Fable 5.2 came out on November 9, 2026, and the question I keep getting is the same: is it better than ChatGPT and Gemini? Short answer: on independent leaderboards it ranks first, with Gemini 4 Carbon right behind. But it does not win everything, and the list price is not what you actually pay. Below is the comparison with the models from OpenAI, Google, xAI, the Chinese labs and the open-weight ones.
If you first want to know what Fable 5.2 is, I wrote about what changed from 5.1.
- Fable 5.265
- Gemini 4 Carbon60
- Claude Opus 5.559
- GPT-6 Astra57
- Kimi K3.1 Pro46
- Grok 4.735
- DeepSeek V4.133
- Mistral Large 426
Artificial Analysis, index v5.0.1, read on November 10, 2026. Each model at its best effort level.
Who the contenders are
A token is a piece of a word: a million tokens is roughly half a million words. The prices below apply when you use the models inside software (via API).
| Model | Made by | Input / output price ($ per million tokens) | Open weights |
|---|---|---|---|
| Claude Fable 5.2 | Anthropic (US) | 8 / 40 | No |
| Claude Opus 5.5 | Anthropic (US) | 4 / 20 | No |
| GPT-6 Astra | OpenAI (US) | 10 / 50 | No |
| GPT-6.1 Sol | OpenAI (US) | 2 / 10 | No |
| Gemini 4 Carbon | Google (US) | 2 / 10 until Nov 30, then 4 / 20 | No |
| Grok 4.7 | xAI (US) | 2 / 6 | No |
| Kimi K3.1 Pro | Moonshot (China) | 2 / 10 | Yes |
| Qwen3.8 Max | Alibaba (China) | 2 / 6 | No |
| GLM-5.3 | Zhipu (China) | 1.40 / 4.40 | Yes |
| DeepSeek V4.1 Flash | DeepSeek (China) | 0.30 / 1.20 (half price off-peak) | Yes |
Open weights means you can download the model and run it on your own server, without going through the vendor.
Who wins, test by test
A benchmark is the same exam given to every model. I mostly use independent leaderboards here, run by people who do not sell models. Numbers published by Anthropic are marked as such.
Overall rankings: Fable 5.2 wins
On Artificial Analysis, which combines nine different tests, Fable 5.2 ranks first out of 74 models with 65 points. On Vals, another independent leaderboard, it is first with 71.31%, ahead of Gemini 4 Carbon (70.45%) and GPT-6 Astra (65.02%).
Coding: Fable wins, but it depends on who measures
On the FrontierCode 2.0 autonomous coding test (measured by Artificial Analysis) Fable 5.2 scores 48.8%, against 42.4% for GPT-6 Astra and 40.6% for Gemini 4 Carbon.
On the terminal test (Terminal-Bench 4.0), though, the leaderboards disagree: Vals puts Fable first, tied with Opus 5.5, while Artificial Analysis puts Gemini 4 Carbon and the Chinese Kimi K3.1 Pro ahead of it. Same test, different rules, different result: that is why I never trust a single number.
Business automation: Gemini 4 Carbon wins
On AutomationBench, which measures office workflows, Gemini 4 Carbon is ahead: 77.8% against 76.4% on Artificial Analysis. Even Anthropic's own table admits it (53.9% against 51.4%). Gemini also accepts audio and video, while Fable and GPT-6 only read text and images.
Reasoning on new problems: it depends on the test
On ARC-AGI-3, new problems that cannot be memorized, Fable 5.2 scores 81.2% against 62.7% for GPT-6 Astra. On the previous version, ARC-AGI-2, Astra is ahead instead (95.0% against 93.8%).
What people prefer: Gemini, for now
On LMArena, where real people vote blind for the better answer, Gemini 4 Carbon is first. Fable 5.2 is too new to be ranked: version 5.1 is seventh.
The list price is misleading
This is the most useful part of the article. On paper Fable 5.2 is cheaper than GPT-6 Astra (8 / 40 against 10 / 50). But at maximum effort it reasons a lot and writes much more: in the Artificial Analysis test about 274 thousand tokens per job, against 52 thousand for Astra.
- Fable 5.2 (max)$26.54
- Gemini 4 Carbon$7.76
- GPT-6 Astra$6.24
- Fable 5.2 (medium)$3.80
- GPT-6.1 Sol$1.27
- DeepSeek V4.1$0.60
Artificial Analysis, November 10, 2026. Effort is how much the model reasons before answering.
The fix is simple: effort. At medium effort Fable 5.2 drops to $3.80 per job and 56 points, almost as good as GPT-6 Astra at maximum (57). And the time to finish a job goes from 44 minutes to 8.
The same goes for waiting: at maximum effort Fable takes almost 4 minutes on average before writing its first word. Fine for a long job, not for a chat reply.
Chinese and open-source models: are they worth it?
The Chinese models have one clear advantage: they cost very little. DeepSeek V4.1 Flash does the same test job for 60 cents, against $26 for Fable at maximum. The price shows in the scores, though: 33 points against 65.
The strongest Chinese model is Kimi K3.1 Pro: 46 points, and surprisingly good at terminal work. It is also the best open-weight model overall.
On open models there is a surprising fact: the best ones are all Chinese. The best European open model, France's Mistral Large 4, scores 26. And running a model like Kimi K3.1 Pro in-house, with 2.8 trillion parameters, takes data-centre hardware, not the office server.
Where your data goes
For a European business this matters as much as the scores, because of GDPR.
- Anthropic (Fable 5.2): through the direct API the model runs in the US or "globally", with no EU-only option. Fable 5.2 retains data for 30 days for safety monitoring.
- OpenAI: offers regional data residency, at a 10% premium.
- DeepSeek: its privacy policy says data is processed and stored in the People's Republic of China. In January 2025 the Italian data protection authority blocked its app in Italy.
- Kimi: servers in Singapore, and its privacy policy allows user content to be used to train models.
- Open-weight models: you can run them on your own or European servers, so data never leaves. But you need powerful hardware.
Which one to choose: my rule of thumb
There is no single winner. This is how I think about it:
- Hard jobs, now and then (a bug nobody can explain, an analysis that must be right the first time): Fable 5.2.
- Everyday work and automation: Claude Opus 5.5, Gemini 4 Carbon or GPT-6.1 Sol. They cost much less and the result is close.
- Lots of non-sensitive text to process (product descriptions, translations, summaries of public documents): a cheap model can be enough.
- Customers' personal data: only vendors with clear data-processing terms and, where possible, servers in Europe.
Fable 5.2 is one day old: some leaderboards (LMArena, Humanity's Last Exam, SWE-bench) have not measured it yet. I will update this article when the new numbers arrive.
My take
Fable 5.2 is the strongest model you can use today, but its lead over the others is smaller than the announcements suggest, and it gets expensive if you leave it at maximum all the time. The right choice is not "the best": it is the right model for each part of the job, with the data kept safe.
That is exactly what I do when I build an AI automation or a custom business tool: I choose the model for the task and the budget, not for the hype. If you want to know which one makes sense for you, write to me.
Sources, read on November 10, 2026: Artificial Analysis, Vals Index, Vals Terminal-Bench 4.0, ARC Prize, LMArena, Anthropic announcement, price lists from Anthropic, OpenAI, Google, xAI, DeepSeek, Z.ai, DeepSeek privacy policy.


