Claude Fable 5.2 Is Here. When Should You Use It Instead of Opus?

Claude has a new option for the job that keeps getting stuck. Fable 5.2 targets demanding reasoning and extended work with tools: investigating conflicting evidence, navigating an unfamiliar codebase or turning a messy brief into a finished deliverable.

Start with Opus 5.5, raise its effort if the result falls short, then consider Fable 5.2 for work valuable enough to justify the extra capacity. That follows Anthropic’s model-selection guidance. The independent results give Fable a stronger case than the launch announcement alone, especially below its most expensive effort setting.

In Claude chat, open the model name beside Send, then Effort; in the Claude Code terminal, run /effort for the slider. Workspace administrators can restrict the available levels.

Our assessment combines current documentation and published evaluations with a small Claude Pro comparison by The Choosely Team, described below.

What changes from Fable 5.1?

Fable 5.2 keeps the large working space: a million-token API context window and up to 128,000 output tokens. Its standard input and output rates fall 20% from 5.1.

Anthropic reports clearer communication, output generation up to 30% faster than 5.1 and fewer safeguard blocks. It describes complex software and device-design work as examples of the model’s reach. Gains are uneven: some AI-development tasks improve only to Opus 5.5 levels. These are the company’s findings; the independent evidence is more useful for comparing the purchase.

Fable versus Opus: effort changes the answer

Artificial Analysis publishes the following results on its Intelligence Index v5.0.1, with default fallback enabled. Costs are its average per index task, in USD.

The useful finding is Fable High: it scores above Opus Max at a lower measured task cost. Max buys Fable another three index points at substantially higher cost. Raising effort can change how much reasoning the model uses, even when the price per token stays the same.

Those are specific evaluation configurations. Anthropic describes Medium as Fable’s app and Claude Code default, while its API default is High. A normal chat session therefore starts at a different setting from the Fable results above.

The November 9 Vals Index also favors Fable: first of 36 at 71.31% and $17.94 per test, against Opus’s 70.14% and $24.77. Vals evaluated Fable at High and Opus at Max. The score gap is modest and the reported uncertainty ranges overlap. Sonnet 5.5 sits between them at 70.85% and $16.18 per test, ahead of Opus and cheaper than either model here; our Sonnet versus Opus guide covers that choice.

The picture changes with the task. On Vals’s Terminal-Bench 4.0, both score 65.15%, with Fable costing $14.56 per test against Opus’s $12.04 in the same Mini-SWE-agent harness. For knowledge work, the Vals Index’s component breakdown puts Fable behind Opus on Tax Agent (32.1% versus 44.5%) and Legal Research (46.2% versus 50.5%). Artificial Analysis’s GDP.pdf component also favors Opus Max: 26.2% against Fable Max’s 24.2%.

Fable’s aggregate lead earns it a serious trial on difficult work. Opus still wins some comparisons. Use these results to choose what to compare, then judge the finished work, corrections and cost on your own task.

Our small comparison: both got the decision right

The Choosely Team gave each model a fictional six-document supplier-selection task covering prices, cancellation rules, deletion terms and a later quote amendment. Both correctly calculated every supplier’s costs and selected the only eligible option, passing all ten checks in our prewritten answer key.

Fable’s answer was shorter; both met the requested 450-word cap. On this task, switching to Fable added no accuracy benefit. This was one response per model; effort settings were not independently verified, and the usage meters could not establish precise cost. The timings describe this pair rather than typical model speed.

Which Claude plans include Fable?

The ordinary plan rules separate access from included usage. The dated launch offer below temporarily changes Pro and standard Team billing.

All Fable versions share that 50% allocation. Fable draws from the existing weekly pool faster than other models. For the wider capacity decision, see Claude Max and Ultra limits.

Select Fable 5.2 in the model picker. Claude Code needs version 2.1.321 or later; Cowork needs the latest Claude Desktop. Workspace controls can restrict availability. Enterprise starts disabled until an admin enables it. Seven days after launch, Fable-enabled roles gain 5.2 and Fable defaults move to it unless an admin opts out.

Enterprise note: Legacy seat-based premium seats include the 50% share; standard seats require organization-enabled usage credits. Usage-based Enterprise bills at API rates. Self-serve Enterprise customers should check their contract.

Launch offer: checked November 10, Sydney time

Under the launch terms, Pro and standard Team seats get Fable within 50% of weekly limits through November 16, 2026, 11:59 p.m. Pacific, plus one eligible reset expiring then. Enabled credits apply afterward.

