Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, cutting list API prices by twenty percent and tightening several agent-facing API rules. The November 9 Claude Platform release notes announce claude-fable-5-2 and claude-mythos-5-2 at $8 per million input tokens and $40 per million output tokens, down from $10/$50 on Claude Fable 5.1 and Mythos 5.1. The models overview now lists Fable 5.2 as the line for demanding reasoning and long-horizon agentic work.
This guide explains what changed for builders, what migration steps Anthropic documents, and how the release fits next to Opus 5.5 and the new Claude electronics design results that Anthropic tested on Fable 5.2 the same day.
TL;DR: the questions teams will ask #
| Question | Answer |
|---|---|
| Model IDs? | `claude-fable-5-2` ,`claude-mythos-5-2` |
| Standard price? | $8 in / $40 out per million tokens |
| vs Fable 5.1? | Same list cut: was $10/$50 |
| Fast mode? | speed: "fast" at $16/$80, up to 2.5x output tok/s |
| Context / output? | 1M context, 128K max output, adaptive thinking on |
| Who gets Mythos? | Verified orgs (Cyber Verification Program, etc.) |
| Computer use? | Must use computer_toolset_20260801 |
| Data retention? | 30-day default; same constraints as 5.1 |
| Where available? | Claude API, Claude Platform on AWS, Anthropic-managed Google Cloud, Google-managed Google Cloud |
Pricing: what actually got cheaper #
The headline is straightforward. Input and output list prices fall twenty percent versus Fable 5.1 and Mythos 5.1. Anthropic does not announce new cache-read prices in the November 9 release note paragraph we cite; if your bill is cache-heavy, re-check the pricing page after you migrate, because cache tiers can dominate agent workloads more than base input list price.
Fast mode is new for Fable 5.2 on the Messages API. Set speed: "fast" with no beta header. Anthropic quotes $16/$80 per million tokens and up to 2.5x higher output tokens per second. That is a latency trade, not a capability tier: same model, higher throughput tariff.
For comparison with the rest of the Claude stack:
| Model | Input $/MTok | Output $/MTok | Role in Anthropic's table |
|---|---|---|---|
| Fable 5.2 | 8 | 40 | Demanding reasoning, long agentic runs |
| Opus 5.5 | 4 | 20 | Long agentic coding and knowledge work |
| Sonnet 5.5 | 2 | 10 | Speed plus intelligence balance |
| Haiku 5.5 | 0.10–1.25 | 0.50–4.25 | High-volume, latency-sensitive work |
Fable 5.2 remains the expensive reasoning line even after the cut. Teams routing "default coding" to Opus 5.5 for cost and only escalating hard tasks to Fable should revisit that ladder with the new numbers. Our OpenAI and Anthropic price-cut timeline shows how fast these list prices have moved in 2026 alone.
Fable 5.2 vs Mythos 5.2: same model, different door #
Anthropic's pattern from Fable 5.1 and Mythos 5.1 continues: same underlying capabilities and list price, but Mythos 5.2 is gated to organizations Anthropic has verified for cyber or life-sciences work. If you are on Mythos 5.1 through the Cyber Verification Program, plan a model ID swap rather than a re-architecture.
For everyone else, Fable 5.2 is the migration target. Anthropic still describes it as best for the most ambitious, open-ended technical work in the same release note that Techmeme quoted on November 9.
API and agent migration: what breaks if you ignore it #
The release note is unusually explicit about footguns. Treat these as hard failures, not warnings.
Prefix checks on every account
Fable 5.2 and Mythos 5.2 enforce the prefix check for every account, whatever its age. Fable 5.1 only returned the error by default on accounts created on or after August 31, 2026. If your harness mutates the system prompt, tool definitions, or an earlier turn and then sends back a Fable 5.2 thinking block from a later turn, you get a 400 error. Anthropic tells you to check whether your code edits the prefix before switching production traffic.
This matters for Claude Code-style sessions that compact context, rewrite tool lists, or replay traces from logs. The METR acceleration post assumed faster Anthropic release cadence; tighter prefix enforcement is the cost side of that speed.
Computer use toolset version
Requests that declare the older computer_20251124 tool return 400. You must migrate to computer_toolset_20260801. Anthropic links a migrate from computer_20251124 guide. Any desktop or browser agent still on the old tool name needs a deploy before you bump the model string.
Advisor tool pairing (beta)
When using the advisor tool in beta, a Fable 5.2 or Mythos 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor. Any other advisor model returns 400. Every executor model still accepts Fable 5.2 as advisor, per the release note.
