Compare AI models
Closed and open-weight models · Separate benchmarks · Exact tested configurations
Closed vs open models · side by side
Select up to four models below. Selections stay while you filter. Best selected result per row is highlighted: higher scores and speed, lower cost and time. A dash means no reported result.
| Benchmark | Claude Fable 5.2 (Max, Default Fallback)Closed weights | GPT-6 Astra (Max)Closed weights | Kimi K3.1 ProOpen weights | DeepSeek V4.1 Flash 1017 (Max)Open weights |
|---|---|---|---|---|
| AA IndexComposite index · points · higher is better | 64.6 | 56.9 | 46.5 | 32.6 |
| HLE-DiamondAccuracy · % · higher is better | 80.2% | 63.8% | 37.3% | 25.3% |
| GPQA DiamondAccuracy · % · higher is better | — | 96.1% | — | — |
| Knowledge accuracyAccuracy · % · higher is better | 79.1% | 61.6% | 48.9% | 39.6% |
| FrontierCode 2.0Mean rubric score · % · higher is better | 48.8% | 42.4% | 25.5% | 28.5% |
| Benchmark task costUSD / task · lower is better | $26.54 | $6.24 | $3.68 | $0.60 |
| Benchmark task timeminutes / task · lower is better | 46.8 | 24.3 | 38.3 | 15.1 |
| Output speedtokens / second · higher is better | 97.6 | 35.6 | 31.4 | 153.1 |
| First answer tokenseconds · lower is better | 267.8 | 232.0 | 172.1 | 45.4 |
| API input priceUSD / 1M tokens · lower is better | $8.00 | $10.00 | $2.00 | $0.30 |
| API output priceUSD / 1M tokens · lower is better | $40.00 | $50.00 | $10.00 | $1.20 |
Task cost and time are weighted Intelligence Index measurements, not estimates for your workload. API token prices are separate. Open weights do not mean free inference.
Saved on this device · Links include model configurations and board filters · Search text stays out of links · Scores refresh with the site
Saved comparisons (0/8)
Selections and filters are saved; this is not a historical score archive. Download a dated PNG to keep the results shown today.
Save a comparison to return to it later. No account needed.
Show quality vs cost chart
Filtered board · 103 configurations with a reported cost · Higher and further left means higher AA Index at lower benchmark cost. Green: open weights; accent: closed weights. Rings mark your selections.
Read exact values and select models in the table below. These are evaluator task costs, not estimates for your workload.
Show selected-model ranking chart
AA Index · composite points, not accuracy · axis 0–70 · green: open weights · accent: closed weights
- Claude Fable 5.2 (Max, Default Fallback)64.6
- GPT-6 Astra (Max)56.9
- Kimi K3.1 Pro46.5
- DeepSeek V4.1 Flash 1017 (Max)32.6
Showing 12 of 111 configurations · Source snapshot 2026-11-12 · Select column headings to sort
| Compare | Model / tested configuration | Details | |||||
|---|---|---|---|---|---|---|---|
| Claude Fable 5.2 (Max, Default Fallback)Anthropic · Closed weights | 64.6 | 80.2% | — | 79.1% | 48.8% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| Claude Fable 5.2 (Xhigh, Default Fallback)Anthropic · Closed weights | 62.2 | 75.8% | — | 78.0% | 47.6% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| Claude Fable 5.2 (High, Default Fallback)Anthropic · Closed weights | 61.8 | 73.8% | — | 77.2% | 49.1% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| Gemini 4 Carbon (High)Google · Closed weights | 60.0 | 69.1% | — | 58.3% | 40.6% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| Claude Opus 5.5 (Max, Default Fallback)Anthropic · Closed weights | 59.1 | 67.9% | — | 65.5% | 46.4% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| Gemini 4 Carbon (Medium)Google · Closed weights | 58.3 | 67.4% | — | 48.9% | 39.5% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| Claude Opus 5.5 (Xhigh, Default Fallback)Anthropic · Closed weights | 57.4 | 64.0% | — | 65.5% | 43.8% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| GPT-6 Astra (Max)OpenAI · Closed weights | 56.9 | 63.8% | 96.1% | 61.6% | 42.4% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| Claude Fable 5.2 (Medium, Default Fallback)Anthropic · Closed weights | 56.4 | 66.3% | — | 77.0% | 41.3% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| GPT-6 Astra (Xhigh)OpenAI · Closed weights | 55.3 | 61.5% | 96.3% | 61.0% | 41.4% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| GPT-6 Astra (High)OpenAI · Closed weights | 54.8 | 59.3% | 94.9% | 61.1% | 40.1% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. | |
| GPT-6.1 Sol (Max)OpenAI · Closed weights | 54.6 | 58.8% | — | 61.8% | 40.6% | Sources & linksFull model analysis ↗Model catalog & task guide → Weights access is publisher-labeled. Availability and license permissions vary. Scores retain the full tested effort and fallback setting in the model name. |
Sources, benchmark definitions & task guides
Artificial Analysis snapshot, checked 2026-11-12. Each row is one tested configuration. Results compare that evaluator’s setup; they do not establish universal accuracy. AA Index is a composite, and FrontierCode is a mean rubric score. Neither is a raw accuracy percentage. Missing results are not zeros.
Costs and speed are from the same dated Artificial Analysis snapshot. Benchmark task cost includes input, cache, reasoning and answer-token costs weighted by the index evaluations. Task time includes inference and reasoning; output tokens/second is a different measure. Unknown costs and times are excluded when a maximum is set, and remain last when sorting. API prices are USD per million tokens, not subscription prices or self-hosting costs.
- Benchmark task cost: Weighted average USD cost per Intelligence Index task, including cached and reasoning tokens. Not a quote for your task.
- Benchmark task time: Inference time per Intelligence Index task, including reasoning. Not your task completion time.
- Output speed: Publisher-measured output tokens per second. This is not end-to-end response time.
- First answer token: Time until the first answer token, including reasoning delay where applicable.
- API input price: Listed uncached API input price per million tokens, not a chat subscription price.
- API output price: Listed API output price per million tokens. Reasoning, tools and retries can affect the bill.
Open weights are downloadable model weights; license permissions differ. This board includes 111 current evaluated configurations with an index score, rather than every model in the catalog.
- AA Index methodology ↗
Artificial Analysis Intelligence Index v5.0.1. Composite index points, not an accuracy percentage.
- HLE-Diamond methodology ↗
Share of 1,000 HLE-Diamond questions judged correct in the publisher’s tool-free evaluation.
- GPQA Diamond methodology ↗
Graduate-level science question accuracy on GPQA Diamond.
- Knowledge accuracy methodology ↗
AA-Omniscience v1.1 accuracy. This is separate from its hallucination-penalized net score.
- FrontierCode 2.0 methodology ↗
FrontierCode-AA 2.0 mean rubric score over three attempts per task in the publisher’s Stirrup harness; not a binary success rate.