Claude Fable 5.2 vs Opus 5.5: When the 2x Price Pays Off

$8 per million input tokens and $40 per million output. That is the price of Claude Fable 5.2, released November 9, and it is exactly double the $4 and $20 Anthropic charges for Opus 5.5. The launch coverage led with the other comparison: Fable 5.2 is 20% cheaper than Fable 5.1, which cost $10 and $50. That matters to the small number of teams already running Fable 5.1. For everyone else, the question is whether a model that costs twice as much per token is ever the cheaper choice per finished task.

Anthropic’s own answer depends on the effort setting, and its own launch page still calls Opus 5.5 “our strongest daily driver” and the default place to start. Dianne Penn, Anthropic’s head of product management for research, told CNBC that Fable 5.2 is for “power users or power use cases.” This guide is about finding out whether your work is one of those cases, and about the integration details that will bite before you get an answer.

Per token it is 2x. Per task, Anthropic says it can be half.

The most useful material in the announcement is not the headline benchmark table. It is the set of effort comparisons, where Anthropic pits Fable 5.2 at a lower effort setting against Opus 5.5 at its highest. All of these are vendor figures from the launch page, not independent results:

Benchmark	Fable 5.2 setting	Opus 5.5 setting	Anthropic’s claim
FrontierCode 2.0 (agentic coding)	High: 45.3%	Max: 45.5%	Roughly half the cost per task
WANDR	Medium	Max	Matches the score for about 40% of the cost
Humanity’s Last Exam	Low	Max	Matches the score for under a tenth of the cost
Chartography (no tools)	Medium: 66.4%	Max: 64.4%	Higher score, cost not stated

The mechanism is token count. A stronger model often reaches the same answer in fewer reasoning steps and fewer tool calls, so a higher rate times a smaller bill of materials can come out lower. Our effort settings explainer covers why the dial moves cost more than the model choice does on most agentic work.

Caching narrows the per-token gap further. Fable 5.2’s cache reads cost $0.20 per million tokens, which is 2.5% of its input price. Opus 5.5 reads cost 5% of $4, which is also $0.20, as the migration guide confirms. On a long agent loop where most of the context is a cached prefix, the two models pay the same for the part that dominates the token count.

Here is a worked example, using our own arithmetic on list prices and ignoring cache writes. Take an agent task that reads 2 million input tokens with 90% served from cache and writes 40,000 output tokens:

Opus 5.5: 200,000 uncached at $4 is $0.80, 1.8 million cached at $0.20 is $0.36, 40,000 output at $20 is $0.80. Total: $1.96.
Fable 5.2: $1.60 plus $0.36 plus $1.60. Total: $3.56.

Same tokens, so Fable costs 1.8 times as much, not 2 times. For Fable to break even on that task it needs to finish with roughly 45% fewer tokens across the board than Opus, or in fewer attempts. Whether it does is an empirical question about your workload, and the “half the cost” claim on FrontierCode is one vendor data point on one benchmark. Build the eval before you believe it.

Where Opus 5.5 still wins on Anthropic’s own table

Read the full benchmark grid and Fable 5.2 does not sweep it. Three rows deserve attention before any team moves its default:

AA-Briefcase v1.1 (Elo): Opus 5.5 scores 1800 and Fable 5.2 scores 1785. On the rubric version of the same benchmark Fable leads, 66.1% to 61.2%, but on the Elo ranking the cheaper model edges it.
AutomationBench: Gemini 4 Carbon scores 53.9% against Fable 5.2’s 51.4%. Carbon is still gated to cyber defenders, so this does not change anyone’s options today, but it does say Fable is not the ceiling on workflow automation.
OSWorld 2.1: Fable 5.2 leads at 89.3%, but Opus 5.5 is at 87.2%. Two points on computer use is not worth double the per-token rate for most desktop automation.

Where the gap is large is in long, novel work: Remote Labor Index (41.3% against Opus 5.5’s 21.3%), RoboDojo Sim (64.9% against 39.4%), and Terminal-Bench-Science (74.6% against 58.7%). Anthropic’s demonstration was a team of 1,000 Fable 5.2 agents that formalized Fermat’s Last Theorem in 30 hours, while Opus 5.5 completed 26% of the work in 18 hours. The pattern is consistent with Penn’s description: Fable earns its rate on long horizons and problems nobody has a template for. On bounded tasks with a known shape, Opus 5.5 at a high effort setting is the rational default, and Sonnet 5.5 is often enough below that.

A blocked request comes back as a success

This is the integration detail most likely to cause a quiet failure. Fable 5.2 runs cyber and biology safeguards, and when a classifier declines a request, the API does not throw an error. Anthropic’s refusals and fallback documentation is explicit: “A refusal is a successful HTTP 200 response with stop_reason: \"refusal\".” A stop_details object carries a category (cyber, bio, frontier_llm, reasoning_extraction or general_harms) and an explanation.

Any pipeline that checks only the HTTP status will treat that as a completed task. If your code writes the response text into a ticket, a document or a downstream agent’s context without reading stop_reason, you will get empty or refusal text flowing through as output. We saw the same failure shape in October, when an HTTP 200 carried another customer’s data; a success code is not a correctness check.

Fallback is on in the apps and off in the API

In Claude’s apps, a flagged message switches to another model automatically. Anthropic’s help article on model switching says blocked requests go to “Opus 5 for biology, chemistry, and life sciences requests, and for frontier LLM development requests,” and to “Opus 4.8 for offensive cybersecurity technique requests.” The user sees a notice and the response is labeled with the model that answered.

