: $1.20 a task. At max effort it passes 75.5% of the APEX-Agents professional tasks.
Sorting an email, routing a request or pulling one field
Claude Haiku 5.5
At high effort it solves 42.3% of CursorBench tasks for $0.32 a task, where Opus 5.5 at high effort costs 12 times as much. At max effort it passes 56.1% of the APEX-Agents tasks. For a narrow decision we check it on the client’s own examples before relying on it.
Document-heavy professional work
Worth a look: Gemini 4 Carbon
The top APEX-Agents score, 84.2%. Gemini 4 Argon, Claude Fable 5.2, Claude Sonnet 5.5 and Claude Opus 5.5 are within the margin of error, so we compare them on the client’s own documents before choosing.
Where extra money buys no more on this test
Claude Opus 5 and Fable 5.1
On CursorBench, Opus 5.5 at high effort scores higher than every setting of both and costs less per task than any of them.
Table view
| Model | Effort | Solved | Cost per task |
|---|---|---|---|
| Claude Haiku 5.5 | low | 30.9% | $0.08 |
| Claude Haiku 5.5 | medium | 36.9% | $0.17 |
| Claude Haiku 5.5 | high | 42.3% | $0.32 |
| Claude Sonnet 5.5 | low | 35.8% | $0.37 |
| Claude Sonnet 5.5 | medium | 39.2% | $0.52 |
| Claude Haiku 5.5 | extra high | 44.3% | $0.56 |
| Claude Haiku 5.5 | max | 48.4% | $1.12 |
| Claude Opus 5.5 | low | 43.7% | $1.17 |
| Claude Sonnet 5.5 | high | 47.8% | $1.20 |
| Claude Sonnet 5.5 | extra high | 53.1% | $2.81 |
| Claude Opus 5.5 | medium | 52.5% | $2.91 |
| Gemini 4 Argon | minimal | 48.8% | $3.54 |
| Claude Fable 5.2 | low | 53.2% | $3.94 |
| Claude Opus 5.5 | high | 56% | $3.97 |
| Gemini 4 Argon | low | 50.6% | $4.09 |
| Gemini 4 Argon | medium | 52.8% | $4.39 |
| Gemini 4 Argon | high | 52.9% | $4.57 |
| Claude Fable 5.2 | medium | 54.8% | $4.66 |
| Claude Opus 5 | low | 40.7% | $4.87 |
| Claude Fable 5.1 | low | 45.1% | $5.44 |
| Claude Opus 5 | medium | 43.3% | $6.94 |
| Claude Opus 5.5 | extra high | 56% | $6.98 |
| Claude Fable 5.1 | medium | 46.8% | $7.05 |
| Claude Sonnet 5.5 | max | 55.5% | $7.05 |
| Gemini 4 Carbon | medium | 60.1% | $8.49 |
| Claude Opus 5 | high | 44.7% | $9.00 |
| Claude Fable 5.1 | high | 49.2% | $9.08 |
| Gemini 4 Carbon | high | 60.4% | $9.22 |
| Claude Fable 5.2 | high | 56.9% | $9.70 |
| Claude Opus 5 | extra high | 46.1% | $11.43 |
| Gemini 4 Carbon | low | 58.6% | $11.89 |
| Claude Opus 5 | max | 46.6% | $11.95 |
| Claude Fable 5.1 | extra high | 51.6% | $13.01 |
| Claude Opus 5.5 | max | 57.8% | $13.43 |
| Claude Fable 5.2 | extra high | 56.3% | $15.48 |
| Claude Fable 5.1 | max | 51.8% | $17.28 |
| Claude Fable 5.2 | max | 60% | $32.79 |
| Model | Passed first try |
|---|---|
| Gemini 4 Carbon | 84.2% |
| Gemini 4 Argon | 82.4% |
| Claude Fable 5.2 | 80.1% |
| Claude Sonnet 5.5 | 75.5% |
| Claude Opus 5.5 | 73.5% |
| Claude Fable 5.1 | 68.6% |
| GPT-6.1 Terra | 68.3% |
| Claude Opus 5 | 65.8% |
| GPT-6 Astra | 64.7% |
| GPT-6.1 Sol | 60% |
| Claude Haiku 5.5 | 56.1% |
Data: Epoch AI, Benchmarking data (CC BY 4.0), archive files dated 11 November 2026. Cost per task is Epoch’s figure for one attempt.