Claude Fable 5.2 lands at No. 1 on an independent ranking, priced 20% lower

On November 9, 2026, Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2. One day later, on November 10, Fable 5.2 appeared in 1st place among 74 models on the Artificial Analysis Intelligence Index, with 65 points. Gemini 4 Carbon scored 60 and Claude Opus 5.5 scored 59. This measurement is independent; it does not come from Anthropic.

For anyone deciding budgets at a mid-sized company, the practical question is just one: does this change the bill and the routine of people who already use AI day to day? It depends on what you use today. Let's get to the facts.

What changed from 5.1 to 5.2

Fable 5.1 arrived in September 2026, about two months earlier. The new version brings three changes that matter to whoever pays the bill.

Lower price per token. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. 5.1 charged $10 and $50. That is a 20% drop. A token, here, is the unit these services bill by: roughly speaking, chunks of text going in (what you send) and coming out (what the AI answers).

More speed. Anthropic says responses come out up to 30% faster than 5.1's. There is also an optional fast mode, with generation up to 2.5 times quicker, billed at double the standard rate: $16 for input and $80 for output. 5.1 did not have this option.

Greater capability, according to the company. In the tests published by Anthropic, Fable 5.2 scored 69.9% on FrontierSWE v2, which measures agents on programming tasks, versus 56.3% for 5.1. On reading charts and figures without auxiliary tools (Chartography), it went from 44.8% to 73.4%. These are self-reported numbers, measured by the manufacturer itself.

Fable and Mythos: the same model, with different locks

Anthropic says Fable and Mythos are the same model underneath. The difference is in the filters. Fable applies extra restrictions to high-risk cybersecurity requests and to biology research. Direct access to Mythos 5.2 is restricted to verified organizations, such as those in the company's cybersecurity and life sciences programs.

For most mid-sized companies, Fable is the one that matters. If your business is not a laboratory or a security consultancy, Mythos probably doesn't enter the conversation.

One detail to consider: queries in these sensitive areas continue to be routed to other models, Opus 4.8 and Opus 5. Anthropic says this results in about 55% fewer blocks than on Fable 5.1. When the handoff happens, the next turn is slower and more expensive, and extended reasoning may not carry over to the destination model.

"20% cheaper" is not the same as "20% smaller bill"

This is the point that deserves the most attention from whoever controls costs. The 20% discount applies to the price per token. The final bill depends on how many tokens each task consumes, and that varies with the level of effort the model applies.

A third-party analysis, published on the martincid site, calculated that, depending on the effort level, a task can end up 56% more expensive ($0.446 versus $0.285). The methodology is not detailed in the material, so treat it as a warning, not a verdict. But the warning is useful: list price is not operating cost.

There is also a comparison with the competition, citing Fortune via Exame: on common tasks, Fable 5.2 costs 20% less than OpenAI's GPT-6 Astra, and, when reading large volumes of documents with caching (memory that reuses what has already been read), the difference reaches 80%. The absolute prices of GPT-6 Astra and Gemini 4 Carbon are not in the material I reviewed.

What about Opus 5.5, which is cheaper still?

Opus 5.5 costs $4 for input and $20 for output, half the per-token price of Fable 5.2. Meanwhile, Fable 5.2's batch processing (Batch API, for tasks that don't need an immediate response) comes in at $4 and $20, the same standard price as Opus 5.5, according to a secondary source. For overnight routines, such as classifying invoices or reviewing accumulated contracts, this can be relevant.

How to test without getting a shock

If your company already uses Claude for code, document analysis, or agents that execute tasks, switching versions tends to be simple, because the model identifier in the API is claude-fable-5-2 and the rest of the flow can stay as is. But "tends to" is not enough to decide. In practice, it works like this:

1. Pick three to five real tasks, the kind your team does every week: answering a quote, checking a spreadsheet, summarizing a contract.
2. Run the same tasks on both versions and compare three things: quality of the answer, time until it's ready, and total cost per task.
3. Test at more than one effort level, since that is where the third-party calculation points to a risk of the bill going up.
4. Decide per task, not for the whole company. It may be worth using 5.2 where quality matters and a cheaper option where it doesn't.

This test fits in a week and costs little compared to what is spent deciding in the dark. The cost of any larger project depends on usage volume and process design, and can only be estimated by looking at the specific case.

Where to use it and how long the bonus lasts

Fable 5.2 has been available since November 9 on Anthropic's platforms and on the Amazon Web Services, Google Cloud, and Microsoft Azure clouds. In the Claude apps, access is available on the paid plans: Pro, Max, Ultra, Team, and Enterprise. The company has not said whether the free plan will receive the model.

There are limited-time launch offers, with extra Fable usage for subscribers on eligible plans. A third-party site cites November 16 as the end date, without official confirmation. If you're going to test, take advantage of the window, but check the terms on Anthropic's support page.

Exame recommends the model for more complex, open-ended technical work. Simple, repetitive tasks may not justify the upgrade, and that is what the test with your own cases will show.