Best model for autonomous agents
Which model should I use to build an agent?
As of 18 September 2026, the model ranking first for this question is Anthropic: Claude Fable 5.1 on OpenRouter, with agentic index 58.0 and an output price of $50.00 per million tokens. Ranked by highest agentic index, among models that support tool calling, across 116 models for which we hold both figures. This ranking ignores speed and reliability: it is a query on price and a relayed capability score, not a verdict on which model serves you best.
How this ranking is built
Criterion: highest agentic index, among models that support tool calling. An agent needs to call tools. Models without tool support are excluded regardless of their score, because they cannot do the job at all.
This is a query, not an opinion. Both figures are shown on every row so you can check the order, and the table sorts on any column if you disagree with our criterion.
Two things to know before trusting this table. Capability scores are published by Artificial Analysis and relayed by the catalogue we read. They are not our measurement, and a score measures the model rather than the seller, so we apply it to every seller of that model. This ranking says nothing about speed or reliability, which we measure only in part: availability is published per upstream host in reliability.json and on the model pages, and we publish no latency at all, for reasons given on the data page. Prices are ours to observe, each carries the date we saw it, and we keep only the cheapest offer per model so that batch and free tiers do not crowd out the answer.
| # | Model | Seller | Agentic index/ 10040 of 40 heldScale from 3.8 to 55 | Input$ / 1M tokens40 of 40 heldScale from $0.0001 to $600 | Output$ / 1M tokens40 of 40 heldScale from $0.0001 to $600 | Contexttokens40 of 40 heldScale from 512 to 2M | Observed |
|---|---|---|---|---|---|---|---|
| 1 | Anthropic: Claude Fable 5.1 | OpenRouter | 58.0 | $10.00 | $50.00 | 1 000 000 | 18 September 2026 |
| 2 | Anthropic: Claude Opus 5 | OpenRouter | 56.2 | $5.00 | $25.00 | 1 000 000 | 18 September 2026 |
| 3 | Qwen: Qwen3.8 Max (0902) | OpenRouter | 56.1 | $2.00 | $6.00 | 1 000 000 | 16 September 2026 |
| 4 | Z.ai: GLM 5.3 | OpenRouter | 53.4 | $1.40 | $4.40 | 1 310 720 | 18 September 2026 |
| 5 | SpaceXAI: Grok 4.6 | OpenRouter | 53.4 | $2.00 | $6.00 | 500 000 | 18 September 2026 |
| 6 | OpenAI: GPT-6 Astra | OpenRouter | 51.5 | $10.00 | $50.00 | 1 050 000 | 18 September 2026 |
| 7 | Z.ai: GLM 5.3 Flash | OpenRouter | 51.2 | $0.090 | $0.300 | 1 310 720 | 18 September 2026 |
| 8 | Anthropic: Claude Fable 5 | OpenRouter | 51.0 | $10.00 | $50.00 | 1 000 000 | 18 September 2026 |
| 9 | MoonshotAI: Kimi K3 | OpenRouter | 50.6 | $2.10 | $10.95 | 1 048 576 | 18 September 2026 |
| 10 | OpenAI: GPT-5.6 Sol | OpenRouter | 50.5 | $2.00 | $10.00 | 1 050 000 | 18 September 2026 |
| 11 | Qwen: Qwen3.8 2.4T A95B | OpenRouter | 50.4 | $2.00 | $6.00 | 1 048 576 | 18 September 2026 |
| 12 | Qwen: Qwen3.8 Max | OpenRouter | 49.9 | $2.00 | $6.00 | 1 000 000 | 4 September 2026 |
