Claude Fable 5.2
Anthropic
95.5CTO Index
- Terminal-Bench
- 65.8%
- WebDev
- 1824
- ARC-AGI-2
- 93.8%
- Text Arena
- 1503
- Price $/M
- $8 / $40
Claude Opus 5.5
Anthropic
95.3CTO Index
- Terminal-Bench
- 65.8%
- WebDev
- 1819
- ARC-AGI-2
- 93.3%
- Text Arena
- 1506
- Price $/M
- $4 / $20
Gemini 4 Carbon
92.4CTO Index
- Terminal-Bench
- 64.5%
- WebDev
- 1769
- ARC-AGI-2
- 94.2%
- Text Arena
- 1527
- Price $/M
- $2 / $10
The full leaderboard
Sort by any column. Click a model for its full profile and sources.
| Model | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| #1Claude Fable 5.2Anthropic | 95.5 | 65.8% | 1824 | 93.8% | 1503 | 81.2% | $8 / $40 | 6.0 | 1M |
| #2Claude Opus 5.5Anthropic | 95.3 | 65.8% | 1819 | 93.3% | 1506 | — | $4 / $20 | 11.9 | 1M |
| #3Gemini 4 CarbonGoogle | 92.4 | 64.5% | 1769 | 94.2% | 1527 | 65.1% | $2 / $10 | 23.1 | Not listed |
| #4GPT-6.2 AstraOpenAI | 88.0 | 62.1% | 1766 | 95.8% | 1468 | 70.3% | $8 / $40 | 5.5 | 1.1M |
| #5Gemini 4 NeonGoogle | 87.6 | 61.2% | 1765 | — | 1513 | — | $1.50 / $7.50 | 29.2 | 1M |
| #6GPT-6 AstraOpenAI | 87.2 | 60.3% | 1764 | 95.0% | 1475 | 62.7% | $10 / $50 | 4.4 | 1.1M |
| #7Claude Sonnet 5.5Anthropic | 87.1 | 63.9% | 1772 | 83.9% | 1473 | 24.3% | $2 / $10 | 21.8 | 1M |
| #8GPT-6.1 SolOpenAI | 86.4 | 60.9% | 1743 | 94.2% | 1483 | 52.7% | $2 / $10 | 21.6 | 1.1M |
| #9Claude Fable 5.1Anthropic | 84.9 | 57.6% | 1742 | 90.0% | 1503 | — | $10 / $50 | 4.2 | 1M |
| #10GPT-6.2 SolOpenAI | 84.6 | 61.5% | 1754 | 85.8% | 1465 | 54.9% | $2 / $10 | 21.2 | 1.1M |
| #11Gemini 4 HydrogenGoogle | 81.4 | 56.4% | 1701 | 92.5% | 1490 | — | $1.50 / $7.50 | 27.1 | Not listed |
| #12Gemini 4 ArgonGoogle | 81.3 | 55.5% | 1682 | 91.3% | 1520 | 49.8% | $2 / $10 | 20.3 | 1M |
| #13Claude Opus 5Anthropic | 78.9 | 53.3% | 1692 | 90.4% | 1490 | 30.2% | $5 / $25 | 7.9 | 1M |
| #14Kimi K3.1 ProMoonshot AIOpen weights | 77.7 | 59.7% | 1683 | — | 1482 | — | $2 / $10 | 19.4 | 1M |
| #15GPT-6 SolOpenAI | 74.3 | 50.3% | 1674 | 89.6% | 1458 | 4.6% | $2 / $10 | 18.6 | 1.1M |
| #16GPT-6.1 TerraOpenAI | 73.3 | 53.3% | 1685 | — | 1474 | — | $1 / $5 | 36.7 | 1.1M |
| #17Claude Fable 5Anthropic | 71.4 | 44.8% | 1626 | 89.2% | 1504 | — | $10 / $50 | 3.6 | 1M |
| #18Muse WonderMeta | 69.3 | 47.0% | 1647 | — | 1511 | — | $4 / $20 | 8.7 | 1M |
| #19GPT-5.6 SolOpenAI | 67.4 | 38.5% | 1615 | 92.5% | 1485 | 7.8% | $4 / $20 | 8.4 | Not listed |
| #20DeepSeek V4.1 FlashDeepSeekOpen weights | 66.1 | 38.5% | 1658 | 78.1% | 1475 | 0.2% | $0.15 / $0.60 | 252.0 | 1M |
| #21GLM-5.3Z.aiOpen weights | 62.0 | 41.2% | 1623 | — | 1482 | — | $1.40 / $4.40 | 28.9 | 1M |
