Claude Fable 5.2 API migration from Fable 5.1 or Opus 5.5: 4 breaking changes, $8/$40 pricing with cache, batch and fast mode, and a worked monthly bill.
Moving to Claude Fable 5.2 is a one-line model ID change from Fable 5.1 plus a four-item review, not a rewrite. Anthropic's docs list four breaking changes from 5.1: failed prefix checks return a 400, 5.1 cannot read 5.2's thinking blocks, the old computer use tool errors, and the advisor tool needs a matching advisor.[2] Sampling parameters, forced tool choice, prefill and refusals-as-HTTP-200 behave as they did on 5.1. From Opus 5.5 the list is longer and the bill doubles: Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, against $4 and $20 for Opus 5.5.[4] Fable 5.1 to 5.2 saves 20 percent; Opus 5.5 to Fable 5.2 doubles the rate. Every figure traces to Anthropic's announcement or Claude API docs, linked in Sources. Below: before and after code, a parameter-by-parameter breaks table, exact pricing with a worked monthly cost, the Fable and Mythos access split, and when to stay put.
One scope note. Everything here comes from Anthropic's announcement and Claude API docs, read on November 16, 2026. The code snippets follow the field names and request shapes in those docs; they were not run against the live API for this post, so test them in your own workspace first. Benchmark figures are Anthropic's own, from its announcement.
The minimum change: before and after
The model ID is claude-fable-5-2 on the Claude API, and claude-mythos-5-2 for the restricted sibling. Context is 1M tokens and max output is 128K on both.[3] Here is the Python call, written to work coming from either Fable 5.1 or Opus 5.5:
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-fable-5-2", # was "claude-fable-5-1" or "claude-opus-5-5"
max_tokens=16000, # thinking tokens count toward max_tokens
output_config={"effort": "high"}, # set it; Opus 5.5 defaulted to medium
messages=[{"role": "user", "content": "Review this diff for race conditions."}],
)
for block in response.content:
if block.type == "text":
print(block.text)Three choices in that call matter. The effort value sits under output_config.[6] It is set explicitly because the API default on Fable 5.2 is high while Opus 5.5 defaults to medium, so an unchanged request runs one level up with no error, only a different bill. max_tokens is large because thinking tokens count against it. And there is no temperature, top_p or top_k: non-default values return a 400, and so does top_p of 1 or sending both temperature and top_p.[8] Same call in TypeScript, with the refusal branch included:
import Anthropic from '@anthropic-ai/sdk';
const client = new Anthropic();
const response = await client.messages.create({
model: 'claude-fable-5-2', // was 'claude-fable-5-1' or 'claude-opus-5-5'
max_tokens: 16000,
output_config: { effort: 'high' },
messages: [{ role: 'user', content: 'Review this diff for race conditions.' }],
});
if (response.stop_reason === 'refusal') {
// HTTP 200, not a thrown error: branch on stop_reason
console.warn(response.stop_details);
} else {
for (const block of response.content) {
if (block.type === 'text') console.log(block.text);
}
}What breaks, parameter by parameter
Each row is a documented change. The "from" column says which starting point it hits, because several items are new only for Opus 5.5 users.[2][3]
| Parameter or behavior | Old behavior | New on Fable 5.2 | Fix |
|---|---|---|---|
| Edited history plus thinking blocks (both) | No prefix-check error before 5.2; the guide lists the 400 as a change from both 5.1 and Opus 5.5 | Failed prefix check returns 400: "The block is bound to a different conversation" | Keep system and tools fixed, only append to messages, echo assistant turns exactly as returned |
| Thinking blocks across models (both) | On the Claude API, 5.2 reads blocks from Fable 5.1 and Opus 5.5, so moving onto 5.2 keeps earlier reasoning | 5.1 and Opus 5.5 cannot read 5.2 blocks; the API drops them, the request succeeds, dropped blocks are not billed | Expect higher cost and latency on the first turn after a router or fallback switch |
computer_20251124 (from 5.1) | Fable 5.1 accepts it with the computer-use-2025-11-24 beta header | 400 on every platform except Amazon Bedrock | Declare computer_toolset_20260801, drop the beta header, update the agent loop |
| Advisor tool (both) | Opus 5.5 executor accepts several advisors | 5.2 executor accepts only Fable 5.2 or Mythos 5.2; anything else is a 400 | Set the advisor model to claude-fable-5-2 in new conversations |
| Data retention (from Opus 5.5) | Opus 5.5 can run under zero data retention | 30-day retention required; ZDR workspaces get a 400 invalid_request_error | Enable 30-day retention for the workspace, or contact your Anthropic account team before you switch |
| Default effort (from Opus 5.5) | medium | high | Pass output_config.effort explicitly and sweep |
