Claude Fable 5.2 (9 Nov 2026): 20% cheaper, and the 400 errors to fix first

Summary. Anthropic released Claude Fable 5.2 (claude-fable-5-2) on 9 November 2026 on the Claude API, Claude Platform on AWS, Amazon Bedrock, both Google Cloud options and Microsoft Foundry. It lists at $8 per million input tokens and $40 per million output tokens, 20% below Claude Fable 5.1's $10 and $50. Cache reads fall from $0.25 to $0.20 per million, batch costs $4 and $20, and fast mode costs $16 and $80 for what Anthropic describes as up to 2.5 times faster output. The context window is 1M tokens, max output is 128k, adaptive thinking is always on and the default effort is high. Anthropic still tells most teams to start on Claude Opus 5.5 ($4 and $20) and move up only when Opus evals at higher effort fall short. Four changes can break Fable 5.1 code: the thinking prefix check returns a 400 error on every account, whatever its age; only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks; the older computer_20251124 tool returns a 400 everywhere except Amazon Bedrock; and a Fable 5.2 executor in the advisor tool needs a Fable 5.2 or Mythos 5.2 advisor. The model also requires 30-day data retention. Benchmark figures below are Anthropic's; we have not tested the model.

About this article. This is for engineering leads and platform teams running Claude in production who are deciding whether, and how, to move from Fable 5.1 or Opus 5.5. We checked every price, model ID, date and API behaviour against Anthropic's release notes, model page, migration guide, preserved-thinking, pricing, rate-limit, deprecation and data-retention documentation and the launch post, on 11 November 2026. Capability figures are Anthropic's claims, not our measurements. Vendors change prices and API behaviour; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What Anthropic shipped on 9 November 2026

The release notes for 9 November 2026 announce two models. Claude Fable 5.2 needs no access approval on the Claude API. Claude Mythos 5.2 (claude-mythos-5-2) has the same capabilities and prices, but only organisations verified through Anthropic's verification programmes, such as the Cyber Verification Program, can call it. The launch post says the two are the same underlying model, and that Fable carries cybersecurity and biology safeguards so it can be offered to all users.

Both models have a 1M-token context window (default and maximum), 128k max output tokens, always-on adaptive thinking and the tokenizer introduced with Claude Opus 4.7. The reliable knowledge cutoff is June 2026. The model ID is claude-fable-5-2 on every platform except Amazon Bedrock, where it is anthropic.claude-fable-5-2. The pricing page says Claude 4.6 and later models, Haiku 5.5 excepted, include the full 1M window at standard pricing, so a 900k-token request costs the same per token as a 9k-token one.

Four lifecycle and capacity details matter for planning:

• Retirement floor. The deprecations page lists claude-fable-5-2 as Active with a tentative retirement date of not sooner than 9 November 2027.

• Fable 5.1 is not deprecated. claude-fable-5-1 stays Active with a floor of 1 September 2027, and Anthropic promises at least 60 days' notice before retiring a publicly released model. Cloud platforms can set earlier dates; our Claude model retirement calendar for 2027 tracks them.

• Separate, smaller rate limit. Fable 5.2 has its own rate limit, outside the combined bucket that Fable 5.1 and Fable 5 share, with lower token limits than Opus 5.5 at each tier. At the Start tier that is 500,000 input and 100,000 output tokens a minute, against 2,000,000 and 400,000 for Opus 5.5.

• No Priority Tier. Neither Fable 5.2 nor Mythos 5.2 is supported on Priority Tier.

Fast mode (the speed field set to fast) works in Messages API requests on the Claude API and on both Google Cloud options, where Google sets its own prices on the Google-managed one. On Microsoft Foundry deployments hosted on Anthropic it is a separate claude-fable-5-2-fast model. It is not offered for Fable 5.2 in Claude Managed Agents or through the Batch API.

Fable 5.2 pricing against Fable 5.1 and Opus 5.5

Every figure below comes from Anthropic's pricing page, in US dollars per million tokens (MTok) at standard first-party Claude API rates.