Max, Ultra, Team premium and legacy Enterprise premium get resets expiring March 8, 2027, 11:59 p.m. Pacific: one below six months’ paid tenure, two at six to 12 months, three above 12 months.

Resets require paid billing before November 9, 10 a.m. Pacific; workspace billing and seat joining must meet that cutoff. See the terms for trial exceptions. Other Enterprise variants are excluded from the offer; HIPAA-ready Enterprise cannot use resets.

Open Settings > Usage > Limit resets > Reset for free, even before reaching a limit. This restores weekly Fable usage and any five-hour window. Limits stay the same. Resets are irreversible; canceling or downgrading forfeits unused resets.

What API use costs

API billing is separate from a Claude subscription. These are Anthropic’s standard/global rates, in USD per million tokens:

Caching reuses processed input. Fable 5.2 and Opus 5.5 share the same cache-read rate, so cache-heavy work has a different price relationship from fresh input.

For a hypothetical 250,000 fresh input tokens and 25,000 total billable output tokens, the bill is $3 on Fable 5.2, $3.75 on Fable 5.1 or $1.50 on Opus 5.5. Fable fast mode makes it $6. This holds consumption constant and excludes caching, tools, taxes and other charges. Actual reasoning and retries change the total.

The full context window carries no separate long-context rate premium. US-only inference adds 10%; tools and partner-cloud terms can add costs. These API changes leave subscription prices and allowances separate.

Fast mode uses the same model with faster inference. Anthropic claims up to 2.5 times the output-generation speed. Availability depends on the hosting route, and batch cannot be combined with fast mode.

Safeguards can change which model answers

Fable 5.2 and Mythos 5.2 share an underlying model, according to Anthropic. Public Fable has additional cybersecurity and biology safeguards.

For users, that can mean an automatic switch to Opus 4.8 for offensive cybersecurity requests, or Opus 5 for biology and certain AI-development work. Anthropic warns of high cybersecurity fallback rates and advises against Fable for professional biology research at present. Routine security help can work. Check the model label: after a fallback, the picker stays on Opus until you switch back manually. Files, search results and memory can also trigger a block.

Switching is on by default in Claude apps; API customers must configure it. Artificial Analysis’s “Default Fallback” label includes this behavior. For API billing, the fallback response uses the answering model’s own rate, while some blocked Fable processing can also be charged.

Mythos’s more permissive access requires verification. The Cyber Verification Program lets qualifying individuals apply for Defense Access; other tiers require organizations. Approval still needs provisioning. Life Sciences access is for qualifying organizations on Team, Enterprise or the API. Claude Security can also deliver Mythos-powered findings within that product’s scope.

Builder checks before switching

Read the migration guide, especially these changes from 5.1:

• Saved thinking: Every account gets conversation-prefix checks. Editing earlier messages, prompts or tools can invalidate preserved thinking and cause a 400 error. Blocks belong to the originating or linked account; only 5.2 siblings can read them.
• Computer use: Adopt computer_toolset_20260801. The older computer_20251124 errors except on Bedrock.
• Advisor pairing: A Fable 5.2 executor requires a Fable 5.2 or Mythos 5.2 advisor.

Always-on adaptive thinking and restrictions on forced tool calls, prefill and sampling are inherited from 5.1. Set effort explicitly. Also check retention eligibility: the default is 30 days unless Anthropic authorizes an exception for this model.

Choosely’s Verdict

Fable 5.2 deserves a place alongside Opus for difficult work. Its High-effort results make a stronger value case than judging it only at Max.

Our small task needed no switch to deliver a correct result. Keep successful routine work where it is. When Opus at higher effort still misses a valuable requirement, give Fable the same material and judge whether it reduces corrections or failed attempts enough to earn its capacity. For lighter jobs, Sonnet remains worth considering before either model.

Before the next difficult job, check Settings > Usage, note your remaining allowance and choose an effort setting deliberately. That gives the comparison a useful cost boundary.

FAQ

Is Fable 5.2 free?

Claude Free has no access. Eligible Pro and standard Team seats receive temporary included usage under the dated offer above; ordinary billing uses credits.

Does Fable 5.1 retire immediately?

Anthropic’s model-status table still lists it as active, with retirement no sooner than September 1, 2027. There is no announced shutdown date.

Should I buy a higher plan just for Fable?

Check your current access and compare a recurring task first. Our premium-plan comparison covers the separate monthly purchase decision.