Claude Fable 5.2: Prefix-Check 400 on Every Account, $8/$40, and Default Effort Is High

Attributed compile + list-price arithmetic on my box

Primary announcement: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (Anthropic, dated November 9, 2026).

Primary docs: What's new in Claude Fable 5.2, Claude Fable 5.2 overview, Claude Fable product page, Claude Platform release notes (November 9, 2026).

Independent index (not a vendor bench): Artificial Analysis: Claude Fable 5.2 (max), plus the high, xhigh, and medium pages.

Discussion (not a source of numbers): HN: Claude Fable 5.2 and Mythos 5.2 (996 points / 8 comments at check); HN: Claude Fable 5.2 scores 65 on Artificial Analysis (379 points / 278 comments).

Local copies: /workspace/reputation/runs/40-fable-5-2/.

Specs and prices below are vendor docs unless marked otherwise. AA figures are from the pages I fetched this morning. This box did not send a Messages API request.

Anthropic opened Claude Fable 5.2 on November 9, 2026, with a twin called Claude Mythos 5.2. The callable id is claude-fable-5-2. List price is $8 / $40 per million input / output tokens — 20% below Fable 5.1's $10 / $50 — and official docs say outputs can be up to 30% faster than 5.1. The official HN thread was at 996 points this morning. The migration trap is not the price. Anthropic's "What's new" page leads with four changes that can break existing code, including a 400 if you replay a thinking block after editing the conversation prefix. That prefix check now fires on every account, not just ones created after August 31, 2026. Swapping "claude-fable-5-1" for "claude-fable-5-2" is not the whole job. Anthropic also still tells you not to start here for daily work: for most coding and knowledge tasks, start with Claude Opus 5.5 ($4 / $20, half of Fable 5.2 at the same token counts) and move up only when you need to.

What shipped

Vendor docs, compressed:

Fable and Mythos are the same underlying model. Fable ships with cyber and biology safeguards and is open to all API customers; Mythos is for organizations in Anthropic's Life Sciences or Cyber Verification Programs. When Fable blocks a request, docs say the reply comes from Opus 4.8 (cyber) or Opus 5 (biology). US-only inference is available at 1.1× input and output pricing. I am not repeating Anthropic's prerelease stories as results I observed. The rest of this post is about what breaks when you switch, the list-price math, and what Artificial Analysis reported at each effort level this morning.

Four changes that return 400 — plus thinking blocks that will not come back

If Claude Code, claude.ai, Managed Agents, or the Agent SDK builds your requests, the prefix-check item does not apply. If your own code builds the messages array, it does.

1. The prefix check is now on every account. The API tests whether anything before a Fable 5.2 or Mythos 5.2 thinking block has changed: the system prompt, tools, or an earlier message. A request that sends back an invalid block, with prefix_mismatch_behavior unset, returns a 400 whose message says The block is bound to a different conversation. On Fable 5.1 that error was the default only for accounts created on or after August 31, 2026, so code that edits history can look fine on an older 5.1 account and fail the day you switch. The official stopgap is the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior set to "drop_block" (dropping blocks can raise output tokens). The durable fix is append-only history: do not edit system or tools mid-conversation, only append to messages, and send each assistant turn back exactly as returned.

2. Computer use must use the new toolset. computer_20251124 returns a 400 invalid_request_error on every platform except Amazon Bedrock. Declare computer_toolset_20260801 instead. Bedrock still accepts the earlier tool if you send computer-use-2025-11-24 in anthropic_beta.

3. A Fable 5.2 executor only accepts a Fable 5.2 or Mythos 5.2 advisor. Any other advisor, including Fable 5.1, returns a 400. Change the advisor's model in the same change as the executor, and do it in a new conversation — editing tools mid-thread fails the prefix check. Every executor model accepts Fable 5.2 as its advisor, so you can upgrade the advisor first.

4. Forced tool use is a 400 (tool_choice of any or a named tool). Keep tool_choice: {"type": "auto"} and use strict: true, or JSON outputs. Organizations with customer-managed encryption keys do not get structured outputs on these two models; validate tool input in your code.

A fifth gotcha does not always 400, but it will surprise a router. On the Claude API and Google-managed Claude on Google Cloud, Fable 5.2 and Mythos 5.2 can read thinking blocks from each other and from earlier models, including Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1. No other model reads theirs. Move a conversation from Fable 5.2 down to Opus 5.5 or back to Fable 5.1 and the API drops those blocks: the request still succeeds, dropped blocks are not billed, and the reasoning from those Fable turns is gone. Thinking blocks are also bound to the account that produced them, or to a linked account.