Three pricing notes are easy to miss. Cache reads on Fable 5.2 cost 0.025 times base input, the same multiplier as Fable 5.1, while Opus 5.5 uses 0.05 times, which is why both land at $0.20 even though Fable's base price is double. US-only inference through the inference_geo parameter adds a 1.1x multiplier to every token category on Claude 4.6 and later models. And fast mode pricing stacks with other modifiers, including prompt caching and data residency multipliers.

The 20% cut holds on list price, but cost per task depends on how much the model thinks; the worked example below shows where the saving disappears.

What Anthropic says Fable 5.2 does better

Everything in this section is a vendor claim from Anthropic's launch post, measured by Anthropic on its own harnesses. We have not reproduced any of it.

Two caveats belong in any business case. The comparisons are effort-specific and depend on Anthropic's scaffolding, so your cost ratio will differ. And the docs say that at high effort and above the model thinks before almost every reply. Run your own evals before you trust any row.

Fable 5.2 keeps the safety classifiers of earlier Fable models, which block high-risk cyber work such as penetration testing and exploit generation, and most research-biology tasks. A declined request returns HTTP 200 with a stop_reason of refusal, and the permitted fallback targets are Claude Opus 4.8 and Claude Opus 5. Authorised security teams should apply for Mythos 5.2 instead.

The four breaking changes, and the 400s that carry over

Anthropic's What's new page names four changes from Fable 5.1 that can break code. Three return HTTP 400 errors; one fails silently.

1. The prefix check now applies to every account

Each Fable 5.2 thinking block is bound to the system prompt, the tools and every message before it. If your code changes any of those between requests and then sends the later thinking block back, the block fails the prefix check. On Fable 5.1 that request gets a 400 only on accounts created on or after 00:00 UTC on 31 August 2026. On Fable 5.2 and Mythos 5.2 it gets a 400 on every account. So code that runs cleanly on an older account today can start failing the moment you change the model ID.

The error message says the block is bound to a different conversation, and retrying the same body does not clear it. Common triggers in home-grown agent loops: rebuilding the system prompt with the current date, clearing or shortening old tool results, editing tools mid-session, dropping a thinking block from the middle of the history, and rebuilding saved sessions from templates.

One trigger is new with Fable 5.2. If you use the MCP connector without pinning a server's tool list, a tool list that changes between requests counts as an edit for Fable 5.2 thinking blocks, though not for Fable 5.1 or Opus 5.5 blocks. Pin the list; pinning is a beta feature of the connector.

Nothing changes if Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds your requests. The risk sits with teams that assemble messages themselves and with gateways that rewrite requests.

2. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks

On the Claude API and Google-managed Google Cloud, Fable 5.2 reads thinking blocks from Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1 and older models, so moving a conversation onto Fable 5.2 keeps its reasoning. The reverse does not hold. If a router, a client retry or a refusal fallback moves a conversation from Fable 5.2 to any other model, Fable 5.1 and Opus 5.5 included, the API drops the Fable 5.2 blocks. The request succeeds and the dropped tokens are not billed, but the next model works without that reasoning, which Anthropic says can raise cost and latency on the first turn after the switch. Fable 5.2 blocks are also tied to the account that produced them, so replaying stored conversations through a different, unlinked account loses them without an error.

3. The older computer use tool returns a 400 outside Amazon Bedrock

Fable 5.1 accepted computer use either as the computer_toolset_20260801 toolset or, with the computer-use-2025-11-24 beta header, as the older computer_20251124 tool. Fable 5.2 accepts only the toolset on every platform except Amazon Bedrock, where the older tool still works. The migration guide says to drop the beta header and declare the toolset with no name or display dimensions:

            {
  "tools": [
    { "type": "computer_toolset_20260801" }
  ]
}
          

The agent loop also changes with the toolset. Our piece on the seven computer-use changes at general availability walks through them.

4. The advisor tool needs a Fable 5.2 or Mythos 5.2 advisor

With the advisor tool, which is in beta, the advisor must be at least as capable as the executor. A Fable 5.2 executor accepts only claude-fable-5-2 or claude-mythos-5-2 as advisor; a Fable 5.1 advisor returns a 400 invalid_request_error. Changing the advisor's model is itself an edit to tools, so make the change in new conversations or the prefix check will reject old thinking blocks. A Fable 5.2 advisor returns its advice encrypted, as Fable 5.1 and Opus 5.5 advisors do, which matters if your code reads the advice text.

