On this page
Claude Fable 5.2 replaces Fable 5.1 and costs less. Anthropic released it today, 9 November 2026. Input now costs $8 per million tokens, down from $10. Output costs $40, down from $50. On the company's own test of long coding jobs, the new model scores 49.1%. The old one scored 32.4%. Its safety filters also block about 55% fewer requests.
If you already run Fable 5.1: change the model ID. You pay less for a stronger model.
If you are weighing it against Opus 5.5: start on Opus, as the vendor advises, and move up only for the jobs it fails.
If you need inference kept in Canada: neither version offers that. The route is an open-weight model on dedicated GPUs in a region inside the country.
Key facts
| Claude Fable 5.1 | Claude Fable 5.2 | |
|---|---|---|
| Released | 1 Sep 2026 | 9 Nov 2026 |
| Model ID | claude-fable-5-1 | claude-fable-5-2 |
| Input, per million tokens | $10 | $8 |
| Output, per million tokens | $50 | $40 |
| Long coding jobs (vendor test, highest effort) | 32.4% | 49.1% |
| Safety-filter blocks | Baseline | About 55% fewer |
| Sibling model for vetted organisations | Mythos 5.1 | Mythos 5.2 |
Fable 5.2 prices, scores and the block figure come from the launch post, so they are the vendor's own numbers. Fable 5.1 prices come from its own launch post.
What changed in Fable 5.2?
Three things changed: the price fell by a fifth, the coding and agent scores rose sharply, and the safety filters refuse far less often. The model is also quicker. The launch post says answers arrive "up to 30% faster" than before, and an optional fast mode writes up to 2.5 times quicker at double the price.
The job the model is meant for stayed the same. The launch post still pitches it at "the most ambitious, open-ended technical work" and still recommends Opus 5.5 as the place to start for most coding. CNBC's report quotes the company's research product lead calling it a model for "power users or power use cases", such as engineering, mathematics and robotics.
The release also includes Mythos 5.2, the same model with lighter safeguards for vetted organisations. A section further down covers it.
How much does Fable 5.2 cost compared with Fable 5.1?
The new model costs 20% less per input token and 20% less per output token. The launch post gives only the new prices, $8 and $40 per million tokens. The old prices were $10 and $50, so both lines fall by a fifth.
| Per million tokens | Fable 5.1 | Fable 5.2 | Opus 5.5, for reference |
|---|---|---|---|
| Input | $10 | $8 | $4 |
| Output | $50 | $40 | $20 |
| Cache read | $0.25 | not in launch post | $0.20 |
| Fast mode, input / output | not in launch post | $16 / $80 | $8 / $40 |
Each column comes from that model's launch post: Fable 5.1, Fable 5.2 and Opus 5.5, as of 9 November 2026. A token is about three quarters of a word.
Here is what the cut does to one coding job, the same one we priced in our earlier post. The model reads 250,000 tokens of repository context, nine tenths of it from cache. It then writes 30,000 tokens. We assume both versions use the same number of tokens, which a newer model may not. The launch post does not give a cache price for 5.2, so the totals leave the cached tokens out.
| One attempt | Fable 5.1 | Fable 5.2 |
|---|---|---|
| Input, 25,000 fresh | $0.250 | $0.200 |
| Output, 30,000 | $1.500 | $1.200 |
| Total before cache reads | $1.75 | $1.40 |
| Input, 225,000 cached | $0.056 | not in launch post |
Output makes up over four fifths of the bill in both columns, so the output price is the cut that counts. The real saving depends on how many attempts each model needs to finish the job. That is why we price models per solved task, not per token.
What do the coding numbers say?
On the vendor's test of long, multi-step coding jobs, the new model finishes about half and the old one about a third. Nobody outside the company has reproduced these runs yet.
| Task (vendor figures) | Fable 5.1 | Fable 5.2 | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Long agentic coding jobs, highest effort | 32.4% | 49.1% | 45.5% (Max effort) | 42.8% |
| Long agentic coding jobs, High effort | not given | 45.3% | 45.5% (Max effort) | n/a |
| Long, multi-file software engineering | 56.3% | 69.9% | 62.3% | 65.5% |
| Automating business workflows end to end | 31.4% | 51.4% | 42.5% | 41.4% |
| Scientific computing in a terminal | 52.6% | 74.6% | 58.7% | 68.1% |
| Very hard expert questions, with tools | 65.6% | 74.2% | 67.7% | 57.2% |
The first row is the headline. At its highest effort setting, the new model leads Opus 5.5 by 3.6 points. Its lead over OpenAI's flagship is 6.3 points. A lead of three or four points on a test the vendor ran itself is real but small.
The second row matters more for a budget. The launch post says 5.2 at high effort scores 45.3%, about level with Opus 5.5 at its top setting, for "roughly half the cost per task". Per token, 5.2 costs twice as much as Opus. So the claim holds only if 5.2 reaches the answer in far fewer tokens. Check it on your own jobs before relying on it.
Two notes on reading the table. Our Fable 5.1 post listed 65.0% on the expert-questions test; the new launch table gives the older model 65.6%, and we quote the newer figure. And every scientific-computing score here comes from version 0.1 of that test. A score from another version will not line up, as we explain in why test versions matter.
Fable 5.2 vs Mythos 5.2: what is the difference?
Fable 5.2 and Mythos 5.2 are one model with two sets of guardrails. The launch post calls them "the same underlying model". Fable carries strong safeguards on cybersecurity and biology and is open to everyone. Mythos goes only to organisations the vendor has approved for cyber or life-science work. Both cost the same.
