This comparison is built from vendor documentation, published pricing and independent measurements from Artificial Analysis and others. I have not run all three models on a shared workload, so every benchmark below is attributed, and vendor-reported numbers are labelled as such.
Claude Fable 5.2 is the strongest model you can rent this month, and it is cheaper than the one it replaced. Anthropic shipped it on 9 November 2026 at $8 per million input tokens and $40 output, 20% below Fable 5.1, and Artificial Analysis now ranks it first on its Intelligence Index.
But “strongest” is not the same as “right for you”. Moonshot’s Kimi K3.1 Pro gives you downloadable weights and a score within ten points for a quarter of the token price. Mistral Large 4 trails both on raw capability, yet it is the cheapest of the three by a distance and the only one with a European lineage and a clear self-hosting story. Here is how to pick, with the numbers that matter and the ones you should treat with suspicion.
Table of Contents
- Claude Fable 5.2, Kimi K3.1 Pro and Mistral Large 4 at a Glance
- What Claude Fable 5.2 Changes
- Kimi K3.1 Pro: Open Weights Near the Frontier
- Mistral Large 4: The European Open Model
- Benchmarks: Claude Fable 5.2 Against the Field
- Cost per Task Beats Price per Token
- Self-Hosting and Licences
- Who Should Pick Claude Fable 5.2 and Who Should Not
- FAQ About Claude Fable 5.2 and Its Rivals
- Verdict
- Sources and Further Reading

Claude Fable 5.2, Kimi K3.1 Pro and Mistral Large 4 at a Glance
| Claude Fable 5.2 | Kimi K3.1 Pro | Mistral Large 4 | |
|---|---|---|---|
| Maker | Anthropic | Moonshot AI | Mistral AI |
| Released | 9 Nov 2026 | API 23 Oct, weights 4 Nov 2026 | Preview 6 Oct, weights later in autumn 2026 |
| Weights | Closed | Downloadable | Downloadable |
| Size | Not disclosed | About 2.8T total, 104B active (MoE) | About 1T total, 49B to 52B active (MoE) |
| Context | 1M tokens, 128K output | 1M tokens | 1M tokens |
| Input / output per 1M tokens | $8 / $40 | $2 / $10 | $0.70 / $2.10 |
| Cached input | $0.20 | $0.20 | $0.07 |
| Artificial Analysis Intelligence Index (v5) | 65 at max effort | 55 | 26 |
Two warnings before you read anything into that table. First, the Intelligence Index changed version in November, and v5 scores are not comparable with older v4 numbers you may have bookmarked. Second, Mistral’s API price and parameter counts have shifted between its preview and open-weight announcements, so check its pricing page before you budget.
What Claude Fable 5.2 Changes
Claude Fable 5.2 is Anthropic’s general-release Mythos-class model. It shares its weights with Claude Mythos 5.2, which carries fewer safeguards and is limited to vetted organisations. For most developers, Fable is the one you can actually call.
The price cut is the headline, and every related rate moved with it. Cache reads fell to $0.20 per million tokens, batch processing is $4 input and $20 output, and there is a new fast mode at $16 and $80 that generates tokens up to 2.5 times faster, according to Anthropic’s figures as compiled by ClaudeLog. Fast mode was not available on Fable 5.1.
Speed improved on the standard tier too. Anthropic says output is up to 30% faster than Fable 5.1 and that the model completes tasks with fewer tokens at medium effort, which is the default in the Claude apps and Claude Code. On the API, the default effort is high.
There are migration traps. Adaptive thinking is always on, and manual budget_tokens is no longer accepted, so code that set a thinking budget explicitly will need changing. BuildMVPFast counts four API changes in total that can break working code. If you are upgrading from Fable 5.1, our Claude Fable 5.1 guide covers the earlier breaking changes, and you will want to recheck both lists together.
Inside Anthropic’s own products the switch is already done. Claude Fable 5.2 became the default in Claude Code from version 2.1.321, and saved Fable 5.1 selections move to the general fable alias, subject to admin settings on Team and Enterprise plans. If you need the old model for a regression comparison, /model claude-fable-5-1 brings it back. On the subscription side, BuildMVPFast reports that Pro and standard Team seats can use Fable for up to half of their weekly limits, so heavy users will still hit ceilings.
Retirement is not a near-term worry. Anthropic’s models overview lists Claude Fable 5.2 as not retiring sooner than 9 November 2027, which gives teams a year of stability to build on it.
Anthropic also reports about 55% fewer safeguard blocks than Fable 5.1 in early testing, with roughly 60% fewer in cybersecurity and 40% fewer in biology. Exploit generation and penetration testing still route to an Opus model.
Kimi K3.1 Pro: Open Weights Near the Frontier
Moonshot released Kimi K3.1 Pro through its API on 23 October 2026, and published the weights on 4 November, on the schedule it had promised. It is a post-trained version of Kimi K3 rather than a model trained from scratch, built with what Moonshot describes as substantially larger post-training compute.