Every 400 to check before you switch

The table draws on the What's new page, the migration guide and Anthropic's thinking documentation.

One more trap sits in organisations with customer-managed encryption keys: structured outputs, covering both strict tool use and JSON outputs, are not available for Fable 5.2 there. Remove strict: true and output_config.format and validate in your own code.

How to find prefix edits before production does

Start by diffing what you send. Capture request bodies for a few normal turns, including a compaction and a tool change, and compare system, tools and the shared messages for each pair of consecutive requests. They should be identical up to the newly appended turns.

Then confirm against the API. Anthropic's test works from any account, because setting the mismatch behaviour opts the request into the check. Run a normal multi-turn session through your integration on claude-fable-5-1, and also on claude-fable-5-2 if you use the MCP connector without pinning, with the thinking-binding-controls-2026-08-01 beta header and the drop behaviour set:

            {
  "model": "claude-fable-5-1",
  "max_tokens": 16000,
  "thinking": {
    "type": "adaptive",
    "block_binding": { "prefix_mismatch_behavior": "drop_block" }
  },
  "messages": []
}
          

Log the top-level input_transformations array on every response. Each entry with the reason prefix_binding_mismatch marks a block your code invalidated.

In production, Anthropic recommends leaving prefix_mismatch_behavior at its default, error, if a mismatch can only mean a bug. Treat drop_block as a stopgap: the model loses the dropped reasoning, the prompt cache restarts at the edit, and output tokens can rise. The lasting fixes are in our guide to Claude mid-conversation tool changes: add tools with tool_addition blocks, put changing context in the newest turn, and use on-demand compaction. For Fable 5.2 on the Claude API, Claude Platform on AWS and Anthropic-managed Google Cloud, tool changes that carry a full definition use the inline-tools-2026-09-15 beta header. The same discipline keeps your prompt cache warm.

Anthropic ties this enforcement to distillation: its launch post says preserved thinking now applies to all accounts, to stop distillers editing earlier context to extract the model's reasoning. Expect the check to stay.

Behaviour changes that cost money without an error

The migration guide also lists changes that alter cost or latency without an error:

• Default effort is one step higher if you come from Opus 5.5. Opus 5.5 defaults to medium and Fable 5.2 to high, so a request that omits effort runs harder and longer. Set effort explicitly.

• Prompt lines asking for less thinking stop working. Fable 5.1 usually obeys a line telling it to answer without deliberating. In Anthropic's testing, Fable 5.2 kept thinking at medium and above. Remove those lines and lower the effort instead.

• Small output limits can produce no text. Thinking counts toward max_tokens, so a request sized for Fable 5.1's short answers can stop with a stop_reason of max_tokens after a thinking block and before any text.

• Self-review subagents. At xhigh and max in an agentic coding product, Anthropic saw Fable 5.2 start subagents to review its own work, which Fable 5.1 did not do on the same tasks. Subagents can be a large share of session cost.

Anthropic's advice is to start at high, step to xhigh only where you have measured a gain, and reserve max for ceiling tests.

Worked cost example: 1,000 long research tasks a month

Assume an analyst agent that runs 1,000 long tasks a month. Each task sends 150,000 uncached input tokens, reads 1,200,000 tokens from the prompt cache, writes 50,000 tokens to the 5-minute cache and produces 60,000 output tokens, thinking included. These are our assumptions; count your own prompts with Anthropic's Token Counting API, since Fable 5.2 uses the tokenizer introduced with Opus 4.7. The table applies the list prices above.

For Fable 5.2 the sum is 0.15 x $8 + 1.2 x $0.20 + 0.05 x $10 + 0.06 x $40 = $1.20 + $0.24 + $0.50 + $2.40 = $4.34. The fast-mode column is our own arithmetic: it applies the standard cache multipliers (0.025x for reads, 1.25x for 5-minute writes) to the $16 fast-mode input price, because the pricing page says caching multipliers apply on top of fast mode pricing.