A first Messages call, as a shape only — this box did not execute it. Send the anthropic-version header your SDK already uses (current SDKs default to 2026-10-21):

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2026-10-21" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-fable-5-2",
    "max_tokens": 4096,
    "effort": "high",
    "messages": [
      {"role": "user", "content": "Outline a safe rollback plan for a failed FastAPI migration, in six bullets."}
    ]
  }'

Pin effort yourself. If you omit it, the API default is high, not the Medium default that Anthropic documents for Claude apps and Claude Code.

$8/$40 is 20% off 5.1 — and still twice Opus 5.5

I ran prices.py on the official list rates (arithmetic, not an invoice and not a measured task cost). For a made-up turn of 25,000 input + 5,000 output tokens, with no cache and no tools:

fable-5.2          $0.4000
fable-5.2-fast     $0.8000
fable-5.2-batch    $0.2000
fable-5.1          $0.5000
opus-5.5           $0.2000

5.2 / 5.1 at same tokens:      0.80  (20% cheaper)
5.2 / opus-5.5 at same tokens: 2.00x
fast / standard 5.2:           2.00x
batch / standard 5.2:          0.50x
cache read vs base input 5.2:  0.025x ($0.20 / MTok)

Equal-token arithmetic is the only comparison I can make without a key. Real sessions will not hold token counts still. Anthropic's announcement says that on FrontierCode 2.0, Fable 5.2 at High effort scored about level with Opus 5.5 at Max (45.3% vs 45.5%) for roughly half the cost per task — a vendor result, on their harness, not something I reproduced. Fast mode is 2× the rate for up to 2.5× output tokens per second, useful when you are waiting on a turn. For offline volume, Batch at half price is the cheaper official lever. Cache reads at $0.20/MTok are what make the 1M window usable; stuffing the whole window into every turn at $8/MTok is not.

Default effort is high — and AA's max score is a different bill

The API default is high. Docs say most of the capability gain is already present there, to step up to xhigh only where you have measured a gain, and to reserve max for ceiling tests. Claude apps and Claude Code default to Medium, so a raw API client and the chat product are not on the same rung unless you set it.

Artificial Analysis published Intelligence Index pages for four effort labels. I fetched them this morning (2026-11-12, America/Toronto). These are AA's figures, on their harness, labeled "Default Fallback" on each page, not a run from this box:

On that table, high and xhigh are the same Index score (62), and xhigh costs more and takes longer. Max is the headline 65, and also the expensive setting: $26.54 and 47.5 minutes per Index task, versus $9.08 and 16.7 minutes at high, with 275M output tokens against AA's comparable-model median of 72.8M. "It topped the Index" is not a reason to pin effort: "max" in production — if you skip the effort field on the API, you are already on high. AA's Max page also calls the variant "particularly expensive," "notably slow per task," and "very verbose," and lists time to first answer token as 252.09s (based on Anthropic's API, per AA).

Prompts that used to work can waste a turn

Official prompting notes: a line such as "answer directly without deliberating" does not, in Anthropic's testing, stop thinking at medium and above — if easy requests get slow or expensive, lower effort, do not argue in prose. On long tasks, especially at low and medium, a turn can end with a list of what remains; ask it to finish the whole task. What the prompt says about code ownership matters: in Anthropic's testing, a prompt that said nothing about who owns the code tended to add to the existing structure and report nearby problems instead of fixing them.

What this box actually did

preflight 2026-11-12 ~09:10–09:11 America/Toronto
NO ANTHROPIC API KEY / NO MESSAGES API CALL / NO WEIGHT DOWNLOAD

HTTP 200: Anthropic announcement; What's new in Fable 5.2; model overview;
  product page; Platform release notes
HTTP 200: AA /models/claude-fable-5-2 (max), -high, -xhigh, -medium
HN Firebase: items 50429906 (996/8) and 50431144 (379/278)

Parsed: $8/$40 (was $10/$50); prefix check on every account; computer
  toolset 400 except Bedrock; advisor 400 unless Fable/Mythos 5.2;
  thinking blocks unreadable by any other model; API default effort high
Local run: prices.py list-price arithmetic only

NOT tested: authenticated completion; cache hits; fast mode latency;
  computer use or advisor tool; mid-conversation model switches;
  Claude Code / claude.ai behavior

For a Toronto product team on 2026-11-12, the rational first move is keep Opus 5.5 as the daily driver, audit your request builder for prefix edits and the old computer-use tool, pin claude-fable-5-2 with effort set explicitly (start at high) on the tasks that actually need it, and treat max as a ceiling test — not a new default.

YongBo Yu — Toronto AI engineer (agent systems / LLM workflows). Site: yongbo-yu.vercel.app · GitHub: YongBoYu1