| #22GLM-5.3-FlashZ.aiOpen weights | 60.2 | 37.0% | 1610 | 65.8% | 1475 | — | $0.02 / $0.07 | 1853.7 | 1M |
| #23Grok 4.7xAI | 60.0 | 38.8% | 1639 | 61.4% | 1443 | 1.8% | $2 / $6 | 20.0 | 500K |
| #24Claude Haiku 5.5Anthropic | 57.1 | 40.0% | 1595 | 61.3% | 1421 | 0.5% | $0.10 / $0.50 | 285.5 | 1M |
| #25Qwen3.8-MaxAlibaba | 57.1 | 27.0% | 1674 | — | 1484 | — | $2 / $6 | 19.0 | 1M |
| #26Gemini 3.8 FlashGoogle | 55.9 | 18.2% | 1584 | 89.2% | 1497 | 10.4% | $0.75 / $3.75 | 37.3 | 1M |
| #27Kimi K3Moonshot AIOpen weights | 55.6 | 23.0% | 1659 | 60.4% | 1488 | — | $3 / $15 | 9.3 | 1M |
| #28MiMo-V2.6-ProXiaomiOpen weights | 53.2 | 27.0% | 1622 | — | 1481 | — | $0.43 / $0.87 | 98.5 | 1.1M |
| #29Claude Opus 4.8Anthropic | 52.8 | 23.6% | 1555 | 72.1% | 1482 | 1.5% | $5 / $25 | 5.3 | 1M |
| #30Qwen3.8-27BAlibabaOpen weights | 52.3 | — | 1590 | 42.4% | 1435 | — | $0.40 / $3 | 49.8 | 262K |
| #31Gemini 3.7 FlashGoogle | 51.9 | 12.4% | 1589 | 84.6% | 1490 | — | $0.75 / $3.75 | 34.6 | 1M |
| #32Grok 4.6xAI | 51.4 | 19.7% | 1618 | 67.1% | 1457 | 2.1% | $1.60 / $4.80 | 21.4 | 500K |
| #33Inkling 2 SmallThinking MachinesOpen weights | 50.1 | 13.0% | 1643 | 81.7% | 1412 | 0.3% | $0.30 / $1 | 105.5 | Not listed |
| #34Muse Spark 1.3Meta | 49.0 | 14.8% | 1653 | — | 1498 | — | $1.25 / $4.25 | 24.5 | 1M |
| #35Mistral Large 4MistralOpen weights | 46.6 | 27.6% | 1573 | — | 1423 | — | $0.70 / $2.10 | 44.4 | 1M |
| #36MiMo-V2.6-FlashXiaomiOpen weights | 45.3 | 16.1% | 1638 | — | 1448 | — | $0.14 / $0.28 | 258.6 | 1M |
| #37GPT-6 LunaOpenAI | 44.9 | 15.2% | 1578 | 59.3% | 1443 | 0.2% | $0.10 / $0.50 | 224.4 | 1.1M |
| #38Claude Sonnet 5Anthropic | 37.9 | 12.1% | 1539 | — | 1461 | — | $2 / $10 | 9.5 | 1M |
| #39Beam 1.0Reflection AIOpen weights | 27.7 | 6.4% | 1437 | 23.5% | 1434 | — | $0.30 / $1 | 58.3 | 1M |
| –Copilot FusionGitHubPartial | 100.0 | 71.2% | — | — | — | — | Not listed | — | Not listed |
| –Qwen4-MaxAlibabaPartial | 79.3 | — | 1751 | — | — | — | $1.50 / $4.50 | — | Not listed |
| –Muse Wonder FullMetaPartial | 71.7 | — | 1723 | — | — | — | Not listed | — | Not listed |
| –Step 5 PreviewStepFunPartial | 55.5 | — | 1570 | — | 1449 | — | $1 / $2.70 | — | 1M |
| –Hy4TencentPartial | 53.4 | 37.9% | 1650 | — | — | — | Not listed | — | Not listed |
| –Muse EmberMetaPartial | 43.7 | 30.6% | 1608 | — | — | — | $0.10 / $0.50 | — | Not listed |
Models need at least 3 of the 4 indexed benchmarks to get a rank; the rest are marked Partial. A dash means the source has not published a result for that model. ARC-AGI-3 is shown for reference and is not part of the index.
