Claude Fable 5.2: what changes and when to choose it over Opus 5.5

In 60 seconds: Fable 5.2, announced on November 9, 2026, lowers standard API pricing to USD 8/40 per million input/output tokens. Anthropic reports up to 30% faster output than Fable 5.1 and fewer unnecessary blocks. It still recommends starting with Opus 5.5 for most coding and knowledge work. Evaluate Fable for difficult cases your agent cannot yet complete, with explicit effort and a budget per accepted result.

Official sources checked on November 9, 2026 (Europe/Madrid). Speed and behavior figures come from Anthropic. Kiia has not run its own Fable 5.2 benchmarks for this article; the checklist proposes tests for your team.

Verdict by profile

• If your operations or sales team already gets acceptable results from Opus 5.5, keep that route as your baseline. Anthropic’s models overview recommends starting with Opus for most workloads.
• If a long technical investigation, complex migration or analysis with many dependencies keeps failing, test Fable 5.2. The model selection guide recommends moving to Fable when Opus at xhigh or max still falls short.
• If you lack a reviewed sample, an owner for the result or permission to process the data, complete that preparation before moving traffic.

What Fable 5.2 is and what changes from Fable 5.1

Fable 5.2 targets demanding reasoning and long agent runs. Mythos 5.2 shares its underlying model with fewer cybersecurity and biology restrictions; access is restricted to organizations approved through the Cyber and Life Sciences verification programs. Anthropic’s announcement makes Fable the generally available offering.

Change	Fable 5.1	Fable 5.2	Evidence checked on Nov 9, 2026
Standard API: input / output	USD 10/50 per MTok	USD 8/40 per MTok, 20% lower	Official pricing
Output generation	Comparison baseline	Up to 30% faster	Anthropic’s measurement in the announcement
Classifier blocks	Comparison baseline	About 55% fewer in early apps, Code and Cowork testing	Anthropic’s reported result in the same announcement
History and thinking	Prefix checking did not cover every Fable 5.1 account	Covers every account; editing earlier context can cause a 400	What’s new in Fable 5.2

Anthropic presents fewer blocks as a false-positive improvement while maintaining safeguards. Fable still blocks most high-risk cybersecurity and biology tasks. The percentage does not measure refusals in your operation. Faster generation also does not establish the total duration of a workflow that includes CRM calls, searches or human review.

Specifications and effort by channel

The official Fable 5.2 model page and models overview, checked on November 9, list:

Specification	Fable 5.2
Claude API / Claude Platform	claude-fable-5-2
Amazon Bedrock	anthropic.claude-fable-5-2; check your deployment’s inference profile
Context / maximum output	1M / 128K tokens
Thinking	Adaptive, always on
Knowledge cutoff	June 2026
Platforms	Claude Platform/API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry on Azure

Defaults depend on the channel:

Channel	Default Fable 5.2 effort	Source
Claude API	high	Models overview
Claude apps	medium (Medium in the announcement)	November 9 announcement
Claude Code	medium, unless user or organization settings override it	Model configuration

Set output_config.effort in the API; Claude Code offers /effort. Fable 5.2 requires Claude Code v2.1.321 or later. Check /model and /effort status: aliases, gateways and policies can change the effective configuration. Record that difference when comparing an app default with an API call.

Pricing, caching and fast mode

Claude API list prices in USD per million tokens (MTok), checked on November 9, 2026 in Pricing. These cover standard API consumption with global routing, excluding negotiated discounts and regional premiums; calculate taxes and provider charges separately.

Usage	Fable 5.1	Fable 5.2	Opus 5.5
Uncached input	10	8	4
Output	50	40	20
Cache write, 5 minutes	12.50	10	5
Cache write, 1 hour	20	16	8
Cache read	0.25	0.20	0.20

Fable 5.2 costs twice as much per uncached input and output token as Opus 5.5; cache reads have the same price. Total tokens, tools and retries also affect the bill. Measure cost per accepted case.

Fable 5.2’s optional fast mode costs USD 16/80 per MTok of input/output. Anthropic reports up to 2.5 times more output tokens per second than standard mode, without promising that improvement in time to first token or total task duration. Caching multipliers apply on top of fast rates; do not price fast cache reads at the standard USD 0.20 rate.

The announcement and documentation list fast mode on Claude Code, Claude API/Platform, Google Cloud (Anthropic managed or Google managed) and Anthropic-hosted Foundry. Foundry uses the separate claude-fable-5-2-fast model; Google sets pricing on its Google-managed service. Check availability and quotas for your channel. Fast mode is unavailable with Batch API; do not extend its availability to Bedrock or Claude Platform on AWS.

