- Key facts
- When to use it
- How to prompt it: 8 rules
- 1. Pick the model before the first message
- 2. Default to high effort, and change effort instead of model
- 3. Give it the whole task up front
- 4. Ask for updates during long runs
- 5. Ask for shorter, clearer reports
- 6. Tell it to finish the work
- 7. Audit the rules you've piled up
- 8. Ask for proof, not reasoning
- What else is in Anthropic's guide
- Why it matters for me
- Further reading
Claude Fable 5.2 is Anthropic's most capable model for general use, released on 9th November 2026 together with Claude Mythos 5.2. Both are the same underlying model. Fable carries strong cyber and biology safeguards and is open to everyone. Mythos drops them for organisations verified through Anthropic's cyber and life sciences programmes.
It costs twice as much as Opus 5.5 and follows instructions far more literally than earlier models. So 2 questions matter: when is it worth the money, and which of your existing prompts now work against it. This note combines Anthropic's launch material and prompting guide with the tests and tips in Ben AI's video.
Key facts
| API id | claude-fable-5-2 (Mythos: claude-mythos-5-2) |
| Price | $8 per million input tokens, $40 per million output. Fable 5.1 was $10 / $50 |
| Fast mode | Up to 2.5x faster generation, at $16 / $80 |
| Context | 1M tokens, 128k max output |
| Thinking | Adaptive thinking always on. The effort level sets the depth |
| Speed | Output up to 30% faster than Fable 5.1 |
| Availability | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and the Pro, Max, Ultra, Team and Enterprise plans |
Selected benchmarks from the launch post, Anthropic's own numbers:
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| FrontierCode 2.0 (coding) | 49.1% | 32.4% | 45.5% |
| AutomationBench (workflows) | 51.4% | 31.4% | 42.5% |
| Remote Labor Index (real work) | 41.3% | 17.9% | 21.3% |
| Humanity's Last Exam (tools) | 74.2% | 65.6% | 67.7% |
| Chartography (charts, vision) | 73.4% | 44.8% | 64.4% |
The Remote Labor Index is the "almost twice Opus" result the video quotes. The cost curves matter as much: at high effort, Fable 5.2 matches the best Opus 5.5 coding score for about half the cost per task. At low effort, it matches Opus 5.5's best Humanity's Last Exam score for under a tenth of the cost.
When to use it
Anthropic is direct about this. Fable is for "the most ambitious, open-ended technical work". For most coding and knowledge tasks, start with Opus 5.5, "our strongest daily driver", and move up only when Opus at a higher effort still falls short.
Ben AI ran both models at high effort on 3 of his own jobs, with the same prompts:
| Job | Opus 5.5 | Fable 5.2 | His verdict |
|---|---|---|---|
| Video editing and motion graphics | 35 min, $21 | 51 min, $35 | Fable much cleaner, closer to the style reference |
| Landing page redesign | 22 min, $10 | 38 min, $18 | Fable close to ready to publish, more animation |
| Marketing report as a deck | 21 min, $20 | 32 min, $34 | Similar. Opus more creative, but more metric errors |
In total, Fable used about 70% more ($87 against $51). His rule of thumb follows Anthropic's:
- Use Fable for big, open-ended jobs you hand over and walk away from, where accuracy matters: long research, a financial model, a report where every number must be right, a security audit. Visual and design work is where 5.2 improved most.
- Stay on Opus 5.5 for everyday work, writing and back-and-forth editing. It's faster and half the price.
- Before you switch to Fable after a bad output, raise the effort level on the current model first.
On the subscription plans (from the video, check your own plan): Fable can use at most half of the weekly limit on Max, and burns it faster than other models. Pro and standard Team include it until 16th November 2026, then it runs on paid usage credits.
How to prompt it: 8 rules
The video's 8 tips, each checked against Anthropic's prompting guide.
1. Pick the model before the first message
Switching away from Fable mid-conversation throws away its reasoning: other models can't read Fable 5.2's thinking blocks, so they only see the messages. Each model also keeps its own prompt cache, so a switch in either direction rereads the whole conversation at full price. A cached message can be up to 40x cheaper, according to the video.
To change model, start a new chat with a short summary, or hand the part off to a subagent.
2. Default to high effort, and change effort instead of model
Anthropic's guide per level:
| Effort | Use it for |
|---|---|
low |
Quick back-and-forth, well-specified tasks, fast subagents |
medium |
Chat, interactive coding, everyday knowledge work |
high |
Hard coding, research, agentic work, long unattended runs. The API default |
xhigh |
The hardest long runs, where you've measured a gain |
max |
Open-ended science and maths, when time and tokens don't matter |
Most of the gain is already at high. Above it, gains are uneven. A mix works well: plan at low or medium, then raise effort for the long part. Changing effort keeps the cache, changing model doesn't.
