AI ranking
Cloud AI — agentic
Cloud models (Claude, GPT, Gemini, Grok, Qwen cloud, DeepSeek cloud, Mistral cloud…) suited to driving agents: models present in both reasoning (general) and code. These are models, not applications or scaffolds — OpenHands, SWE-agent, Aider, OpenCode, TRAE… stay under agentic systems.
Visual top 10
- 01 Claude opus 4.6 (high) 97.52
- 02 Claude opus 4.7 (high) 97.16
- 03 Claude opus 4.6 96.45
- 04 Claude opus 4.7 95.92
- 05 Claude Fable 5 High 95.36
- 06 Claude opus 4.8 (high) 93.62
- 07 Claude Fable 5 1 Max 93.11
- 08 Claude sonnet 5.5 (xhigh) 93.09
- 09 Gemini 3 pro 93.07
- 10 Deepseek V4.1 Flash 20260910 (max) 92.73
Compare models. Tick up to 3 models in the table below, then run the comparison: who wins on which use, by how much, and how solid the measurement is.
| # | Model | Vendor | Score | Monthly subscription | Assessment | |
|---|---|---|---|---|---|---|
| 01 | Anthropic | 97.52 / 100 | Not available | Partial | ||
| 02 | Anthropic | 97.16 / 100 | Not available | Partial | ||
| 03 | Anthropic | 96.45 / 100 | Not available | Partial | ||
| 04 | Anthropic | 95.92 / 100 | Not available | Partial | ||
| 05 | Anthropic | 95.36 / 100 | €23/month | Reliable | ||
| 06 | Anthropic | 93.62 / 100 | Not available | Partial | ||
| 07 | Anthropic | 93.11 / 100 | €23/month | Reliable | ||
| 08 | Anthropic | 93.09 / 100 | Not available | Partial | ||
| 09 | 93.07 / 100 | €22/month | Reliable | |||
| 10 | DeepSeek | 92.73 / 100 | Free | Partial | ||
| 11 | Anthropic | 92.38 / 100 | Not available | Partial | ||
| 12 | Anthropic | 92.37 / 100 | Not available | Partial | ||
| 13 | Anthropic | 92.02 / 100 | Not available | Partial | ||
| 14 | xAI | 92.01 / 100 | Free | Partial | ||
| 15 | 91.03 / 100 | €22/month | Reliable | |||
| 16 | xAI | 90.96 / 100 | Free | Partial | ||
| 17 | Mimo V2.5 Pro | Xiaomi | 90.95 / 100 | Not available | Partial | |
| 18 | DeepSeek | 90.78 / 100 | Free | Partial | ||
| 19 | xAI | 90.77 / 100 | Free | Partial | ||
| 20 | OpenAI | 89.72 / 100 | €23/month | Reliable | ||
| 21 | 89.54 / 100 | €22/month | Partial | |||
| 22 | xAI | 89.37 / 100 | Free | Reliable | ||
| 23 | Anthropic | 89.18 / 100 | Not available | Partial | ||
| 24 | DeepSeek | 89.01 / 100 | Free | Partial | ||
| 25 | DeepSeek | 88.80 / 100 | Free | Reliable | ||
| 26 | Moonshot AI | 88.12 / 100 | Not available | Partial | ||
| 27 | OpenAI | 87.59 / 100 | €23/month | Partial | ||
| 28 | xAI | 87.39 / 100 | Free | Reliable | ||
| 29 | Alibaba | 87.16 / 100 | Not available | Reliable | ||
| 30 | Z.ai (zhipu ai) | 87.04 / 100 | Not available | Reliable | ||
| 31 | Inkling | Thinky | 86.70 / 100 | Not available | Partial | |
| 32 | OpenAI | 86.25 / 100 | €23/month | Reliable | ||
| 33 | Anthropic | 86.23 / 100 | €23/month | Partial | ||
| 34 | DeepSeek | 86.17 / 100 | Free | Partial | ||
| 35 | OpenAI | 86.10 / 100 | €23/month | Reliable | ||
| 36 | 86.06 / 100 | €22/month | Reliable | |||
| 37 | 85.82 / 100 | Not available | Partial | |||
| 38 | Beam 1.0 501b a23b | Reflection | 85.64 / 100 | Not available | Partial | |
| 39 | 85.47 / 100 | €22/month | Reliable | |||
| 40 | Z.ai | 85.28 / 100 | Not available | Partial | ||
| 41 | Mistral | 85.27 / 100 | €15/month | Partial | ||
| 42 | DeepSeek | 85.26 / 100 | Free | Reliable | ||
| 43 | Meta | 84.93 / 100 | Not available | Reliable | ||
| 44 | Moonshot AI | 84.75 / 100 | Not available | Partial | ||
| 45 | Moonshot AI | 84.24 / 100 | Not available | Reliable | ||
| 46 | OpenAI | 84.14 / 100 | €23/month | Reliable | ||
| 47 | DeepSeek | 83.76 / 100 | Free | Reliable | ||
| 48 | Alibaba | 83.72 / 100 | Not available | Reliable | ||
| 49 | DeepSeek | 83.69 / 100 | Free | Partial | ||
| 50 | Meta | 83.68 / 100 | Not available | Partial |
Indicative score, weighted from public sources. Transparent methodology. See the AI ranking page for the full methodology and the other categories.
Cloud AI models suited to agentic use
This view gathers the cloud models capable of powering an agent: those that hold up in both general reasoning and code. These are models, not applications. Agentic scaffolds and environments — OpenHands, SWE-agent, Aider, OpenCode — belong to the Agentic systems ranking.
The distinction matters in practice. An agent's performance depends first on the model driving it, but an excellent model plugged into a poor scaffold produces poor results, and the reverse is true too. Separating the two avoids crediting the model with what comes from the tooling.
The score is the one from the general ranking: this view does not reassess anything, it filters. You will therefore find the models whose dual competence in reasoning and code makes them credible for long tasks, keeping the usual caveats of any ranking with partial coverage.
What this ranking is for
This view helps you choose the model that will drive an agent, setting aside those that excel in a single area and drop out as soon as a task mixes reasoning and code execution.
It is also a reminder that a model ranked well here does not guarantee a reliable agent: autonomy, supervision and tooling remain decisive factors, assessed separately.
How to read the ranking
- Simultaneous presence in general reasoning and in code (entry condition).
- Overall quality and reasoning, weighted as in the general ranking.
- Coding competence, decisive for executing real tasks.
- Cost and speed, which weigh heavily on long, repeated tasks.
- Stability, a condition for prolonged autonomous execution.
FAQ — Cloud AI — agentic
What is the difference with the AI agents ranking?
This view ranks models; the AI agents ranking assesses complete systems (agent plus tooling plus supervision). The same model can give very different results depending on the system using it.
Why is the score identical to the general ranking?
Because this is a filtered sub-view, not a reassessment. The filter is dual presence in reasoning and in code; the score remains the one from the general ranking.
Will a model ranked well here make a good agent?
It is a favourable condition, not a guarantee. The tooling, the permissions granted and the level of supervision change the final result radically.
Why are some well-known models missing?
Because they are present in only one of the two strands, reasoning or code. The filter requires both.
How often is it updated?
About every 25 hours, automatically. The update date is shown at the top of the page.