Small max_tokens (from 5.1) | A "don't deliberate" prompt line kept replies short | That line no longer stops thinking at medium and above; a request can end in max_tokens after thinking with no text | Raise max_tokens or lower effort; delete the line |
| Rate limits (both) | 5.1 shares a limit with Fable 5; Opus 5.5 has its own | Fable 5.2 has a separate limit, with lower token limits per tier than Opus 5.5 | Check your tier before moving traffic |
| Structured outputs under CMEK (from Opus 5.5) | Supported | Strict tool use and JSON outputs unavailable | Remove strict: true and output_config.format, validate in code |
The prefix check is the one that bites
If Claude Code, claude.ai, Managed Agents or the Agent SDK builds your requests, nothing changes for you.[3] If your own code assembles messages, any code that rewrites a system prompt per turn, trims old messages, reorders tools, or reconnects an MCP server with a different tool list can trigger it. Retrying the same body does not clear the error. For a staged rollout, the docs describe a stopgap: send the beta header thinking-binding-controls-2026-08-01 and set prefix_mismatch_behavior to drop_block. The response then lists each dropped block in a top-level input_transformations array.[2]
{
"thinking": {
"type": "adaptive",
"block_binding": {
"prefix_mismatch_behavior": "drop_block"
}
}
}That fragment is the thinking setting as the docs show it, not a full request body. Send it with the header anthropic-beta: thinking-binding-controls-2026-08-01. In the SDK, the docs call the beta method, client.beta.messages.create, and pass betas: ["thinking-binding-controls-2026-08-01"] alongside the same thinking object. Treat it as a diagnostic, not a setting: Anthropic warns that dropping blocks can raise output tokens. Log every input_transformations entry, fix the edit that caused it, then return to the default "error".[2]
What does not break
These carry over from Fable 5.1 unchanged: adaptive thinking is always on, and both thinking: {"type": "disabled"} and budget_tokens return a 400; prefilling the assistant message returns a 400; forced tool choice (any or tool) returns a 400, leaving auto and none.[2] Opus 5.5 shares the always-adaptive and forced-tool rules, so those two do not hurt an Opus migration either.[2] For schema-valid output, keep auto and add strict: true tools or JSON outputs.
Refusals arrive as HTTP 200
A declined request is a normal response with stop_reason: "refusal", an empty or partial content, and a stop_details.category of cyber, bio, frontier_llm, reasoning_extraction or general_harms. The category and explanation can be null, so branch on stop_reason, not on those fields.[5] Your error-rate dashboards will never see these. Server-side fallback (beta) retries on another model inside one call; for Fable 5.2 the permitted targets are Claude Opus 4.8 and Claude Opus 5.[2]
response = client.beta.messages.create(
model="claude-fable-5-2",
max_tokens=16000,
messages=messages,
fallbacks="default",
betas=["server-side-fallback-2026-07-01"],
)
if response.stop_reason == "refusal":
# no recommended fallback for this category, or the fallback declined too
log_refusal(response.stop_details)
elif any(i.type == "fallback_message" for i in response.usage.iterations or []):
log_fallback(response.model)log_refusal and log_fallback are placeholders for your own logging. Two billing details to know. A pre-output refusal is billed only in the bio, frontier_llm and reasoning_extraction categories, and every refusal still counts against rate limits. And reasoning_extraction has no fallback at all: it fires on prompts that ask for reasoning in the output, such as a "scratchpad" section or a reasoning field in a tool schema. Change the prompt instead.[5]
Exact pricing, including cache, batch and fast mode
All figures are USD per million tokens, from the pricing page.[4] Cache multipliers stack with batch and fast mode; batch and fast mode do not combine with each other.
| Price per MTok | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| Input | $8 | $10 | $4 |
| Output | $40 | $50 | $20 |
| Cache write, 5 minute | $10 | $12.50 | $5 |
| Cache write, 1 hour | $16 | $20 | $8 |
| Cache read | $0.20 | $0.25 | $0.20 |
| Batch input / output | $4 / $20 | $5 / $25 | $2 / $10 |
| Fast mode input / output | $16 / $80 | not listed | $8 / $40 |
Fable 5.2's cache read is 0.025 times base input, the same multiplier as 5.1, which is why the read price falls from $0.25 to $0.20 while base input falls from $10 to $8. Opus 5.5 reads at 0.05 times a lower base and lands on the same $0.20.[4] Fast mode is exactly 2x the standard price for up to 2.5x output speed, and it is set with speed: "fast" on the Claude API.[7]
Worked monthly bill
Assume an agent service running 30,000 requests a month. Each request has 50,000 cached-read input tokens, 8,000 fresh input tokens, 2,000 five-minute cache-write tokens and 6,000 output tokens. Monthly totals: 1,500M cache reads, 240M fresh input, 60M cache writes, 180M output.
