Claude Fable 5.2 Complete Guide: Pricing, Fast Mode, Benchmarks and Migration

The short answer

Claude Fable 5.2 is Anthropic’s most capable generally available model for hard technical work, and it costs less than Fable 5.1. As of November 11, 2026, it lists at $8 per million input tokens and $40 per million output tokens, down from $10 and $50. It runs up to 30% faster than Fable 5.1 by Anthropic’s count, offers an optional fast mode that Anthropic says is up to 2.5 times quicker at $16 and $80, and has a 1M-token context window with up to 128K tokens of output. The Claude API model ID is claude-fable-5-2.

Anthropic does not tell most people to start here. Its announcement recommends beginning with Claude Opus 5.5, its strongest daily driver, and moving up to Fable only for the most ambitious, open-ended work. Fable 5.2 can also act as an Advisor that smaller models consult, or as an Orchestrator that hands simpler tasks to cheaper ones. Claude Mythos 5.2 is the same underlying model with fewer safeguards, open only to verified cyber and life-sciences organizations.

For developers the price cut is the easy part. The harder part is four documented breaking changes, listed in the migration section, that can turn a working Fable 5.1 integration into a stream of 400 errors. This guide is built from Anthropic’s announcement, the model overview, the what’s new page, the prompting guide, the release notes, the launch-offers help article and the AWS launch post. Benchmarks are labeled as vendor-reported or independent. We did no hands-on testing; the cost arithmetic is ours, from published prices. For the neighboring models, see our Claude Fable 5.1 guide and the sibling posts on Gemini 4 Carbon and the Gemini 4 Carbon vs Claude Fable 5.2 head-to-head.

A creator roundup comparing Gemini 4 Carbon and Claude Fable 5.2 (watch on YouTube). Treat it as commentary; every number in this guide is checked against a vendor or leaderboard page.

What changed from Fable 5.1

Anthropic frames 5.2 as an upgrade in coding, reasoning and vision, with clearer communication. The changes that matter in practice fall into four groups: money, speed, capability and API behavior.

Two of those rows deserve a note. The 30% speed figure is a ceiling, “up to,” and Anthropic gives no baseline hardware or prompt shape for it. The 55% reduction in blocks comes from early testing across claude.ai, Claude Code and Cowork, and is Anthropic’s own measurement. Neither has an independent check that we could find.

The announcement also describes the model’s training. Anthropic says it took steps during training to limit advances in the research capabilities that could most directly accelerate AI development, such as accelerator kernel optimization and improving large language model architectures, where Fable 5.2 performs at about the level of Opus 5.5. It says agentic coding, hardware development and traditional machine learning research were not affected, and that nothing is altered at inference time. That is a notable carve-out for anyone evaluating the model for AI-engineering work.

The launch posts

Anthropic announced the models on X on November 9, 2026. The @claudeai thread opened with the headline claim and a launch video, then listed the safety testing, the benchmark step above Opus 5.5, and the price.

The @AnthropicAI account followed minutes later with the availability post, linking to the announcement.

The @claudeai thread states the price directly: available on all paid plans, the Claude Platform and the major cloud platforms, at $8 per million input tokens and $40 per million output tokens, down from $10 and $50. A developer-focused account, @ClaudeDevs, posted the same price and the up-to-30% speed claim for the Claude Platform and Claude Code. Both launch posts are also linked in the sources list below.

Pricing and cost math

Fable 5.2 and Mythos 5.2 share one price list. These are USD list prices per million tokens as of November 11, 2026, from the model overview and the fast mode page.

The cut is 20% on both input and output. The cost examples below are our arithmetic from those list prices, not measured bills. They assume a single agent turn with the stated token counts and ignore tool charges, data residency multipliers and refusals.

Worked example: 200,000 input tokens at $8 per million is $1.60, and 20,000 output tokens at $40 per million is $0.80, so $2.40 per task, against $2.00 plus $1.00, or $3.00, on Fable 5.1. At 1,000 such tasks a month that is $2,400 against $3,000. With a stable 180K-token prefix served from cache, the input side drops to 180,000 times $0.20 per million, or $0.036, plus 20,000 fresh tokens at $8 per million, or $0.16, and the task costs about $1.00 including the $0.80 of output. Fast mode doubles every line, so it pays only where a person is waiting on the answer.