Size is the first thing people notice. The K3 family is a mixture-of-experts design with about 2.8 trillion total parameters and roughly 104 billion active per token. That makes it the largest downloadable model available, and also one of the hardest to run yourself.
The price is the second thing. At $2 input and $10 output, with cache hits at $0.20, Kimi K3.1 Pro costs a third less per token than Kimi K3 did. Moonshot also charges for writing to the cache: $2 per million for a five-minute cache and $4 for an hour.
Moonshot made one unusual decision for the open release. It says it identified and removed cyber-related data from both training stages while keeping general coding tasks, and that its offensive cyber capability sits below frontier proprietary models. For a model anyone can download, that is a reasonable trade.
A faster sibling, K3.1 Pro Highspeed, runs two to three times faster with the same capabilities, according to Moonshot. Artificial Analysis measured the standard model at around 73 to 77 output tokens per second through Kimi’s API, which is on the slow side for its class.
Mistral Large 4: The European Open Model
Mistral Large 4, nicknamed “Le Chonk”, entered public preview on 6 October 2026 and moved to open weights afterwards. Reports differ on the exact date the weights went up, so check Mistral’s own open-weight announcement if the timing matters to you. Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres.
It is roughly a trillion parameters with about 49 billion active per token, rising to 52 billion if you count embeddings. Mistral’s own pages have quoted both figures without explaining the change, which is a small but telling sign of how fast this release moved.
On capability, Mistral Large 4 is a tier below the other two. Its Artificial Analysis Intelligence Index v5 score of 26 is the highest among Western open-weight models, but Chinese open models score far higher. Its strength is coding efficiency: Artificial Analysis puts it at 48% on its Coding Agent Index while using 47% fewer tokens per task on average.
The price is where it wins outright. At $0.70 input, $0.07 cached and $2.10 output per million tokens, it is less than a tenth of Claude Fable 5.2 on input and about a twentieth on output. Reasoning tokens bill as output, so heavy thinking narrows that gap a little.
Benchmarks: Claude Fable 5.2 Against the Field
Most published numbers come from the vendors themselves. Anthropic’s benchmark table for Claude Fable 5.2 has not been reviewed by any external organisation, and Cognition’s FrontierCode benchmark is private, so nobody outside can rerun it. Read the table with that in mind.
| Benchmark | Claude Fable 5.2 | Kimi K3.1 Pro | Mistral Large 4 | Source |
|---|---|---|---|---|
| AA Intelligence Index v5 | 65 (max), 62 (high) | 55 | 26 | Artificial Analysis |
| FrontierCode 2.0 | 49.1% (xhigh) | Not published | 19.7% | Vendors |
| Terminal-Bench 4.0 | Not published here | 67.2% | 41% | Vendors |
| FrontierSWE v2 | 69.9% | Not published | Not published | Anthropic |
| Remote Labor Index | 41.3% | Not published | Not published | Anthropic |
The only row where all three sit on the same independent scale is the Artificial Analysis index, and it tells a clear story. Claude Fable 5.2 leads, Kimi K3.1 Pro is a solid fourth among all models, and Mistral Large 4 plays in a different league.
The Kimi Terminal-Bench figure comes from Alibaba’s Qwen4-Max launch chart, which listed K3.1 alongside its rivals, so treat it as a competitor’s reported number. For Mistral, Terminal-Bench jumped from 28.3% in preview to 41% at the open-weight release, which suggests the extra weeks of reinforcement learning paid off.
Cost per Task Beats Price per Token
Token prices make Claude Fable 5.2 look four times more expensive than Kimi and nearly twenty times more than Mistral. Cost per finished task tells a more interesting story.
Artificial Analysis measures what each model spends to complete its index. Claude Fable 5.2 at low effort costs about $3.10 per task, the cheapest Fable setting, and prices vary up to 8.6 times across its effort levels. At the default high effort, Capital and Compute reports a score of 62 for $9.08 a task, against Claude Opus 5.5 at max effort scoring 59 for $10.04. So the bigger model beats Opus on both quality and cost at its default setting.
Kimi K3.1 Pro costs about $2.11 per index task. That is cheap for its score, but note what happened against its own predecessor: the token price fell by a third while the model used roughly 14% more output tokens per task, so cost per task barely moved.
The lesson applies to all three. A model that thinks for longer on a cheaper token can cost the same as a pricier model that answers briefly. Measure on your own workload. Our guide to cutting LLM API costs shows how to set up that comparison, and caching alone can change the ranking for agent loops with long, repeated prompts.
Self-Hosting and Licences
Claude Fable 5.2 is API-only. You can reach it through Anthropic, Amazon Bedrock, Google Cloud and Microsoft Foundry, but you cannot download it.