How the CTO Index works
Public leaderboards each measure one thing, and model vendors quote whichever one flatters them. The CTO Index puts the ones that predict real engineering value side by side, weighted toward what engineering teams actually buy models for.
- Primary sources only. Each score is copied from the benchmark's own leaderboard on the snapshot date, never from an aggregator or a vendor press release.
- Normalise. The best tracked model on each benchmark scores 100. Percentages scale linearly. Elo ratings become the expected head-to-head win rate against the leader, so a 100-point Elo gap is about 72.
- Weight. Terminal-Bench 4.1 35%, WebDev Arena 25%, ARC-AGI-2 20%, Text Arena 20%. Agentic coding gets the most weight because that is where most engineering spend on models is going.
- Require coverage. The weighted mean is taken over the benchmarks a model has. Models with fewer than 3 of the four are listed but not ranked.
- Value. Index points per blended dollar, using a 3:1 input-to-output token mix typical of coding and RAG workloads.
What this does not tell you: latency, rate limits, data residency or how a model handles your codebase. Use the index to pick a shortlist of two or three, then run them on twenty of your own real tasks.
Benchmarks we track
Terminal-Bench 4.1
Whether a model, driving an agent harness, can finish real multi-step tasks in a Linux terminal: building software, fixing environments, running and debugging code. The score is the share of tasks resolved.
See rankings Coding (human-judged) · 25% of indexWebDev Arena
Head-to-head votes from developers who ask two anonymous models to build the same web app and pick the better result. Votes are turned into an Elo-style rating.
See rankings Reasoning · 20% of indexARC-AGI-2
Novel visual puzzles that require inferring an abstract rule from a few examples. Each task is new, so memorised knowledge does not help.
See rankings General chat (human-judged) · 20% of indexText Arena
Blind head-to-head votes on everyday prompts, from writing and analysis to Q&A, turned into an Elo-style rating.
See rankings Interactive reasoning · not indexedARC-AGI-3
Interactive game-like environments the model has to explore, learn the rules of and solve, scored on the standard (no provider adapter) track.
See rankingsHead-to-head comparisons
Frequently asked questions
What is the best AI model in November 2026?
Claude Fable 5.2 from Anthropic leads the Diary of a CTO Index with 95.5 points, followed by Claude Opus 5.5 (95.3) and Gemini 4 Carbon (92.4). The top three are close enough that price, latency and your own task evaluations should decide between them.
Which LLM is best for coding agents?
On Terminal-Bench 4.1, which measures agents finishing real tasks in a terminal, Copilot Fusion has the top published score at 71.2%. Scores within about three points of each other are statistical ties.
What is the best value LLM?
GLM-5.3-Flash gives the most CTO Index points per blended dollar, at $0.02 / $0.07 per million input/output tokens. Value matters most for high-volume workloads such as classification, extraction and support.
What is the best open-weight model?
Kimi K3.1 Pro from Moonshot AI is the highest-ranked model whose weights you can download and self-host, released under the Kimi K3.1 Pro License.
Where does this leaderboard get its data?
Every score is taken from the benchmark's own public leaderboard: Terminal-Bench, ARC Prize and LMArena. We do not copy aggregator scores or run private evaluations. Prices and context windows are list prices as shown on LMArena on the snapshot date.
How is the CTO Index calculated?
Each benchmark is normalised so the best tracked model scores 100. Elo ratings become the expected win rate against the leader. We then take a weighted mean: Terminal-Bench 35%, WebDev Arena 25%, ARC-AGI-2 20%, Text Arena 20%. A model needs at least 3 of those four to be ranked.