Per-token price is not per-task price. Output length and thinking depth vary with the effort setting, and Anthropic says Fable 5.2 at Medium effort can complete challenging tasks with fewer tokens than Fable 5.1. The independent cost-per-task figures in the Artificial Analysis section below are a better guide to real spend than the per-token list.

Vendor-reported benchmarks

The table below reproduces a selection of rows from Anthropic’s launch table. Every number is vendor-reported, published November 9, 2026, with safeguards on and each model at its highest available effort unless noted. Anthropic’s table also lists GPT-6 Astra and Gemini 4 Carbon; those cells are Anthropic’s numbers for rival models, not figures from OpenAI or Google. The full table, including cyber, protein-design and robotics rows, is on the announcement page.

Read the table with three cautions. First, Gemini 4 Carbon beats Fable 5.2 on AutomationBench (53.9% against 51.4%) even in Anthropic’s own chart, so “a step above Opus 5.5 in almost all the capabilities we evaluated” does not extend to every row against every rival. Second, the knowledge-work block for AA-Briefcase lists Fable 5.2 ahead on the rubric score (66.1% against 61.2% for Opus 5.5), but a second Elo-scored line below it shows Opus 5.5 ahead, 1800 to 1785. Third, the comparison cells for other vendors were run by Anthropic and may not reflect how those vendors would configure their own models.

Anthropic also publishes cost-efficiency claims, which matter more than raw scores for budgeting. On FrontierCode 2.0, it says Fable 5.2 at High effort scores about level with Opus 5.5 at Max (45.3% against 45.5%) for roughly half the cost per task. It says that on Perplexity’s WANDR benchmark, Fable 5.2 at Medium matches Opus 5.5’s top score for about 40% of the cost per task, and that on Humanity’s Last Exam it matches Opus 5.5’s top score at Low effort for less than one-tenth of the cost. All three are vendor-reported.

The showcase results

The announcement leads with demonstrations rather than scores. Anthropic reports that a team of 1,000 Fable 5.2 agents formalized the proof of Fermat’s Last Theorem in Lean in 30 hours, against 11 days for its first formalization, while Opus 5.5 with the same setup completed 26% in 18 hours. It describes a prototype operating system coordinated across more than 1,000 subagents over eight days, a Python-to-Rust service migration that now starts about 26 times faster, and three working electronic devices, each with a circuit board ready for manufacturing in under a day, built with the circuit-board maker Diode. These are single-run, vendor-described showcases. They show what the model can do under heavy scaffolding; they do not predict what it will do on your repository.

Independent evidence

Two independent signals landed within hours of launch, and both are on X. Artificial Analysis said Fable 5.2 took the number one spot on its Intelligence Index v5 at 65, and Cursor said the model topped its own CursorBench.

The Artificial Analysis leaderboard, read on November 11, 2026, shows how the score moves with effort. This is the table to use when you pick an effort level, because it pairs each score with a measured cost and time per task.

Three readings from the table, all ours. High and xhigh tie at 62, but xhigh costs about 63% more per task and takes about 50% longer, so High is the sweet spot if you want most of the lift. Max buys three more index points for roughly three times the cost of High. And Fable 5.2 at High (62, $9.08) beats Opus 5.5 at Max (59, $10.04) on both score and cost per task. Cost per task depends on how many tokens each model spends, which Artificial Analysis measures, so it is not derivable from list prices alone. Artificial Analysis also reports a first-answer-token latency of about 234 seconds at max, a reminder that max is for batch-like work.

Cursor’s numbers come from its own benchmark and are the vendor’s (Cursor’s) report, not an independent run. They do line up with Anthropic’s direction of travel: 49.2% to 56.9% at default effort is a jump of 7.7 points over Fable 5.1. One more caution: a creator video we embed above says Gemini 4 Carbon edges Fable 5.2 on CursorBench, and that claim is attributed to the video; our sibling posts on Carbon and the head-to-head carry the sourced comparison.

Fable 5.2, Opus 5.5 or Mythos 5.2?