Mid-conversation model switches
If you upgrade a thread to Fable 5.2 mid-flight on the Claude API or Google-managed Claude, reasoning from Opus 5.5, Sonnet 5.5, and Haiku 5.5 is preserved. If you downgrade away from Fable 5.2 or Mythos 5.2, thinking blocks those models produced are dropped, but the request succeeds. Logged replay pipelines that expect thinking blocks to survive arbitrary model hops will see silent loss on downgrade.
Preserved thinking and data retention
Thinking blocks from Fable 5.2 and Mythos 5.2 work only in the account that produced them or a linked account. Both models keep 30-day data retention requirements like 5.1 and are not available under zero data retention unless Anthropic authorizes it.
Platform availability #
Anthropic lists Fable 5.2 on:
- Claude API
- Claude Platform on AWS
- Claude Models on Google Cloud (Anthropic managed)
- Claude Models on Google Cloud (Google managed)
Mythos 5.2 follows the same verification gate as earlier Mythos releases. Bedrock and Microsoft Foundry availability for Fable 5.2 is not in the November 9 paragraph we read; confirm your cloud's model table before cutover.
What Anthropic did not publish on day one #
The November 9 release note emphasizes price, fast mode, and API contracts. It does not publish a fresh benchmark table for 5.2 versus 5.1 in the section we fetched. Same-day electronics research does cite Fable 5.2 scores on Diode PCB Gym and BenchCAD 2.1, which gives one concrete capability signal beyond token pricing.
If you need a scored comparison against GPT-6 Astra or Meta Muse Wonder, wait for third-party harness runs or Anthropic's own tables. Do not assume a twenty percent price cut implies a uniform twenty percent quality gain.
What builders should do this week #
- Pin model IDs in staging:
claude-fable-5-2and run your agent regression suite with prefix checking enabled. - Search your repo for
computer_20251124and update tocomputer_toolset_20260801before raising Fable model versions. - Reprice internal chargeback using $8/$40 list; add a fast-mode budget line if you enable
speed: "fast". - Mythos customers : confirm verification status still covers
claude-mythos-5-2. - Leave Fable 5.1 only if you depend on looser prefix behavior on older accounts; Anthropic has not announced a Fable 5.1 retirement date in the November 9 note.
For agent security when you widen tool and computer-use surfaces, AgentBeam stops AI agents before they take dangerous actions; pair policy there with Anthropic's prefix and tool-version rules.
How Fable 5.2 fits your model routing table #
Most production stacks already run a three-tier router: Haiku or another small model for classification, Sonnet or Opus for default coding, and Fable only when benchmarks or user tier demand it. The November 9 price change shifts the breakeven point where Fable is worth the premium over Opus 5.5.
Rough list-price math for a balanced 3:1 input-to-output agent turn (illustrative only):
| Model | Approx $ per 1M-token mix (75% in / 25% out) |
|---|---|
| Opus 5.5 | $7.00 | | Fable 5.2 | $16.00 |
| Fable 5.1 (old) | $20.00 | Fable still costs more than Opus per token, but the 20% cut matters when a single long-horizon job burns hundreds of millions of tokens across cache hits. Teams that moved workloads to Fable 5.1 for Terminal-Bench science gains should re-run cost traces this week; do not assume cache-read pricing moved in lockstep without checking Anthropic's pricing doc.
Fast mode is the second dial. Use it when latency dominates, such as interactive demos or blocking CI steps, and you accept double list input and output on Fable. It is not a separate model ID; it is a request flag. Log speed: "fast" usage separately so finance can compare spend against wall-clock wins.
Incident patterns teams reported on earlier Fable releases #
Anthropic's API notes read like a post-mortem checklist from prior launches:
- Prefix mismatch 400s after prompt compaction in agent frameworks
- Computer tool version mismatches when one microservice lagged behind another
- Thinking block loss when A/B testing Fable against Sonnet mid-session
If you operate Claude Code remote sessions on AgentCore, schedule a staged rollout: canary tenant, compare error rate on prefix_mismatch and invalid_tool classes, then widen traffic. Mythos 5.2 customers should run the same tests even if traffic is low, because verification programs often touch the highest-risk tools.
What is still unknown about capability deltas #
Anthropic's November 9 release note does not attach a fresh Terminal-Bench or GDPval table for 5.2 versus 5.1. The same-day electronics research is the first public capability anchor: 69% PCB Gym and 83.5% BenchCAD 2.1. That suggests geometric and layout reasoning moved, not just price.
Until independent harnesses publish, treat 5.2 as "same line, cheaper list, stricter API" unless your eval suite shows a clear win. That is conservative and avoids surprise regressions on writing style or refusals, issues that showed up in the Fable 5.1 mannered-prose debate.
Related reading #
Model IDs, prices, and API rules are accurate as of November 9, 2026, per Anthropic's Claude Platform release notes and models overview.