| 13 | Qwen: Qwen3.8 27B | OpenRouter | 46.5 | $0.214 | $2.55 | 1 000 000 | 18 September 2026 |
| 14 | Anthropic: Claude Sonnet 5 | OpenRouter | 44.3 | $2.00 | $10.00 | 1 000 000 | 18 September 2026 |
| 15 | OpenAI: GPT-5.4 | OpenRouter | 44.2 | $2.50 | $15.00 | 1 050 000 | 18 September 2026 |
| 16 | Meta: Muse Spark 1.2 | OpenRouter | 44.0 | $1.25 | $4.25 | 1 048 576 | 18 September 2026 |
| 17 | OpenAI: GPT-5.6 Terra | OpenRouter | 43.7 | $2.00 | $12.00 | 1 050 000 | 18 September 2026 |
| 18 | OpenAI: GPT-5.6 Luna | OpenRouter | 42.7 | $0.200 | $1.20 | 1 050 000 | 18 September 2026 |
| 19 | Anthropic: Claude Opus 4.8 | OpenRouter | 42.6 | $5.00 | $25.00 | 1 000 000 | 18 September 2026 |
| 20 | DeepSeek: DeepSeek V4 Pro 0813 | OpenRouter | 42.3 | $0.57816 | $1.73448 | 1 048 576 | 18 September 2026 |
| 21 | SpaceXAI: Grok 4.5 | OpenRouter | 42.1 | $2.00 | $6.00 | 500 000 | 18 September 2026 |
| 22 | DeepSeek: DeepSeek V4 Flash 0731 | OpenRouter | 41.7 | $0.060 | $0.120 | 1 310 720 | 18 September 2026 |
| 23 | Google: Gemini 3.8 Flash | OpenRouter | 41.1 | $0.750 | $3.75 | 1 048 576 | 18 September 2026 |
| 24 | Anthropic: Claude Opus 4.7 | OpenRouter | 39.5 | $5.00 | $25.00 | 1 000 000 | 18 September 2026 |
| 25 | Z.ai: GLM 5.2 | OpenRouter | 39.4 | $0.5544 | $1.7424 | 1 048 576 | 18 September 2026 |
| 26 | OpenAI: GPT-5.5 | OpenRouter | 37.3 | $5.00 | $30.00 | 1 050 000 | 18 September 2026 |
| 27 | Google: Gemini 3.7 Flash | OpenRouter | 36.4 | $0.750 | $3.75 | 1 048 576 | 18 September 2026 |
| 28 | Upstage: Solar Pro 4 | OpenRouter | 33.6 | $0.090 | $0.360 | 524 288 | 18 September 2026 |
| 29 | Anthropic: Claude Sonnet 4.6 | OpenRouter | 33.1 | $3.00 | $15.00 | 1 000 000 | 18 September 2026 |
| 30 | Nex AGI: Nex-N2-Pro | OpenRouter | 31.2 | $0.250 | $1.00 | 262 144 | 8 September 2026 |
| 31 | MiniMax: MiniMax M3 | OpenRouter | 30.8 | $0.300 | $1.20 | 1 048 576 | 18 September 2026 |
| 32 | Google: Gemini 3.6 Flash | OpenRouter | 30.2 | $0.750 | $3.75 | 1 048 576 | 18 September 2026 |
| 33 | inclusionAI: Ling 3.0 Flash VL | OpenRouter | 30.0 | $0.060 | $0.180 | 131 072 | 18 September 2026 |
| 34 | inclusionAI: Ling 3.0 Flash Fin | OpenRouter | 29.3 | $0.060 | $0.180 | 262 144 | 18 September 2026 |
| 35 | Qwen: Qwen3.6 Plus | OpenRouter | 29.0 | $0.325 | $1.95 | 1 000 000 | 18 September 2026 |
| 36 | SpaceXAI: Grok Build 0.1 | OpenRouter | 28.9 | $1.00 | $2.00 | 256 000 | 18 September 2026 |
| 37 | DeepSeek: DeepSeek V4 Flash 0423 | OpenRouter | 27.9 | $0.04984 | $0.09968 | 1 048 576 | 18 September 2026 |
| 38 | DeepSeek: DeepSeek V4 Pro 0423 | OpenRouter | 27.7 | $1.60 | $3.20 | 1 048 576 | 18 September 2026 |
| 39 | Meta: Muse Spark 1.1 | OpenRouter | 27.5 | $1.25 | $4.25 | 1 048 576 | 18 September 2026 |
| 40 | Google: Gemini 3.5 Flash | OpenRouter | 27.3 | $1.50 | $9.00 | 1 048 576 | 18 September 2026 |
116 models meet this criterion. The table shows the first 40, so 76 ranked below the cut are not listed here; the full catalogue holds them all.
A model missing from this table is missing one of the two figures, not judged. We would rather show a shorter list than fill it with guesses.