Anthropic’s own guidance is unusually direct: begin with Opus 5.5 for most coding and knowledge tasks, then move up to Fable or Mythos for scope and complexity, or down to Sonnet 5.5 or Haiku 5.5 for volume. The platform docs repeat it: move to Fable 5.2 when your evals on an Opus model at higher effort still do not meet your requirements.

The Advisor role comes with an API rule. With the advisor tool, which is in beta, a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor, and any other advisor returns a 400. If you want to keep a cheaper executor and add Fable as the advisor, set the advisor’s model to claude-fable-5-2, which every executor model accepts. A Fable 5.2 advisor returns its advice encrypted, so code that reads the advice text needs to handle the result variant described in the docs.

If you are weighing Fable against other tiers of the same family, see our Claude Opus 5.5 vs Sonnet 5.5 coding comparison and best AI model for coding in October 2026, which predate the 5.2 release. For the previous generation, our Fable 5.1 guide covers the last generation, and the Fable 5 benchmark and prompting guide covers the first release. For rivals, see GPT-6 Astra vs Sol vs Luna pricing and the Gemini 4 Argon guide.

Access, plans and launch offers

Fable 5.2 is available on the Claude API as claude-fable-5-2, on Claude in Amazon Bedrock as anthropic.claude-fable-5-2 and on Claude Platform on AWS, on Google Cloud (both Anthropic-managed and Google-managed) as claude-fable-5-2, and on Claude in Microsoft Foundry. Anthropic’s tweet says it is available on all paid plans. Mythos 5.2 uses the same platforms with the IDs claude-mythos-5-2 and anthropic.claude-mythos-5-2, but only for organizations verified through the Cyber Verification Program or the Life Sciences Verification Program. Enterprise customers can also reach Mythos 5.2 through Claude Security, which Anthropic says scans codebases for vulnerabilities.

Fast mode is available for Fable 5.2 on Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry (Anthropic-hosted), per the announcement. Mythos fast mode is limited to verification-program organizations. The Claude API gives Fable 5.2 a default fast mode rate limit in the Start, Build and Scale usage tiers, so there is no waitlist, and organizations in the Evaluation tier can use it once they move up.

Launch offers by plan

The Claude Help Center article describes two time-limited offers that began on launch day, November 9, 2026. They cover Pro, Max and Ultra plans, standard and premium Team seats, and premium seats on legacy seat-based Enterprise plans. They are not available on the Free plan, standard seats on legacy seat-based Enterprise, self-serve Enterprise, or usage-based Enterprise.

A Fable reset returns the weekly usage you spent on Fable models, leaves other models’ usage in place, and refills the five-hour window if your plan has one. It does not raise limits. Resets went only to accounts with paid billing in place before November 9, 2026 at 10:00 am PT. The help article also warns that Fable models use your limits faster than other Claude models and that your default model does not change, so you pick Fable from the model picker. Conditions can change, so check the article before relying on a reset.

API quickstart

The model ID is the only change on the surface. This Python example uses the Anthropic SDK with the effort parameter set explicitly to the documented default. Note that adaptive thinking is always on, so you do not pass a thinking block, and that the response may start with a thinking block, so select the text block instead of reading the first item.

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

response = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=4096,
    output_config={"effort": "high"},
    messages=[
        {"role": "user", "content": "Review this migration plan and list the three riskiest steps."}
    ],
)

print(next(b.text for b in response.content if b.type == "text"))

The same call in curl uses the standard Messages API headers. Fast mode is a single field, speed, with no beta header, per the release notes.

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-fable-5-2",
    "max_tokens": 4096,
    "output_config": {"effort": "high"},
    "speed": "fast",
    "messages": [{"role": "user", "content": "Summarize the failing tests in this log."}]
  }'

Drop the speed line for standard speed. The fast mode docs show the same field as speed="fast" in the SDK examples; check your SDK version for where it is exposed. The Python example above passes output_config as shown in the AWS post’s effort guidance and is our adaptation, not a copied snippet. On Amazon Bedrock the ID is global.anthropic.claude-fable-5-2 in the AWS launch post’s Boto3 example, which also reminds you that the response can contain a thinking block before the text. To see how the model reasoned, set thinking.display to summarized; the raw chain of thought is not returned, and prompts that ask the model to write its reasoning into the answer should be removed.