Comparison with the same scope: Fable 5.2 and Opus 5.5

This table compares models integrated into your own agent through an API. Subscription and connector capabilities depend on the surrounding product.

Criterion	Fable 5.2	Opus 5.5
Verified entry price	USD 8 per input MTok, standard global API	USD 4 per input MTok, same scope and date; taxes and cloud charges separate
Where processing and data live	Cloud service. Its model page requires 30-day retention unless Anthropic expressly authorizes ZDR	Cloud service. Confirm your account’s region, retention and contract; this guide does not certify ZDR
MCP status	Not a model attribute; check each connector and its GA/beta status	Same requirement
Authentication	API/cloud credentials plus credentials for each agent tool	Same requirement; changing models does not configure identities
Respects user permissions	Your application must check ACLs and restrict actions	Same requirement; model choice does not demonstrate this control
Integrations	Tool use; your integration must validate each contract	Tool use; validate the same tools and permissions
Limits	1M context, 128K output; quotas vary by account and platform	Same token limits; quotas vary by account and platform
LATAM support	Confirm supported country, billing and contracted support; test your Spanish	Confirm the same fields; this guide verifies no regional SLA

Sources: models, Fable’s model page and API retention, checked November 9. Permission and integration rows are team evaluation requirements, not Kiia test results.

Advisor and Orchestrator

Anthropic proposes two roles for Fable: an Advisor that other models consult on difficult problems, or an Orchestrator that delegates simpler work to smaller models. You could evaluate a targeted consultation from a sales agent or coordination of technical research, keeping permissions unchanged and counting every call’s cost. These are candidate designs, not implementations tested here. The Advisor tool is in beta and has executor/advisor compatibility rules.

Plans and the launch promotion

The official offer, checked November 9, distinguishes:

• Pro and standard Team: Fable up to 50% of weekly limits through Nov 16, 2026, 11:59 pm PT; an eligible reset expires then.
• Max, Ultra, premium Team and premium legacy seat-based Enterprise: one to three resets by subscription age, through Mar 8, 2027, 11:59 pm PT.

Resets require paid billing before Nov 9, 10:00 am PT, and prior seat membership. Trial conditions apply. Free and several Enterprise variants are excluded; check your plan in the source. This is subscription usage, not API credit or a permanent quota increase.

Checklist before routing real traffic in LATAM

These are proposed tests for your operation. Keep their results separate from Anthropic’s reported percentages.

1. Select the same cases for Fable 5.1, Fable 5.2 and Opus 5.5: reconciliations, sales drafts or investigations your team can review. Keep prompts, tools, permissions and rubric unchanged. Assign an owner to accept each result.
2. Record model ID, provider, explicit effort and standard/fast speed for every run. Also test Opus at higher effort before attributing a capability gap’s resolution to Fable. Each model calibrates effort differently.
3. Review parsing: select blocks by type, without assuming content[0] is text. Preserve complete thinking blocks and avoid editing history. The migration guide details prefix checking and the loss of reasoning when moving back to models that cannot read Fable 5.2 blocks.
4. Validate tools. tool_choice with any or tool returns 400; use auto or none and validate arguments. For computer use, check migration to computer_toolset_20260801, with the documented Bedrock exception. What’s new separates these constraints from changes specific to 5.2.
5. Check stop_reason, including refusal, and stop_details.category. An HTTP 200 can contain a refusal. Record the actual responding model and fallback costs using the refusals guide. Also test incomplete output caused by max_tokens, which includes thinking.
6. Measure accepted completions, human corrections, tool errors and total latency. Add tokens, caching, subagents, retries and infrastructure. Divide total sample cost by accepted cases and report USD plus local currency, recording the conversion date and exchange rate used.
7. Include your customers’ Spanish, records mixed with English, product names, currencies and authorized internal documents. Evaluate by country and process; general multilingual ability does not validate your operational vocabulary.
8. Confirm region, retention and quotas. Start with limited traffic, read actions or human approval, alerts and a rollback route. Expand according to criteria your team has defined.

When to choose neither

If the workflow lacks scoped permissions or a verifiable way to accept results, keep the current process while you resolve that. If it requires zero retention, clarify the contract before testing real data: Fable 5.2 is unavailable under ZDR without Anthropic’s express authorization. Choosing Opus does not certify your deployment by itself either.

Next step

Pick one case that currently needs human correction and write down what the agent would have to deliver for acceptance. Use the Opus 5.5 guide for teams to review your current integration. The Opus 5 guide and Sol versus Opus comparison provide context on earlier generations; their prices and plans have their own review dates. For Fable 5.2, use this article’s verified rates and record a fresh review before choosing the production route.