I think a new model is a migration. You diff what you send, fix what breaks, re-run your evals, and change the model string last.

Claude Fable 5.2 launched Nov 9 and costs less than Fable 5.1, so the one-line swap is tempting. Anthropic's docs list four changes that can break your code, and one can turn an agent loop that ran clean last week into a 400 error.

The Signal

Anthropic launched Claude Fable 5.2 (claude-fable-5-2) on Nov 9, "built for demanding reasoning and long-horizon agentic work." It requires 30-day data retention. On price, it's $8 per million input tokens and $40 per million output, down from $10 and $50 on Fable 5.1. (Claude release notes)

Anthropic doesn't push most teams onto it. "For most workloads, start with the latest Claude Opus model," and move to Fable 5.2 "when your evals on a Claude Opus model at higher effort still don't meet your requirements."

The prefix check is the change that bites hand-rolled agent loops. It tests whether anything before a Fable 5.2 thinking block changed since the model produced it: the system prompt, the tools, or an earlier message. On 5.2 an invalid block gets a 400 on every account. On 5.1 that error only hit accounts created on or after August 31, 2026, "so code on Claude Fable 5.1 that edits its history can run without errors on an older account before the switch and fail after it." Retrying the same request body doesn't clear it.

The other three are quieter. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks, so a router or fallback that moves a conversation to another model, Opus 5.5 and Fable 5.1 included, drops that reasoning without an error. The old computer_20251124 tool returns a 400 everywhere except Amazon Bedrock; declare computer_toolset_20260801 instead. And with the advisor tool (beta), a Fable 5.2 executor accepts only a Fable 5.2 or Mythos 5.2 advisor.

"Nothing changes for you if Claude Code, claude.ai, Claude Managed Agents, or the Claude Agent SDK builds your requests." If your own code builds them, keep system and tools fixed for the whole conversation, only append to messages, and send each assistant turn back exactly as the API returned it. People who run your tool with their own keys can hit the 400 even when your own runs are clean. (What's new in Fable 5.2, preserved thinking docs)

The Stack

  1. Decide if you need it. If the latest Opus at higher effort passes your evals, stay on Opus.
  2. If Claude Code, the Agent SDK or Managed Agents builds your requests, the prefix check isn't your problem. Skip to step 5.
  3. If your code builds messages itself, capture a few real request bodies, including a compaction and a tool change, and diff system, tools and the shared messages between consecutive requests. Anything that changed besides the new turn is an edit.
  4. Fix each edit with the documented pattern: a mid-conversation system message, on-demand compaction, tool_addition and tool_removal blocks, replaying what you actually sent, pinned MCP tool lists. Then run test traffic with the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior set to "error" so the 400 shows up on your own account. Treat "drop_block" as a stopgap only. (migration guide)
  5. Search for computer_20251124, forced tool_choice, budget_tokens and advisor tool models; each can return a 400 on 5.2. Check any router or fallback that could move a conversation off 5.2.
  6. Re-run your evals with an effort sweep that starts at high, the API default. Delete prompt lines that tell the model not to think and lower the effort level instead.

One decision: no production switch this week until the request diff is clean and one eval sweep is done.

Prompt of the Day

Paste this with a short description of your agent:

Here is how my agent builds each Claude request: Agent SDK, Claude Code, or my own code; whether the system prompt or tools change between turns; whether I trim tool results or summarize old turns; whether I use computer use, forced tool choice, budget_tokens or the advisor tool; whether a router or fallback switches models.

Using only those facts, give me: (1) a risk level for switching to Claude Fable 5.2 (low, medium, or high), tied to something I wrote; (2) each likely break point mapped to the fix Anthropic documents for it; (3) a test plan for this week using request diffs, the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior; (4) an eval plan with an effort sweep that starts at high.

Do not invent error rates, benchmark numbers, or features I didn't mention. Write "unknown" for anything missing.

If the SDK builds your requests and you use none of those features, just re-run your evals.

Bring your test plan to AI Freedom Lab, where builders launch companies and web and mobile apps with AI.

Tool of the Day

Zapier AI, featured in the AI Tools Index.

Describe an automation in plain language and Zapier AI turns it into a multi-step workflow across 7,000+ app integrations. Webhooks and multi-step Zaps need Professional, $19.99 a month for 750 tasks. The index's caveat: complex workflows can require paid tiers. (AI Tools Index)

A build to try during migration week: have your agent POST a webhook whenever a Claude request returns a 400 or a response lists dropped blocks in input_transformations. A Zap logs the model, error text and conversation id to a sheet and pings your team channel.

Quick Hits

Starting December 8, 2026, GitHub turns off Dependabot alerts on most repositories with no activity in the past year (announced Nov 10). Existing alerts are archived, and any push turns them back on. Worth checking on old client repos still running in production. (GitHub changelog)

Cursor's Nov 6 Voice Chat with Agents lets you talk with your agents in a realtime voice call. With Privacy Mode on, it runs with zero data retention. (Cursor changelog)

@GitHub in Slack and Teams went generally available on Nov 6. Investigate bugs, create issues and open pull requests without leaving the conversation. (GitHub changelog)

Closing the loop

Pick the agent you'd move first, diff a few real request bodies this week, and fix every edit before you touch the model string. If Opus already passes your evals, stay there and keep building.


Join AI Freedom Lab, where builders launch companies and web and mobile apps with AI: skool.com/aifreedomlab

Recommended for you

View all
caret-right