Migration checklist from Fable 5.1

Changing the model string is step zero. The release notes and what’s new page list the rest. Work through these in order.

1. Test for history edits. The prefix check now applies to every account. If your code changes system, tools or earlier messages after a thinking block was produced, the request can return a 400 whose message says the block is bound to a different conversation. Nothing changes if Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds your requests, or if your code only appends and returns each assistant turn exactly as received. As a stopgap you can set thinking.block_binding.prefix_mismatch_behavior to drop_block with the thinking-binding-controls-2026-08-01 beta header, but Anthropic warns that dropping blocks can raise output tokens.
2. Check model switches. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks, so a router that falls back to Opus 5.5, Fable 5.1 or Opus 4.8 loses the reasoning for those turns. Dropped blocks are not billed, and the request still succeeds.
3. Move computer use to the toolset. Declaring computer_20251124 returns a 400 invalid_request_error on every platform except Amazon Bedrock. Use computer_toolset_20260801.
4. Update the advisor. A Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor. Changing the advisor model is an edit to tools, so do it in new conversations.
5. Remove forced tool use and manual thinking. tool_choice of any or a named tool returns a 400; keep auto and use strict: true or JSON outputs for schema-valid results. thinking: {"type": "enabled"} with a budget, and thinking: {"type": "disabled"}, both return 400 errors. Note that these two rules are not new if you come from Fable 5.1.
6. Re-run your evals with an effort sweep that starts at the API default, high, then tests medium and xhigh.
7. Plan the retention requirement. Fable 5.2 requires 30-day retention for safety review and is not available under zero data retention unless Anthropic authorizes it.

Anthropic also says the claude-api skill in Claude Code can walk an upgrade, updating model IDs, thinking settings and outdated parameters, and suggesting effort levels. Treat its output like any automated diff and review it.

Effort and prompting notes

The prompting guide reads as a list of behaviors to watch. We summarize the ones likely to cost you money or time. All of the behavior claims below are from Anthropic’s own testing.

• Start at high. It is the API default, and Anthropic says most of the gain is already present there. Medium suits interactive coding and chat, low suits quick turns and subagents, xhigh is for hard work where you measured a gain, and max is for measuring the ceiling.
• The model thinks at medium and above, even on a greeting. An instruction to answer without deliberating did not stop it in Anthropic’s tests. Use a lower effort level for easy turns.
• It can stop early on long tasks, more often at low and medium effort, with a list of what remains or an offer to continue. Anthropic supplies a “finishing work” system prompt block; in replayed sessions at medium effort, a version of it roughly halved sessions that ended with doable work left, at about 30% more output tokens. Keep your own confirmation step for risky or irreversible actions.
• Say who owns the code. With the sentence “This codebase is yours,” the model merged repeated code in about half of small-task tries, against under a tenth without it, at roughly 1.5 times the output tokens and with more unrequested changes. Leave it out where the user wants minimal diffs, and review the diff instead of trusting the summary.
• Subagents can dominate cost. In harnesses that invite delegation the model delegates readily. Set expectations in the prompt if sessions spawn more subagents than the task needs.
• Edit prompts in new conversations only. Because of the prefix check, changing instructions mid-conversation can trigger a 400 when thinking blocks are sent back. Put changes in the newest turn instead.

Safeguards, refusals and data retention

Fable 5.2 runs classifiers that can decline a request. A declined request returns HTTP 200 with stop_reason: "refusal" and a stop_details.category of cyber, bio, frontier_llm, reasoning_extraction or general_harms. In Claude apps, a blocked cyber query is answered by Opus 4.8 and a blocked biology query by Opus 5, per the announcement. In the API you handle it: the permitted fallback targets for Fable 5.2 are Opus 4.8 and Opus 5, and server-side fallback with fallbacks: "default" is in beta on the Claude API. On Bedrock the retry is your code’s job, and the AWS post says a refusal that hit the prompt cache includes a fallback credit token that makes the retry bill as cache reads.

