Claude Fable 5.2 vs Opus 5.5: When to Use Which

Claude Fable 5.2 vs Opus 5.5: what Anthropic announced, what I measured running the same tasks through the API, and which model to pick for which job.
Anthropic has announced Claude Fable 5.2 and Mythos 5.2. The published benchmarks beat Opus 5.5 on nearly every item, and API pricing drops to $8 input / $40 output. This post covers what was announced, what happened when I pushed the same tasks through the API myself, and how I would split work between the two models.
The short answer first: Opus 5.5 is still enough for everyday work, and Fable 5.2 earns its keep on jobs that run for hours. Benchmark rank and real-world cost-effectiveness turned out to be two separate questions.
What Claude Fable 5.2 and Mythos 5.2 actually are
Fable and Mythos are the same underlying model. What differs is how they are released. Mythos is gated behind review for cyber and life-science work; Fable is the one open to everyone. Research use sits inside the review process, and ordinary day-to-day use is served by Fable.
Four things to know before you start:
- The API model ID is claude-fable-5-2
- Pro and Team standard seats can spend up to 50% of their weekly limit on Fable
- Output is up to 30% faster than Fable 5.1
- API pricing drops to $8 input / $40 output
That 50% cap reads like a footnote, but it shapes rollout. You cannot assume every person on the team will keep Fable running all week. You have to decide which jobs go to Fable before you hand out access.
Reading the official benchmark table
The numbers that stood out to me:
- FrontierCode 2.0 coding: 49.1% (Opus 5.5: 45.5%)
- OSWorld 2.1 computer use: 89.3%
- Humanity’s Last Exam with tools: 74.2%
- Chart reading: up from 44.8% on 5.1 to 73.4%
Top of the table on nearly every row, but not a clean sweep. On AutomationBench, which covers business workflows, Gemini 4 Carbon is ahead at 53.9%. And every figure here is vendor-measured. Whether any of it transfers to your own work is something only your own tasks can tell you.
The biggest single jump is chart reading. Going from 44.8% to 73.4% is the kind of move that changes the success rate of reading figures and dashboards out of documents. If you tried that before and gave up, this row alone is worth re-testing.
Max is no longer the sensible default
The effort chart is the part that matters most in practice. High-effort Fable 5.2 lands at about the same score as Max-effort Opus 5.5, at roughly half the cost per task. The top score is 49.1% at Xhigh, and pushing from there to Max moves the point slightly down on the chart while cost jumps about 3x.
So “just set it to Max” is no longer the right reflex. Start at the default Medium, raise to High when you get stuck, and treat Max as an exception you reach for knowing what it does to the bill. Designing in that order lowers your total spend.
What I found running the same tasks through the API
Announcement numbers alone are not enough to decide with, so I sent the same tasks to the API. Four kinds of work, including coding, math problems, and building a spreadsheet engine in Claude Code.
- Every configuration got every answer right; the only differences were cost and time
- Opus 5.5 Medium cost 60-80% of Fable 5.2 Medium
- Opus 5.5 Max spent up to 13x the cost and over 200 seconds for the same correct answer
- In Claude Code, Opus finished in about 50 seconds and Fable in about 70
The conclusion is that accuracy did not separate them. On one-shot and mid-sized problems there were even cases where Opus 5.5 was both cheaper and faster. Anthropic itself says Fable shows its strength on long, difficult work, and my comparison sits below that threshold, so the two statements do not contradict each other.
Put differently, I found no reason to rip out a workflow that already runs fine on Opus 5.5. The question that decides a migration is not “does accuracy improve” but “do I have work that needs to run autonomously for hours.”
Where the gap shows up: long autonomous runs
Three examples from the announcement show how different the long end looks.
In math, 1,000 Fable 5.2 agents fully formalized the proof of Fermat’s Last Theorem in 30 hours — rewriting it into a form a computer can verify one step at a time. The same setup on Opus 5.5 reached 26% at 18 hours and was still at 27% at 30 hours. It was a single run, but the shape of the gap is unmistakable.
Robot control improved sharply too. In the RoboDojo simulation it scored 64.9%, above Opus 5.5 at 39.4% and GPT-6 Astra at 29.8%, and above many models built specifically for robotics. It is reported to drive a dual-arm robot even on tasks it was not trained on. The trend of a language model driving a robot directly has reached this benchmark as well.
In biology, Mythos 5.2 found a new way to predict the whole structure of capsids — the shells that carry drugs in gene therapy. It modeled 28 of the 34 it attempted accurately, with the stated possibility that researchers could design capsids faster and more cheaply. Use cases like this are why the gated Mythos exists.
What the three share is that they run for a long time without a human sitting on top of them. A request that finishes in seconds will not reveal the difference; work that runs for hours or days will. That is where I would draw the line between the two models.
How this lands in Japan
One piece of local context for readers outside Japan: in the Japanese companies I work with, adopting a new model usually means getting internal budget sign-off first, and the people reviewing that paperwork care about predictable total cost far more than peak capability.
That is exactly why this release matters here. The common reaction to a benchmark table in Japan is to jump straight to the top model at the top effort setting. These numbers argue against that: if Max scores slightly lower than Xhigh while costing about 3x, turning the dial up is not a quality guarantee.
Three things I would focus on when rolling this out in a Japanese organization:
- Compare on your own tasks. Send the same problem to several models and several effort levels, then line up accuracy, cost and time
- Set Medium as the default, with a written rule that you only raise it when you are stuck
- Decide which jobs go to Fable first, given the 50% weekly-limit condition
If you are an overseas AI company communicating into Japan, change the order of emphasis. “Which setting is cheapest” and “what you do not have to change in your current setup” land better with Japanese practitioners than a first-place overall score. Leading with the fact that it is not a clean sweep, and that the figures are vendor-measured, also buys trust over the long run.
Putting this to work
The practical steps are simple. Move your default back to Medium and take Max out of everyday use. Then find one job inside your organization that you want running autonomously for hours, hand only that to Fable 5.2, and keep records of what it costs and how long it takes. If no such job exists, the only things this release changes for you are the price and the chart-reading accuracy.
The more models there are, the more the choosing itself decides the outcome. Time spent measuring on your own tasks pays off faster than time spent reading leaderboards.
FAQ
Should I switch from Opus 5.5 to Claude Fable 5.2?
Not for everyday work. When I ran the same four tasks through the API, every configuration got every answer right and the only differences were cost and time. Fable 5.2 earns its place on jobs that run autonomously for hours.
Which effort setting should I use with Fable 5.2?
Start at the default Medium and raise it to High only when you get stuck. High-effort Fable scores about the same as Max-effort Opus 5.5 at roughly half the cost per task, and the top score is 49.1% at Xhigh — pushing to Max nudges the point slightly down while cost jumps about 3x.
What is the difference between Fable and Mythos?
They are the same underlying model. Mythos is gated behind review for cyber and life-science work, and Fable is the generally available one. The API model ID is claude-fable-5-2, and Pro and Team standard seats can spend up to 50% of their weekly limit on it.
Based on a post on XOriginal post ↗