Models
Claude Fable 5.2 is built for multi-day agent work
Anthropic released Claude Fable 5.2 on November 9 with a 1M-token context and 128K output. What that buys hardware teams running long intake loops.

On November 9, Anthropic shipped Claude Fable 5.2 and Mythos 5.2 — frontier models aimed at technical work that runs for hours or days, not single chat turns. API id: claude-fable-5-2. Context window: 1M tokens. Max output: 128K. List pricing landed at $8 / MTok input and $40 / MTok output, below Fable 5.1.
Anthropic positions Fable for coding, finance, and science tasks where the model must question its own work, spawn subagents, and keep state across tool calls. Fast mode doubles throughput at 2× price for latency-sensitive paths.
Why 1M tokens matter to hardware
A serious board repo is not one file. It is schematics, multiple PCB revisions, a BOM with alternates, markdown specs, test logs, and the last email from your CM about a wrong footprint.
Older context limits forced you to chop the project before the agent could see how the power tree, RF keepout, and test jig relate. A 1M window is room to keep the program coherent while an agent:
- Runs KiCad MCP tools on a live layout.
- Cross-checks the BOM against distributor stock.
- Calls manufacturing preflight and explains the blocker in plain language.
That is the shape of work we see on Gate — long threads, not one-shot quotes.
Migration details that bite
Fable 5.2 enforces prefix checks on every account. If your integration edits the system prompt or tools after emitting a thinking block, the API returns 400. Computer use moved to computer_toolset_20260801. Data retention defaults to 30 days unless you have explicit zero-retention approval.
Read the platform release notes before you swap model strings in production agents.
Models do not release radios
Fable can reason about your schematic for a week. It still cannot start SMT without a human price, a confirmed file set, and — when the board transmits — a named release on the build. That policy does not change because the model got smarter.
Pair Fable with structured tools: KiCad MCP in the editor, Gate MCP at intake, and one build thread from first articles to volume. The November 9 release is the reasoning engine. The line is still the line.