The safeguards changed more than anything else for daily users. In early testing across the company's apps, 5.2 blocks about 55% fewer requests than 5.1. The drop is about 60% for cybersecurity requests and about 40% for biology.
A blocked request still gets an answer, from an older model. Opus 4.8 answers a blocked cybersecurity question, and Opus 5 a blocked research biology question. If your team does security work, find out which of your requests trip the filter, so you know which answers came from the weaker model.
Should you upgrade from Fable 5.1 to Fable 5.2?
Yes, if you already use 5.1. The new version is cheaper per token, scores higher on every test published for both, and refuses less. Switching means changing the model ID from claude-fable-5-1 to claude-fable-5-2.
Three checks before you move everything:
- Rerun your own test cases. Vendor tests measure the vendor's tasks. A stronger model can still change the format or length of answers your code reads.
- Try high effort before the top setting. If the cost-per-task claim holds on your work, high effort gives most of the gain for less.
- Ask whether you need this tier at all. The launch post points most coding work at Opus 5.5, which costs half as much per token. Our cost-per-task method shows how to compare the two on your own jobs.
We have not run either version on our own task sets. Every score on this page is the vendor's. The launch pages do not state the new model's context window, so we leave it out.
Can you use Claude Fable 5.2 with data kept in Canada?
No. Fable 5.2 is a closed model, so your prompts are processed on servers run by Anthropic or its cloud partners. No published route keeps its inference inside Canada.
The cloud routes do not close the gap. Amazon's in-country Bedrock profile, launched on 27 October 2026, keeps requests inside the country, but it covers only the small Haiku 5.5 model. Our residency guide walks through each provider, including options that store data in the country but process prompts elsewhere.
For inference in the country at this level, the route is an open-weight model on dedicated GPUs. The closest this month is Kimi K3.1 Pro, whose weights Moonshot published in the first week of November. You can download them from its model page.
- Size. It has 2.8 trillion parameters, with about 104 billion used per token. The files come to roughly 1.56 TB.
- Hardware. The model card names three serving engines (vLLM, SGLang and TokenSpeed) but no GPU count, so your workload sets the hardware.
- Licence. The model's own licence allows commercial use and changes. Products with over 100 million monthly users, or over $20 million in monthly revenue, must show the model's name in their interface. Resellers of model access earning over $20 million a year need a separate agreement. Internal use is exempt from both rules.
- Where it trails. On the scientific-computing test, version 0.1, its maker reports 38.1%. The new Fable scores 74.6% by its own vendor's count. The gap is large.
- Trade-offs. Graphics chips are harder to get in ca-central-1 and ca-west-1, the two in-country regions, than in larger ones. You pay for them while they sit idle, and you run the serving yourself.
XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account, or on-premises, so prompts and outputs stay in the country. Then we build the workflow on top. See how we deploy and the Canadian residency guide for the provider-by-provider detail.
If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.
FAQ
Is Claude Fable 5.2 cheaper than Fable 5.1?
Yes. Input falls from $10 to $8 per million tokens. Output falls from $50 to $40, the same 20% cut. On our example coding job, one attempt drops from about $1.75 to $1.40 before cache reads, if both use the same number of tokens. Prices are from Anthropic's launch posts as of 9 November 2026.
Is Fable 5.2 better than Opus 5.5 for coding?
On the vendor's own test of long coding jobs, Fable 5.2 at its highest effort scores 49.1%. Opus 5.5 scores 45.5%. The gap is small and nobody has reproduced it yet. Anthropic still recommends starting most coding work on Opus 5.5, which costs half as much per token, and moving up when it falls short. Test both on your own jobs before you decide.
What happens when Fable 5.2 blocks a request?
The request goes to an older model instead of being refused. Opus 4.8 answers cybersecurity questions, and Opus 5 answers research biology questions. In early testing, the new version blocks about 55% fewer requests than 5.1. Vetted organisations can apply for Mythos 5.2, the same model with lighter safeguards.
What is the model ID for Claude Fable 5.2?
The API model ID is claude-fable-5-2, replacing claude-fable-5-1. The launch post says the model is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, and in the Claude apps. Fast mode costs $16 per million input tokens and $80 per million output, and is offered on fewer of those routes.
Can I run Claude Fable 5.2 on my own servers?
No. The weights are not released, so the model runs only on the vendor's platform and its cloud partners. If you need a model on your own hardware or in your own cloud account in the country, pick an open-weight model such as Kimi K3.1 Pro. You can download it and serve it yourself under its own licence terms.
Sources
- Introducing Claude Fable 5.2 and Claude Mythos 5.2, Anthropic, 9 Nov 2026 (benchmarks: FrontierCode 2.0, FrontierSWE v2, AutomationBench, Terminal-Bench-Science 0.1, Humanity's Last Exam with tools)
- Introducing Claude Fable 5.1 and Claude Mythos 5.1, Anthropic, Sep 2026 (Fable 5.1 prices)
- Claude Opus 5.5, Anthropic, 22 Sep 2026 (Opus 5.5 prices)
- Anthropic releases Claude Fable 5.2 and Mythos 5.2, CNBC, 9 Nov 2026
- Claude Haiku 5.5 cross-region inference in Canada, AWS, 27 Oct 2026
- Kimi K3.1 Pro model card and licence, Moonshot AI on Hugging Face, revision of 5 Nov 2026 (commit d9e99b0)
Written by
Cho Yin Yong
Principal AI Solutions Engineer, XY Space
Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.
More from Cho Yin YongShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project