At high and above, Fable thinks before almost every reply, even "hi". A line like "answer directly without deliberating" no longer stops it. Lower the effort for those turns instead.
3. Give it the whole task up front
Step-by-step prompting wastes this model. It's built for long, end-to-end runs. Ben's 4-part checklist:
- The job: the whole task, not the first step.
- The why: what it's for and who it's for. On a long task the model makes 100s of small decisions, and the intent lets it make them your way.
- The rules: spend 1 minute on what could go wrong. Fable follows rules almost word for word, so write the exceptions out too.
- What done looks like: so it knows when to stop. These models now tend to do too much, not too little.
He dictates prompts with a voice tool and walks through the 4 parts as he talks.
4. Ask for updates during long runs
Fable 5.2 can go quiet for long stretches of tool calls. Anthropic's one-line fix for a system prompt or CLAUDE.md:
Every few tool calls, give the user a one-line update on where things stand.
In the Claude Code desktop app, a side chat can also ask what the main chat is doing without interrupting it.
5. Ask for shorter, clearer reports
Final reports can come back dense and over-technical. Anthropic's fix opens with "Write for an intelligent but low-context reader" and asks for the key points first, plain words, no invented terms, and mostly ASD-STE100 Simplified Technical English. The full block is in the guide. The video's alternative is a short custom output style.
6. Tell it to finish the work
Fable 5.2 sometimes stops early with a list of what's left, more often at low and medium. Anthropic published a "Finishing work" block for the system prompt. It starts:
Please avoid stopping while you still owe work the user asked for. Be persistent and do not stop unless needed.
The rest tells it to work through ordinary errors and timeouts, and to leave deliberate blockers alone. In Anthropic's tests it halved the sessions that ended with work left, for about 30% more output tokens. Keep your own confirmation step for risky actions, because the model asks less.
7. Audit the rules you've piled up
This is the one that matters most for a long-lived setup. Fable takes instructions literally, so old cautions in CLAUDE.md, memory and skills now bite:
- "Never retry" or "stop at the first error" blocks autonomous work.
- "Never X" is absolute. Write "ask before deleting files", not "never delete files".
- Hard style limits ("1 instruction per sentence, 20 words or fewer") make the prose choppy.
- Standing instructions to delegate can spawn more subagents than the task needs.
- A token budget shown in the prompt changes how much it spends. Removing a large one cut cost with the same results.
The model writes some of these rules itself over time. Ben's fix: run /doctor in Claude Code, ask for a prompt audit against the Fable 5.2 guide, and paste in the guide's link.
8. Ask for proof, not reasoning
Asking Fable to explain its reasoning can trigger a refusal: Anthropic doesn't let it reproduce its internal thinking in the answer (refusal category reasoning_extraction). In testing, it also sometimes stated guesses as facts, so the explanation would prove little.
Ask for evidence instead. Not "explain how you got these numbers", but "for every number in this summary, show me the exact source and the rows it came from". Each number can then be checked in seconds.
What else is in Anthropic's guide
3 points the video skips:
- Say who owns the code. "You are the owner and long-term maintainer of this codebase" raised the rate at which it merged duplicated code from under 10% to about 75%, with 70 to 80% more output tokens and changes nobody asked for. Leave it out when the code is someone else's.
- Be specific when you approve. Reply "Yes, run the migration on staging", not "go".
- Don't edit earlier turns over the API. Fable 5.2 keeps its thinking from turn to turn. A harness that rewrites the system prompt, the tools or an earlier message invalidates that thinking, and the Messages API now returns an error. Keep the history append-only, or use compaction. Claude Code and claude.ai handle this for you. Details in What's new in Fable 5.2.
Why it matters for me
My NicAI setup is exactly the case rule 7 warns about: a long CLAUDE.md, 100s of memory notes, dozens of skills, many of them written as "never" rules. Some are real hard limits and should stay that way. Others are old cautions without the reason written next to them. Before I run Fable on long jobs, the setup needs the audit.
Rule 5 is a nice confirmation too: Anthropic's own report prompt asks for ASD-STE100, the style I already ask Claude to use with me. Rule 7 is the caveat: only as a style, never as hard word counts.
For daily work, Opus 5.5 stays the default (see Claude Opus Best Practices). Fable gets the long, unattended jobs where a wrong number costs more than the tokens.
Further reading
The launch post, with the full benchmark table:
The prompting guide, with every prompt block quoted in this note:
The API changes and breaking changes from Fable 5.1:
The safety testing and evaluation details: