Anthropic cuts Claude Fable 5.2 price by 20%, and on the coding test it nearly matches Opus 5.5
Anthropic has cut the price of Claude Fable 5.2 by 20%: one million input tokens now cost $8 instead of $10, and one million output tokens cost $40 instead of $50. On the FrontierCode 2.0 programming benchmark, the model scored 45.3%, close to Opus 5.5 at maximum effort (45.5%), and the task cost about half as much. That figure is the cost of the whole task, not the price per token, and per token Opus 5.5 is cheaper. Calculate the actual bill on your own workloads.
AI-processed from Хабр: Машинное обучение; edited by Hamidun News
Anthropic has released Claude Fable 5.2 and cut its price by 20%: one million input tokens (what you send to the model) now cost $8 instead of $10, and one million output tokens (what it writes in response) cost $40 instead of $50. On the FrontierCode 2.0 programming benchmark, Fable 5.2 solved 45.3% of tasks. The senior model, Opus 5.5, solved 45.5% at its maximum effort level. The gap is two-tenths of a percentage point, and according to Anthropic, completing a task cost roughly half as much. It sounds like a simple case of "take Fable and save money," but there's a catch, and it's worth discussing.
Two models on one foundation
Fable 5.2 and Mythos 5.2 are built on the same base model. They differ in their safety restrictions. Fable is available to ordinary users and developers. Mythos is intended for organizations that have passed Anthropic's vetting, meaning those who need broader capabilities under vendor control.
For most teams, this means one simple thing: they will work with Fable. Mythos remains a niche option, for example for companies with special requirements and a separate contract with Anthropic.
Pricing and Fast Mode
Fable 5.2 is faster and cheaper than the previous version. The 20% reduction applies to both input and output tokens: $10 became $8, and $50 became $40.
Take a hypothetical request with 1 million input and 200 thousand output tokens. On the old version it would cost about $20 ($10 for input plus $10 for output). Now it comes to about $16 ($8 plus $8). For an application handling millions of requests per month, that difference is significant.
Fast Mode is a separate option. Generation runs up to 2.5 times faster, but the price is double: $16 per million input tokens and $80 per million output tokens. The same hypothetical request would now cost about $32. Paying for speed makes sense where a person is waiting for a response right now, such as an assistant inside an IDE or a customer-facing chat. For background jobs like an overnight code review, the standard mode is cheaper.
What FrontierCode 2.0 showed
FrontierCode 2.0 is a benchmark, meaning a set of programming tasks used to compare models. Fable 5.2 in High effort mode (where the model spends more effort on reasoning) scored 45.3%. Opus 5.5 in Max effort mode scored 45.5%. A gap that small is smaller than any reasonable margin of error. In other words, the more affordable model solves tasks almost as well as the more expensive one.
Now for the catch. The "roughly twice as cheap" saving refers to the cost of completing the task as a whole, not to the price of tokens. Opus 5.5 is actually cheaper than Fable 5.2 at API list prices. So a token costs less on Opus, but per solved task, by Anthropic's estimate, Opus comes out more expensive. The material does not show exactly where the difference comes from, but the logic is clear: the final bill depends not only on the price per token, but also on how many tokens the model spends on a task, how many steps it takes, and how many times it retries.
What to do in practice
First, don't take the word "cheaper" at face value without checking. Anthropic's figure comes from a single benchmark. On your codebase, with your prompts and tools, the ratio may be different.
Second, run several dozen real tasks through both Fable 5.2 and Opus 5.5. Compare not only the share of solved tasks but also the total cost of each run: how many tokens each task consumed and how many retries were needed. That metric determines the budget.
Third, treat the effort level as a separate cost setting. The 45.3% result was achieved at High effort. At a lower level the model will probably be faster and cheaper, but quality may drop. This also needs to be tested on your own data.
A few limitations. The comparison rests on Anthropic's own claims and one benchmark; there are no independent measurements of task cost yet. But the direction is clear: the model market is gradually shifting from a race for the highest score to a competition over the price of a solved task. Teams that know how to measure that price will come out ahead.
- Nov 21, 2026California Sues OpenAI: Its AI Agents Escaped Control at Least 80 Times
- Nov 21, 2026Dreamteam Raises $53M for a CRM That Fills Itself from Email and Calls
- Nov 21, 2026Claude Opus 5.5 Builds Out Dimension 9 Horror Game From the Author's 15 Rooms
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
Key Questions Answered
What are the key details on Two models on one foundation?
Fable 5.2 and Mythos 5.2 are built on the same base model. They differ in their safety restrictions. Fable is available to ordinary users and developers. Mythos is intended for organizations that have passed Anthropic's vetting, meaning those who need broader capabilities under vendor control. For most teams, this means one simple thing: they will work with Fable. Mythos remains a niche option, for example for companies with special requirements and a separate contract with Anthropic.
What are the key details on Pricing and Fast Mode?
Fable 5.2 is faster and cheaper than the previous version. The 20% reduction applies to both input and output tokens: $10 became $8, and $50 became $40. Take a hypothetical request with 1 million input and 200 thousand output tokens. On the old version it would cost about $20 ($10 for input plus $10 for output). Now it comes to about $16 ($8 plus $8). For an application handling millions of requests per month, that difference is significant. Fast Mode is a separate option.
What FrontierCode 2.0 showed?
FrontierCode 2.0 is a benchmark, meaning a set of programming tasks used to compare models. Fable 5.2 in High effort mode (where the model spends more effort on reasoning) scored 45.3%. Opus 5.5 in Max effort mode scored 45.5%. A gap that small is smaller than any reasonable margin of error. In other words, the more affordable model solves tasks almost as well as the more expensive one. Now for the catch.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.