Kimi K3.1 Pro’s weights are on Hugging Face. Licence reports differ: one tracker lists Apache 2.0, while others describe a Kimi K3.1 Pro licence. Read the licence file in the repository before you build a product on it. Hardware is the bigger obstacle. At 2.8 trillion parameters, even aggressive quantisation leaves you needing a multi-GPU server, and as of mid-November no community GGUF conversion had appeared.
Mistral Large 4 ships under the Mistral Ecosystem Open-Weight Licence, a modified MIT licence. Internal use is fine, but if you or an affiliate sell it as a hosted model service, you need a separate commercial licence. Mistral’s official examples for vLLM, SGLang and Transformers all assume eight GPUs with tensor parallelism of eight, and there is an NVFP4 checkpoint aimed at Blackwell cards.
In practice, neither open model is a desktop project. If you want something you can run locally, our guide to the best local LLMs covers models that fit on one or two consumer GPUs, and LLM inference servers compared covers how to serve the big ones if you do have the hardware.
Who Should Pick Claude Fable 5.2 and Who Should Not
| If you are | Start with | Why | Link |
|---|---|---|---|
| Running long, high-stakes agent tasks | Claude Fable 5.2 at high effort | Top independent score and lower cost per task than Opus 5.5 at max | Claude pricing |
| Needing open weights with near-frontier quality | Kimi K3.1 Pro | Fourth on the AA index, downloadable, $2 / $10 via API | Kimi on Artificial Analysis |
| Building in Europe or on a tight budget | Mistral Large 4 | Lowest token price, token-efficient coding, European lineage | Mistral announcement |
| Upgrading from Fable 5.1 | Claude Fable 5.2, after a regression run | 20% cheaper, but the API changes can break thinking settings | Fable 5.1 guide |
A mixed setup often makes more sense than picking one. Route the hardest planning steps to Claude Fable 5.2 and let a cheaper model handle routine tool calls and summaries. That is the same pattern that makes most agent stacks affordable, and it works with any of these three.
Related on this site: for more context, see Claude Opus 5.5 and Mistral Large 4.
FAQ About Claude Fable 5.2 and Its Rivals
How much does Claude Fable 5.2 cost?
$8 per million input tokens and $40 per million output tokens, with cache reads at $0.20 and batch at half price. Fast mode costs $16 and $80 and runs up to 2.5 times faster.
Is Claude Fable 5.2 better than Opus 5.5?
On Artificial Analysis, yes. At its default high effort it scores 62 for about $9.08 a task, against Opus 5.5 at max effort scoring 59 for about $10.04. Opus 5.5 is still cheaper per token at $4 and $20.
Can I download Claude Fable 5.2?
No. It is available only through Anthropic’s API and cloud partners. If you need downloadable weights, Kimi K3.1 Pro and Mistral Large 4 are the closest options.
Which is the best open-weight model right now?
On the Artificial Analysis index, Kimi K3.1 Pro leads the downloadable models with a score of 55. Mistral Large 4 is the strongest Western open model but scores much lower.
Will my Fable 5.1 code work on Claude Fable 5.2?
Maybe not. Adaptive thinking is always on and manual thinking budgets are rejected, among other changes. Test your prompts and parameters before switching production traffic.
Can I run Kimi K3.1 Pro or Mistral Large 4 on a gaming PC?
Not realistically. Mistral’s official examples assume eight data-centre GPUs, and Kimi is nearly three times larger. Both are server-class models.
Verdict
Pick Claude Fable 5.2 if output quality on hard, long tasks is what you are paying for. It leads the only independent index all three share, and at its default effort it is cheaper per completed task than Anthropic’s own Opus 5.5 at max. The 20% price cut makes the upgrade from Fable 5.1 easy to justify.
The caveat is that most of the dramatic benchmark gains are Anthropic’s own numbers, unreviewed and in some cases on private tests. And if you need weights you control, Claude Fable 5.2 is simply not an option, which makes Kimi K3.1 Pro the realistic frontier alternative and Mistral Large 4 the budget and sovereignty pick.
Your next step: take twenty real tasks from your own logs, run them through Claude Fable 5.2 at high effort and Kimi K3.1 Pro through its API, and compare cost per correct answer, not cost per token. Add Mistral Large 4 if your work is mostly coding. One afternoon of that will tell you more than this whole table.
Sources and Further Reading
- ClaudeLog’s Fable 5.1 vs 5.2 breakdown, for pricing, fast mode, benchmarks and migration notes
- Artificial Analysis Fable 5.2 release page, for the index scores and cost per task by effort level
- Capital and Compute on Kimi K3.1 Pro, for the price history and cost-per-task comparison
- BrainDetox on the Mistral Large 4 weights, for licence terms, hardware and changes since preview
- Unite.AI on the Kimi K3.1 launch, for Moonshot’s cyber data removal and release details
For more model coverage on this site, read our Claude Fable 5.1 guide, the GPT-6 Astra breakdown and GLM-5.3-Flash. For a wider view, see AI models compared.