| Scenario | Cache reads | Fresh input | Cache writes | Output | Total |
|---|---|---|---|---|---|
| Opus 5.5 | $300 | $960 | $300 | $3,600 | $5,160 |
| Fable 5.1 | $375 | $2,400 | $750 | $9,000 | $12,525 |
| Fable 5.2 | $300 | $1,920 | $600 | $7,200 | $10,020 |
| Fable 5.2 batch | $150 | $960 | $300 | $3,600 | $5,010 |
| Fable 5.2 fast | $600 | $3,840 | $1,200 | $14,400 | $20,040 |
Fable 5.1 to 5.2 saves $2,505 a month, exactly 20%, because every price line drops 20%. Opus 5.5 to Fable 5.2 adds $4,860. Two caveats keep this honest. It holds token counts constant, and they will not stay constant: you are moving from medium to high effort, which changes thinking volume. Anthropic reports that in early testing on FrontierCode 2.0, Fable 5.2 at High scored 45.3% against 45.5% for Opus 5.5 at Max, at roughly half the cost per task.[1] That is a vendor-reported figure on one benchmark, not your workload. Every row also assumes the stated cache reads all hit. Plug your own volumes into the AI model cost calculator, and count real tokens with the token counter. Fable 5.2 uses the tokenizer introduced with Claude Opus 4.7, which the pricing page says produces approximately 30% more tokens than Claude Sonnet 4.6 and earlier for the same text.[3][4] For how the cache discount played out on the previous generation, see our Fable 5.1 cache pricing breakdown.
Fable 5.2 versus Mythos 5.2: the access split
Anthropic says Fable and Mythos are the same underlying model. The difference is safeguards and who may call it.[1]
| Claude Fable 5.2 | Claude Mythos 5.2 | |
|---|---|---|
| API model ID | claude-fable-5-2 | claude-mythos-5-2 |
| Who can call it | All Claude API customers, no approval | Organizations verified through Anthropic's Cyber or Life Sciences Verification Programs |
| Specs and price | 1M context, 128K output, $8 / $40 | Identical |
| Safeguards | Classifiers block high-risk cyber work and most research biology | Without certain cyber and biology safeguards |
| Fast mode | Yes, $16 / $80 | Yes, same price, verified organizations only |
Fable 5.2 produces about 55% fewer classifier blocks than Fable 5.1 in Anthropic's early testing across claude.ai, Claude Code and Cowork (about 60% fewer in cyber, 40% in biology). That is a rate, not zero.[1] A separate limit is deliberate: during training Anthropic held back advances in AI-development research skills, so accelerator kernel optimization and LLM architecture work sit at about Opus 5.5 level on both models. Anthropic says agentic coding, hardware work and conventional ML research were not reduced.[1]
When not to upgrade
- Opus 5.5 already passes your evals. Anthropic's own guidance is to start with Opus 5.5 for most coding and knowledge work and move up only when higher effort still falls short.[1][3]
- You run under zero data retention. Requests return a 400 until 30-day retention is enabled or Anthropic expressly authorizes an exception.[2]
- You depend on Priority Tier. Neither Fable 5.2 nor Mythos 5.2 is supported there.[2]
- Your product touches security, biology or ML research. Benign work can trip the cyber, bio or
frontier_llmclassifiers, and Anthropic says accelerator kernel optimization and LLM architecture work sit at about Opus 5.5 level on Fable 5.2, at twice the Opus 5.5 price. - You serve short, latency-sensitive turns. Thinking cannot be disabled, and at
highand above the model thinks before almost every reply.[3] - Rate-limit headroom is tight. Fable 5.2's per-tier token limits are lower than Opus 5.5's.[2]
Pre-flight checklist
- Confirm retention: 30-day on the workspace, or a confirmed arrangement with your Anthropic account team.
- Swap the model ID and set
output_config.effortexplicitly. - Run a staging session with
drop_block, then fix every logged prefix mismatch. - Replace
computer_20251124and update any advisormodelfield. - Add a
stop_reason: "refusal"branch and a fallback path on every request route, subagents included. - Re-run your effort sweep from
high, then re-baseline cost and latency against Opus 5.5 using the Opus 5.5 effort settings as your reference point.
Sources
- Introducing Claude Fable 5.2 and Claude Mythos 5.2, Anthropic, November 9, 2026 (accessed November 16, 2026)
- Claude Fable 5.2 and Claude Mythos 5.2 migration guide, Claude Platform Docs (accessed November 16, 2026)
- What's new in Claude Fable 5.2, Claude Platform Docs (accessed November 16, 2026)
- Pricing, Claude Platform Docs (accessed November 16, 2026)
- Refusals and fallback, Claude Platform Docs (accessed November 16, 2026)
- Effort, Claude Platform Docs (accessed November 16, 2026)
- Fast mode, Claude Platform Docs (accessed November 16, 2026)
- Thinking, sampling parameters section, Claude Platform Docs (accessed November 16, 2026)
Written by
WOWHOW
The WOWHOW team brings 14+ years of production engineering experience. Every tool and product in the catalog is built, tested, and curated in-house.
Monday memo · free
One insight, every Monday. 7am IST.
1 field report, 3 links, 1 tool we actually use. No fluff, no spam.
Beyond the writing
Need production-ready templates?
Free browser tools with no signup, plus premium dev templates and starter kits.
Comments · 0
Beta: comments are stored locally on your device and not visible to other readers.
No comments yet. Be the first to share your thoughts.