
Current AI Models: Which One to Pick for Each Task
Comparison guide to 12 AI models as of November 15, 2026: release date, context, price, open or closed weights and which one to try first for your task.
Which models are the most recent?
This sheet covers the 12 models that have their own page in the wiki. Prices are USD per million tokens, standard input and output rates.
| Model | Lab | Date | Context | Input / output | Weights |
|---|---|---|---|---|---|
| Claude Fable 5.2 | Anthropic | Nov 9, 2026 | 1M | 8 / 40 | Closed |
| Claude Opus 5.5 | Anthropic | Sep 22, 2026 | 1M | 4 / 20 | Closed |
| Claude Haiku 5.5 | Anthropic | Oct 7, 2026 | 1M | 0.10 / 0.50 (0.50 / 2.50 above 100,000 prompt tokens) | Closed |
| GPT-6 Astra | OpenAI | Sep 3, 2026 (system card) | 1,050,000 | 10 / 50 (up to 272K input) | Closed |
| Gemini 4 Carbon | Nov 10, 2026 on the API | 1M | 2 / 10 until Nov 30; 4 / 20 from Dec 1 | Closed | |
| Gemini 4 Hydrogen | Nov 12, 2026 | 1M | 1.50 / 7.50 | Closed | |
| Grok 4.7 | xAI | Sep 21, 2026 | 500K | 2 / 6 (4 / 12 from 200K) | xAI does not announce downloadable weights |
| Muse Ember | Meta | Nov 9, 2026 | 1M | 0.10 / 0.50 (0.30 / 1.50 above 200K input) | Closed |
| DeepSeek V4.1 Flash | DeepSeek | Sep 10, 2026 (version 1017: Oct 17) | 1M | 0.15 / 0.60 off-peak (0.30 / 1.20 at peak), uncached input | Open, MIT |
| Qwen3.8-Max | Alibaba | Aug 3, 2026 | 1M | 2 / 6 (Singapore and US) | Weights under its own license |
| Mistral Large 4 | Mistral | Oct 6, 2026 (preview) | 1M | 0.70 / 2.10 | Open weights, Mistral license |
| Kimi K3.1 Pro | Moonshot | not confirmed (weights by Nov 4, 2026 at the latest) | 1M | not confirmed | Open weights, its own license |
How do they compare on price and context?
- Context: almost all list 1M tokens. GPT-6 Astra reaches 1,050,000 and Grok 4.7 has 500K.
- Cheapest: Haiku 5.5 and Muse Ember, both positioned by their labs for repetitive, high-volume work. See small vs. large models.
- Most expensive: Fable 5.2 (8 / 40) and GPT-6 Astra (10 / 50), presented for the most demanding work.
- Price depends on prompt length: Haiku 5.5, GPT-6 Astra, Grok 4.7 and Muse Ember charge more once the prompt passes a threshold, and it applies to the whole request. Gemini 4 Hydrogen says it keeps a flat price across its full million tokens.
- Price depends on date or time of day: Carbon goes up on December 1 and DeepSeek charges half off-peak.
To turn this into a real bill, read cost per token.
Which model fits each task?
Each row is a wiki suggestion to try first, backed by a fact from its page. It is not a quality ranking.
| If you need… | Start with… | Why (fact from the wiki) |
|---|---|---|
| To classify or extract from thousands of texts | Haiku 5.5 or Muse Ember | 0.10 / 0.50 and positioned for high volume |
| A default model for code and long agents | Opus 5.5 | Anthropic recommends it as the starting point; 4 / 20 |
| More power if Opus 5.5 falls short | Fable 5.2 | Anthropic suggests it after trying Opus at high effort |
| A lot of material in one request | Hydrogen or Astra | Hydrogen: flat price across 1M; Astra: 1.05M, with a surcharge above 272K |
| Code inside Cursor | Grok 4.7 | Available in Cursor; 2 / 6 under 200K prompts |
| Cybersecurity or long engineering work | Carbon | Google positions it there; price rises on December 1 |
| Fast everyday agents | Hydrogen | Google calls it its daily model; 1.50 / 7.50 |
| Weights under the MIT license | DeepSeek V4.1 Flash | Weights on Hugging Face; 0.15 / 0.60 off-peak |
| Open weights with a published price | Mistral Large 4 | 0.70 / 2.10 and 1M context |
| Code with weights and image input | Qwen3.8-Max or Kimi K3.1 Pro | Qwen: 2 / 6; Kimi: its price is not in its announcement or documentation |
Which are open and which are closed?
- Closed, API or product only: the three Claude models on the list, GPT-6 Astra, both Gemini 4 models and Muse Ember. For Grok 4.7, xAI does not announce downloadable weights.
- With downloadable weights: DeepSeek V4.1 Flash (MIT), Qwen3.8-Max (its own license), Mistral Large 4 and Kimi K3.1 Pro (its own license).
Open does not mean unconditional. Qwen and Kimi ask for a separate agreement if you offer the model as a service to third parties at high revenue (Qwen: over 50 million dollars in 12 months; Kimi: 20 million). Mistral Large 4 has two license descriptions (modified MIT and Apache 2.0): ask legal to review the actual file. More in open vs. closed models.
Why can’t I compare their benchmarks directly?
Because each lab measures with its own harness, its own settings and, sometimes, a different version of the same benchmark. Examples from the pages:
- The same name gives very different numbers. DeepSeek publishes 35.4 on Agents’ Last Exam; Meta publishes 61.6 for Muse Ember, on a Linux subset it describes as an internal evaluation.
- Versions differ. Alibaba reports Terminal Bench 2.1; other labs report Terminal-Bench 4.0.
- Settings change the number: reasoning effort, with or without tools, the harness used.
Use scores to see a model’s trend against its own previous version, in the same table from the same lab. To choose between labs, run your own tests: how to read AI benchmarks.
Example
Illustrative example: a junior designer needs to tag user comments and does not know which model to use. A plan to test three:
- Gather 30 real comments and write the correct tag for each by hand.
- Pick three: a cheap one (Haiku 5.5), a default one (Opus 5.5) and one from another lab (Hydrogen).
- Run the same prompt on all three and note hits, tokens used and cost using this guide’s prices.
- Review the misses: did it invent tags or fail on ambiguous cases?
- Keep the cheapest one that is accurate enough, and recheck prices in a month.
Common mistakes
- Choosing by list price without measuring accuracy on your data.
- Treating first place on a lab’s table as the best for your case.
- Comparing scores from different labs as if they were the same exam.
- Missing that a long prompt changes the rate for the whole request.
- Assuming open means free of conditions.
- Budgeting with today’s prices and never looking again.
Sources
- Anthropic: Claude Fable 5.2
- Anthropic: Claude Opus 5.5
- Anthropic: Claude Haiku 5.5
- Anthropic: Claude pricing
- OpenAI: GPT-6 Astra
- OpenAI: API pricing
- Google: Gemini 4 Carbon is now available
- Google: introducing Gemini 4 Hydrogen
- Google AI Studio: pricing
- xAI: Grok 4.7
- xAI: models and pricing
- Meta: Muse Ember announcement
- Meta: pricing and rate limits
- DeepSeek: API models and pricing
- Alibaba Cloud: Model Studio pricing (Qwen3.8-Max)
- Hugging Face: Qwen3.8-2.4T-A95B
- Mistral: Mistral Large 4
- Moonshot AI: Kimi K3.1 Pro