Billing for refusals depends on category. The release notes say refusals before any output are billed again for bio, frontier_llm and reasoning_extraction, mid-stream refusals were already billed, and other categories before output are not. Refusals count against rate limits either way. Benign security or life-sciences work can trigger the classifiers, so the verification programs exist for legitimate teams.

On privacy, Fable 5.2 is a Covered Model with 30-day retention for safety monitoring. Anthropic says Enterprise Frontier Safeguards, which run monitoring on data stored in the customer’s own infrastructure and so work with zero data retention, are in private beta with customers it names, including Bridgewater, Salesforce and Stripe. The AWS post says eligible customers can use Fable 5.2 with zero data retention for internal use through at least June 30, 2027 until EFS ships. On the safety side, Anthropic also says the model has its strongest anti-distillation protections to date, and that preserved thinking now applies to all accounts, which is the reason for the prefix check above.

Claim and evidence table

Limitations

• We did no hands-on testing of latency, tool-calling reliability, long-context accuracy or fast mode speed. The 2.5x and 30% speed figures are Anthropic’s upper bounds.
• Most benchmark numbers come from Anthropic’s launch table. The rival-model cells were run by Anthropic. Only the Artificial Analysis rows and the Cursor post are third-party, and Cursor’s benchmark is its own.
• Artificial Analysis values change as it adds models and index versions. The table here is dated November 11, 2026.
• We did not fetch Opus 5.5 list prices for this post, so we compare it only on Artificial Analysis cost per task. Check the current pricing page before budgeting a switch.
• The migration list covers the changes in the docs we read. It is not a substitute for the full migration guide, especially if you come from Mythos 5.1 or an Opus model rather than Fable 5.1.
• The cost examples are single-turn arithmetic. Real agent sessions include many turns, tool outputs, subagents and retries.
• Launch offers are time-limited and conditional on billing dates and plan type; the help center page was our only source for them.
• The creator video embedded above is secondary commentary. We did not use any of its numbers as facts.

Verdict

Fable 5.2 is a cheaper, faster and, on the available evidence, stronger Fable. The independent Artificial Analysis score of 65 at max and 62 at high is the best evidence we have, and the High setting beating Opus 5.5 at Max on both score and cost per task is the single most useful finding for a budget. Anthropic’s own advice still stands: for most work, begin with Opus 5.5 and move to Fable when your evals say Opus at high effort is not enough.

If you do move, spend the migration effort before the model switch. Audit anything that edits history, fix computer use and advisor settings, remove forced tool use, and run an effort sweep starting at High. Treat fast mode as a latency purchase at double the price, and treat Max as a measurement tool. For how it stacks up against Google’s new model, continue with the Gemini 4 Carbon vs Claude Fable 5.2 comparison or the Gemini 4 Carbon complete guide.

FAQ

Sources

Sources checked November 11, 2026.

• Claude Fable 5.2 and Mythos 5.2 : Anthropic announcement, November 9, 2026 (pricing, fast mode, benchmark table, safety, availability, launch promotion)
• Claude Fable 5.2 overview and What’s new in Claude Fable 5.2 : Claude Platform docs (model ID, context, pricing, breaking changes, feature support)
• Prompting Claude Fable 5.2 and Fast mode : Claude Platform docs
• Claude Platform release notes : November 9 entry (prefix check, fast mode launch, refusal billing)
• Claude Fable 5.2 launch offers : Claude Help Center
• Introducing Claude Fable 5.2 on AWS : AWS Machine Learning Blog, November 9, 2026
• Anthropic on X: @claudeai launch post, @AnthropicAI availability post, @ClaudeDevs developer post
• Artificial Analysis leaderboard (read November 11, 2026) and the @ArtificialAnlys post
• @cursor_ai post on CursorBench
• Gemini 4 Carbon vs Claude Fable 5.2: The Internet Tested Both : YouTube creator roundup, secondary commentary only

Related reading: Gemini 4 Carbon complete guide, Gemini 4 Carbon vs Claude Fable 5.2, Claude Fable 5.1 complete guide, Claude Fable 5 benchmark and prompting guide and best AI model for coding, October 2026.