On the API, the same article says “Automatic switching isn’t active by default.” You opt in by sending "fallbacks": "default" (or a list of up to three models) with the server-side-fallback-2026-07-01 beta header. That beta runs on the Claude API and on Claude Models on Google Cloud (Anthropic managed), and not on the Batches API. When fallback fires, the top-level model field reports the model that actually answered, and the response contains a fallback content block naming the original model, the replacement and the trigger category.

Two consequences for anyone running evals or audits:

Log the model field per response, not per request. A security team running Fable 5.2 over log analysis or detection rules may find a share of its answers came from Opus 4.8, an older model, and its quality numbers will be a blend.
Decide which you want for each workload: a refusal you can route yourself, or a silent downgrade. For regulated work, an explicit refusal that your own router sends to a model you chose is easier to defend than a substitution you discover later in the logs.

Anthropic says Fable 5.2 produced about 55% fewer blocks than Fable 5.1 in early testing, about 60% fewer in cybersecurity and 40% fewer in biology. Fewer is not none. Organizations doing legitimate offensive security or life sciences work can apply for Mythos 5.2, the same underlying model with looser safeguards, through the Cyber Verification Program or the Life Sciences program.

What breaks when you swap the model ID

The migration guide calls the move from Fable 5.1 “mostly drop-in.” Coming from Opus or Sonnet, there is more to check. From the Fable 5.2 model page and the migration guide:

Data retention is a hard requirement

“Claude Fable 5.2 requires 30-day data retention and isn’t available under zero data retention unless expressly authorized by Anthropic.” Without retention, requests fail with a 400 invalid_request_error. Organizations on zero data retention for Opus will not get Fable 5.2 by changing a string. Anthropic’s answer is Enterprise Frontier Safeguards, which stores the data in customer-controlled cloud storage and was rolling out in phases this fall. We covered the policy behind this when Anthropic first made 30-day logging mandatory; it now applies to the 5.2 generation as well.

Request parameters that now return errors

Forced tool use: tool_choice set to any or a named tool returns a 400. Agent code that forces a structured output through a tool call needs a different pattern.
Sampling parameters: omit temperature, top_p and top_k.
Thinking: adaptive thinking is always on and cannot be turned off. The guide warns that prompt lines asking for less thinking do not stop it at medium and above, and that it may start subagents at xhigh and max. Budget for that.
Computer use: the computer_20251124 tool version returns a 400 everywhere except Bedrock; use computer_toolset_20260801.

Capacity and speed

Fable 5.2 has its own rate limit, separate from the one Fable 5.1 and Fable 5 share, and it is not supported on Priority Tier. The optional fast mode gives up to 2.5 times the output speed at $16 and $80 per million tokens, double the standard rate, and does not work with the Batch API. Batch itself is a 50% discount at standard speed, which makes overnight Fable runs at $4 and $20 the same list price as interactive Opus 5.5.

Seats, plans and the November 16 admin date

For teams on Claude’s apps rather than the API, two dates matter. Under the launch offer, Pro and Team standard seats can use Fable models up to 50% of their weekly usage limits, plus one Fable limit reset, through November 16 at 11:59 pm PT. After that, ongoing plan rules move Pro, Team standard and Enterprise standard seats onto usage credits for Fable, while Max, Ultra and premium seats keep up to 50% of weekly limits for Fable at no extra cost.

The admin date is the one to put in a calendar. On Enterprise, Fable 5.2 stays off until an admin enables it, but seven days after launch it is added automatically unless an admin opts out. That is November 16. If your organization is not ready for a model that cannot run under zero data retention, or whose blocked requests switch models by default in the apps, the opt-out has to happen before then. Claude Code users also need version 2.1.321 or later to select it.

One more default to note: the launch page says Medium is “the default in Claude apps and Claude Code,” while the API defaults to high. The same prompt will spend differently in the two places.

A routing rule to start with

Anthropic’s own guidance, the benchmark grid and the integration constraints all point to the same split. A starting rule set we would test before committing any budget:

Keep Opus 5.5 as the default for bounded coding, document and analysis work. Anthropic recommends it, and on AA-Briefcase its Elo number is higher.
Send long-horizon, novel work to Fable 5.2 at high, not max: multi-day agent runs, research problems without a known solution path, large migrations. The FrontierCode comparison that justifies the price was at high.
Try Fable 5.2 at low or medium on hard reasoning tasks you currently run on Opus 5.5 at max. Anthropic’s HLE and Chartography figures suggest this is where the per-task saving is largest, if it holds on your data.
Use Batch for anything that can wait. Half price takes Fable to Opus 5.5’s interactive list price.
Read stop_reason and the model field on every response, and decide per workload whether fallback is on.
Re-run your evals on cost per completed task, not cost per token, with at least three attempts per item so a single lucky run does not decide the routing.

The release that makes Fable 5.2 interesting is not the 20% price cut. It is that Anthropic published effort-level comparisons that let a buyer price a frontier model per task, and then told most of its customers to keep using the cheaper one. Both statements can be true. The work is finding out which of your tasks sit on each side of that line, and our earlier guides on routing between Opus 5 and Fable 5 and on getting the most from Fable 5 still describe the method; the prices and defaults above are what changed.

When Is Claude Fable 5.2 Worth Twice the Price of Opus 5.5?