What changed vs Claude Fable 5.1
Context, max output, and the knowledge cutoff stay the same. Generation speed and measured coding, vision, and long-task results change.
Anthropic says token generation is up to 30% faster than Fable 5.1. That is output speed. At high effort the model thinks before almost every reply, so the wait for the first answer token is still long.
Anthropic classifies Fable 5.2 as a Covered Model. It requires 30-day retention at the provider and is not available under zero data retention unless Anthropic expressly authorizes it. Follow the Tess Privacy Policy and Anthropic’s retention guidance. Don’t send secrets, credentials, or regulated personal data unless the task needs them.
What Anthropic measured
The numbers below are from the announcement. The headline score is the highest available effort unless the cell says otherwise. Safeguards on.Benchmarks depend on the harness, the tools, the safeguards, and the effort. A small gap between frontier models does not always show up the same way in your flow. Validate before you change a production setup.
What Artificial Analysis measured
On the max variant, Artificial Analysis records an Intelligence Index of 65 and places the model at the top of the intelligence sample they published. The same run shows the other side of max effort: 221 seconds to the first answer token and 44.6 minutes per index task. Output speed was 102 tokens per second. That run produced 275 million output tokens.The 65 and the wait are for the max variant, not the high effort the API uses by default. Anthropic says most of the capability gain is already present at high, and that max is for measuring the ceiling.
Where it does not lead
- On AutomationBench, Gemini 4 Carbon scored 53.9%, above Fable 5.2’s 51.4%.
- On AA-Briefcase Elo, Opus 5.5 scored 1,800 and Fable 5.2 scored 1,785. On the same benchmark’s rubric, Fable leads: 66.1% versus 61.2%.
- Anthropic points to Opus 5.5 for most coding and knowledge work.
- Fable’s safeguards block most cybersecurity tasks. Claude Mythos 5.2 is the same model for organizations Anthropic has verified. It is not the model in the public picker.
- At medium effort and above, a line such as “answer directly, without deliberating” does not stop thinking. Fable 5.1 usually followed that kind of line.
Where it creates the most value
Code that Opus still does not close
The change crosses the repository, the tests, and several review rounds. Fable 5.2 is the next step when Opus 5.5, at a higher effort, still returns a diff that does not pass. Expected result: the change implemented, the tests for that step, and a list of what was left out.Research that has to become a document, spreadsheet, or deck
The work gathers sources, keeps the plan across tools, and ends in a deliverable. Anthropic describes this as a long agent session through to a finished document, spreadsheet, or presentation. Expected result: the file you asked for, with the source of each number and a written gap when a figure is missing.A chart or figure the model has to read
The Chartography jump, from 44.8% on Fable 5.1 to 73.4% on 5.2, is the case where the image is the data. The model returns what the figure shows and what it does not show. Expected result: value, unit, and the region of the figure each number came from.How to apply it in Tess
1
Move to Fable only when Opus falls short
Use Opus 5.5 for recurring work. Choose Fable 5.2 for the coding, research, or visual reading that Opus still does not close.
2
Define what done means
State the deliverable, what must not change, and the check that ends the task. On long tasks, ask it to finish the whole job: at low and medium effort the model sometimes stops with a list of what remains.
3
Say who owns the code
When the prompt says nothing about code ownership, Anthropic saw the model add to the existing structure and report nearby problems instead of fixing them.
4
Keep the thread append-only
Editing, reordering, or deleting an earlier turn invalidates thinking blocks on every account. Switching models mid-conversation also drops Fable 5.2 reasoning: only Fable 5.2 and Mythos 5.2 read those blocks.
5
Start at high effort
High is the API default, and Anthropic says most of the gain is already there. Step up to xhigh when you have measured a gain on coding or a long agent run. Reserve max for measuring the ceiling, when time and tokens do not matter.
Three playbooks to start
1. A code change a lighter model did not close
Input: the repository, the expected behavior, what already failed on Opus, and the test commands.2. Research with a deliverable
Input: the question, the sources, and the final format (document, spreadsheet, or slides).3. Chart reading
Input: the figure and the list of what to extract.Limitations and best practices
- Adaptive thinking stays on. Disabled thinking, or thinking with
budget_tokens, is not this model’s mode. - An instruction to think less does not turn thinking off at medium effort or above.
- A 1M context window does not replace selection. Send what the task uses.
- On long tool runs with little user-facing text, ask for a short status before each block of tools.
- Very hard style rules, such as a strict cap on words per sentence, make the prose choppy. The model follows the rule closely.
- The API offers a fast mode, with faster generation. This page does not confirm that mode in the Tess picker.
- Anthropic will not retire the model before 9 November 2027.
- Review code, numbers, and citations before you publish or run them in production. Keep a human approval for financial, legal, clinical, or high-impact decisions.