A lot changed in the two weeks since our October ranking. Artificial Analysis rebuilt its Intelligence Index, Google announced Gemini 4 Carbon, OpenAI released the cheaper GPT-6.1 Terra, and Mistral finally published the weights of Mistral Large 4. This guide uses the new index, version 5, so scores are lower than last month and not directly comparable.
What changed in the leaderboard
Artificial Analysis launched Intelligence Index v5 on 28 October. About 60% of it now uses private tests or hidden answers, because labs can train on public benchmarks and inflate their scores. It adds two new tests, FrontierCode 2.0 for real-world coding and Terminal-Bench-Science for scientific tasks, and replaces Humanity's Last Exam with a harder subset. The result: lower, more spread-out scores, and a wider gap between the best closed and open models.
Source: Artificial Analysis via BenchLeader, 3 November 2026. Best reasoning setting per model. Kimi K3.1 Pro's weights are not yet released; Mistral Large 4 is shown at high effort.
Prices and cost per task
Token prices only tell half the story, because models use very different amounts of tokens to finish the same work. Where Artificial Analysis has published a cost per index task, we show it.
| Model | Input / output per million tokens | Cost per index task | Can you use it? |
|---|---|---|---|
| Gemini 4 Carbon | $2 / $10 introductory | $7.76 | Cyber defenders only |
| Claude Opus 5.5 | $4 / $20 | $10.04 | Yes |
| GPT-6 Astra | $10 / $50 | Not published | Yes |
| GPT-6.1 Sol | $2 / $10 | About a sixth of Carbon | Yes |
| Claude Sonnet 5.5 | $2 / $10 | Not published | Yes |
| GPT-6.1 Terra | $1 / $5 | $0.96 | Yes |
| Gemini 4 Argon | $2 / $10 until 30 Nov, then $4 / $20 | Not published | Yes |
| Claude Haiku 5.5 | $0.10 / $0.50 | Not published | Yes |
| Mistral Large 4 | $0.70 / $2.10, or free to self-host | Not published | Yes, open weights |
Who leads at what
- Overall
- Gemini 4 Carbon (not yet available to most), then Claude Opus 5.5.
- Real-world coding (FrontierCode 2.0)
- Claude Opus 5.5 at 46%, then GPT-6 Astra, GPT-6.1 Terra and Claude Sonnet 5.5 at 42%.
- Science in the terminal
- GPT-6 Astra at 63%, then Claude Opus 5.5 at 59%.
- Hard reasoning (CritPt v2)
- GPT-6 Astra at 72%, then GPT-6.1 Sol at 69%.
- Office-style tasks (AA-Briefcase)
- Claude Sonnet 5.5 leads.
- Business automation (AutomationBench)
- Gemini 4 Argon leads at 78%.
- Fewest made-up facts (AA-Omniscience)
- Gemini 4 Carbon at 54, ahead of Claude Opus 5.5 at 48.
Open weights: downloadable, but further behind
Two big open-weight stories this month. Mistral Large 4, about 1 trillion parameters, was published on Hugging Face on 29 October under a modified MIT licence that allows commercial use, except for companies selling it as a hosted model service. On the tougher index it scores 26, behind several Chinese open models. Kimi K3.1 Pro from Moonshot AI scores 46, and Moonshot plans to release its weights in early November. When it does, it becomes the best model you can run yourself, still about 12 points behind the leader.
Closed models
Best quality- Top of every leaderboard
- No hardware to manage
- Your data goes to the provider, unless you have zero-retention terms
Best for complex analysis, coding and agents.
Open weights
Full control- Run on your own servers; data never leaves
- Very cheap through hosted providers
- 12 or more points behind the leader on the new index
- The largest models need racks of top-end GPUs
Best for private data and simple, high-volume tasks.
Which model for which job
- 1
Best quality you can use today
Claude Opus 5.5 at high or max effort, especially for coding and long analysis.
- 2
Best value for a team
GPT-6.1 Sol or Claude Sonnet 5.5 at $2 / $10, both near the top.
- 3
High-volume agents and automation
GPT-6.1 Terra at $1 / $5, the cheapest model in the top ten per task.
- 4
Cheap classification and extraction
Claude Haiku 5.5 or an open model such as DeepSeek V4.1 Flash.
- 5
Private or regulated data
Mistral Large 4 on your own infrastructure, or wait for Kimi K3.1 Pro's weights.
- 6
Security reviews
Ask about Gemini 4 Carbon access through Google's Fairwind Program.
Check your understanding
1/3Why are November scores lower than October's?
Version 5 uses about 60% private tests and harder questions, so scores dropped across the board.
2/3Which model is cheapest per task among the top ten?
Terra costs about $0.96 per index task at max effort.
3/3What does Mistral Large 4's licence restrict?
Use, fine-tuning and commercial use are allowed; model-as-a-service providers need a commercial licence.
You got 0 out of 3.
Go deeper: Gemini 4 Carbon, GPT-6.1 Terra, Mistral Large 4 and Claude Haiku 5.5. Last month's edition: the best AI models in October 2026.
Sources: Artificial Analysis Intelligence Index v5 via BenchLeader (3 November 2026) and OfficeChai's coverage of the v5 launch (28 October 2026); Tom's Hardware on Gemini 4 Carbon costs (3 November 2026); Mistral AI and Hugging Face on the Mistral Large 4 open weights (29 October 2026); provider price lists. Prices and scores change often; check each provider before committing.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 3 November 2026. About me