On these assumptions, moving from Fable 5.1 to Fable 5.2 saves $1,085 a month ($5,425 minus $4,340), exactly the 20% list cut. The saving holds only if output stays flat. Fable 5.2's non-output cost is $1.94 a task, so it costs more than Fable 5.1's $5.425 once output passes about 87,100 tokens a task ($5.425 minus $1.94, divided by $40 per million), roughly 45% above the 60,000 assumed. A model that thinks before almost every reply at high can get there. Anthropic's counter-claim is that at Medium effort Fable 5.2 completes challenging tasks with fewer tokens than Fable 5.1; measure it on your traffic.

Two more adjustments: US-only inference at 1.1x lifts the Fable 5.2 month to $4,774 ($4,340 x 1.1), and the Batch API halves base input and output for non-urgent tasks, though fast mode is not available with it. If Opus 5.5 at high or xhigh passes your evals, it remains the cheaper line at $2,290.

Should you move? A decision table

For broader model choice by workload, see our best LLM for coding in 2027 cost comparison.

Data retention, India and the EU

Fable 5.2 is one of Anthropic's Covered Models. It requires 30-day data retention and is not available under zero data retention (ZDR) unless Anthropic expressly authorises it. On the Claude API, a request from an organisation or workspace without 30-day retention returns a 400 invalid_request_error. Organisations with a ZDR arrangement can turn on standard 30-day retention for a single workspace while the others keep ZDR; the Default Workspace always follows the organisation's setting, so use another one. On Amazon Bedrock and Google Cloud's Agent Platform, retained data stays within your cloud provider's environment, while on the Claude API, including Claude Platform on AWS, Anthropic handles it.

The launch post adds a route for enterprises that cannot accept retention: Enterprise Frontier Safeguards, in private beta, runs automated safety monitoring on prompts and responses stored in the customer's own infrastructure. Eligible customers can keep using Fable 5.2 under ZDR before they implement it.

For Indian and EU teams, the practical question is where prompts containing personal data are retained and for how long. For those workloads the retention setting belongs in the data-protection assessment and processor terms under the DPDP Act and the GDPR, not only in an engineering ticket. This article is not legal advice. Use of Fable 5.2 is also governed by Anthropic's Usage Policy, whose updated version takes effect on 12 November 2026; our guide to Claude's November 2026 Usage Policy for HR, lending and GEO teams covers what changes.

Migration checklist

1. Confirm 30-day retention on the workspace that will call Fable 5.2, and check its separate rate limit.

1. Run the prefix test with drop_block and fix every prefix_binding_mismatch entry, even if Fable 5.1 never returned the error for you.

1. Pin MCP connector tool lists, move computer use to computer_toolset_20260801, and point advisors at claude-fable-5-2.

1. Audit routers and fallbacks: any path from Fable 5.2 to another model drops its reasoning.

1. Handle a stop_reason of refusal, and name only Opus 4.8 or Opus 5 in explicit fallback lists.

1. Set effort explicitly, raise small max_tokens values and compare cost per completed task.

1. Change the model ID last, behind a flag, on a slice of traffic.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI evals and observability service measures quality, refusals and cost per completed task on your own traffic, so a model switch like this one rests on evidence rather than list prices. Talk to us at /contact-us/.

Last updated: 11 November 2026.

References

1. Claude Platform release notes, 9 November 2026 entry — Anthropic

1. Pricing — Anthropic

1. Introducing Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026 — Anthropic

1. What's new in Claude Fable 5.2 — Anthropic

1. Claude Fable 5.2 and Claude Mythos 5.2 migration guide — Anthropic

1. API and data retention — Anthropic

1. Claude Fable 5.2 model overview — Anthropic

1. Model deprecations — Anthropic

1. Rate limits — Anthropic

1. Service tiers — Anthropic

1. Fast mode — Anthropic

1. Preserved thinking — Anthropic

1. MCP connector — Anthropic

1. Advisor tool — Anthropic

1. Thinking — Anthropic

1. Usage Policy, effective 12 November 2026 — Anthropic