Claude Fable 5.2 in Claude Code: Advisor, Fast Mode and Agents

👁 113 reads

At Uber, a security audit of one system took 10 to 12 hours. It was carried out not by one person or one model, but by around 170 reviewer agents led by Claude Fable 5.2. They produced a report of their findings, ready-to-use pull requests with fixes, and a proposed new policy. This was one of the examples Anthropic used to illustrate Fable 5.2’s launch on November 9, 2026. It shows what the model is intended to do in Claude Code: not just write a single function, but lead an entire team.

I reviewed the announcement, the Claude Code and Claude API documentation, and the system card. There are several changes, and one of them, advisor, may matter more to the average developer than the new model itself.

Fable 5.2 in the terminal: version, effort and cost

Fable 5.2 requires Claude Code version 2.1.321 or later. Its default effort level in Claude Code and Claude apps is Medium. The API differs: its default is high, so if you call Fable from your own code, you should set effort manually. The model has a one-million-token context window and can generate responses of up to 128,000 tokens.

Fable is not free with a subscription. In the Pro plan, after the one-week promotion (November 9-16), its use draws on usage credits. Before you use it for the first time, Claude Code asks you to approve this billing method once: enter /model fable and confirm. In Max and Ultra, Fable is included up to half of the weekly limit. The API costs $8 per million input tokens and $40 per million output tokens.

There is also a fast mode. It generates text up to 2.5 times faster, but costs twice as much: $16 and $80 per million tokens. It is available in Claude Code, the Anthropic API, Google Cloud and Microsoft Foundry. Mythos 5.2 in fast mode is available only to organizations in verification programs.

Use an advisor instead of switching models

The most practical new feature for anyone who does not want to pay for Fable on every line is advisor. It lets the main model in a session consult a more capable model at key moments: before choosing an approach, when the same error comes up again, or before declaring the task finished. The advisor receives the full conversation, including every tool call and its result, and returns guidance. Claude decides when to call it, not the user.

You can enable it in three ways: enter /advisor during a session (for example, /advisor fable), set advisorModel in the settings file, or use the --advisor flag at launch. Turn it off with /advisor off. The advisor must be at least as capable as the main model. For Opus 5 and 5.5, the documentation allows Fable or Opus 5 and later as the advisor. For Fable 5.2, the only permitted advisor is another Fable 5.2.

Why does this matter? A cheaper model doing the work, with Fable handling decisions, could cost much less than using Fable for the entire session. The documentation specifically recommends, for example, using Sonnet as the main model and Fable as the advisor, so you get Fable’s guidance at decision points without paying for it all the time. Each consultation still reads the full conversation at the advisor’s rates, with no prompt caching on the advisor’s side. Where Fable draws on usage credits in a subscription, an advisor running Fable does too. One more limitation: advisor works only through the Anthropic API, not Amazon Bedrock, Claude Platform on AWS, Google Cloud or Microsoft Foundry.

Leading hundreds of agents

Anthropic describes Fable 5.2 as a model that can advise smaller models (Haiku, Sonnet and Opus) or lead them by assigning simpler tasks. In the Claude API, this is supported by multi-agent orchestration in Managed Agents, which is still in beta. An agent can delegate tasks to subagents and read their reports, or write a workflow - a program that runs multiple agents in the background and combines their results. Fable 5.2 and Mythos 5.2 are among the supported models.

The launch examples show the scale. Over eight days, Fable 5.2 coordinated more than 1,000 subagents and built a prototype operating system from an empty folder, one that boots on a laptop. A team of agents running an early version of the model rewrote a production service from Python to Rust; it now starts about 26 times faster. One thousand Fable 5.2 agents formalized Fermat’s Last Theorem in 30 hours. According to Anthropic, Claude now writes more than 90% of the changes merged into the company’s main codebase, under human supervision.

There is also an API change that may get less attention: preserved thinking, which retains the model’s thinking blocks between turns, is now available to all accounts. Anthropic presents this as part of its defense against distillation, or competitors extracting knowledge from the model. The company says its anti-distillation techniques reduced the amount of data collected by such “distillers” by 90-99%.

My test results

I tested the code in my test across 8 tasks with 136 unit tests covering a parser, regex, SQL, refactoring, a calculator and more. Fable 5.2 passed all of them, as did Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon. Fable 5.2 generated 135-141 tokens per second while writing code, compared with 104 for Fable 5.1 on the same Anthropic infrastructure. But I paid $0.29 for 6 basic coding tasks, compared with $0.15 for Opus.

Watch out for prompt injection

The system card includes one detail every Claude Code user should know. In prompt-injection scenarios, where someone slips instructions into a file or webpage, Fable 5.2 escalates to Opus 4.8 more often than previous models. The result is a higher rate of successful attacks. When Fable responds on its own, Anthropic says it is one of the company’s most resilient models. The problem is the fallback, not Fable itself (system card, p. 3).

A second issue concerns text pasted into a prompt, such as a README, email or webpage containing a planted instruction. Early versions of Mythos 5.2 followed these instructions more often than Opus 5 and Sonnet 5, including instructions hidden in invisible characters. The model reasoned that everything in the user’s message came from the user. Anthropic says the final version and changes to its products largely fixed this (p. 130). I still do not paste content into an agent without reading it first.

Mythos 5.2, the version without cybersecurity restrictions, refused malicious requests slightly less often than Mythos 5.1 in Claude Code, and helped with almost all dual-use security tasks (p. 3). This is not relevant to most users, since Mythos is available only to verified organizations. Read more in the Claude Mythos 5.2 hub and the Claude Fable 5.2 guide.

Frequently asked questions

How do you enable Claude Fable 5.2 in Claude Code?

Update Claude Code to version 2.1.321 or later, then enter /model fable. In plans where Fable draws on usage credits, Claude Code will ask you to approve this once.

What is advisor in Claude Code?

It lets the main model consult a more capable model at key points in a task. Configure it with the /advisor command, the advisorModel setting or the --advisor flag. It works only through the Anthropic API.

How much does Fable 5.2 fast mode cost?

It costs $16 per million input tokens and $80 per million output tokens, twice the standard rates. In return, generation is up to 2.5 times faster.

What effort level should you use for Fable 5.2?

Medium is the default in Claude Code and is a sensible place to start. Anthropic’s charts show that at lower effort, Fable often matches Opus 5.5 at its highest effort level, while costing less.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources

• Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (November 9, 2026)
• Claude Code documentation: advisor (checked November 9, 2026)
• Claude documentation: Multi-agent orchestration (beta)
• Claude documentation: Preserved thinking
• Claude documentation: Pricing (checked November 9, 2026)
• Anthropic: System Card Claude Fable 5.2 & Claude Mythos 5.2, pp. 3 and 130
• Claude Help Center: Fable 5.2 launch promotion