Fable 5.2 is two model names, one model, and a price cut dressed as restraint

Anthropic published Claude Fable 5.2 and Mythos 5.2 on 9 November, and the two names describe a single model. Base pricing is $8 per million input tokens and $40 per million output, which CNBC reports is 20% below the 5.1 generation. The same reporting says the company is preparing an IPO that could price as soon as next week, at a $1 trillion to $1.5 trillion valuation rather than the $2 trillion floated earlier. Dianne Penn, Anthropic's head of product management for research, framed the release to CNBC as "we called for pacing, not pausing", picking up the argument chief executive Dario Amodei made in September, when he asked the industry to slow down. A 20% discount on the frontier tier is not a pause. It is a price cut with a philosophical footnote, and the fine print is the more interesting document.

One checkpoint, two doors

Fable and Mythos are the same weights under different admission rules. Fable carries stronger cybersecurity and biology safeguards and is open to all users. Mythos is limited to organizations verified through Anthropic's Cyber Verification Program or Life Sciences Verification Program, per the Fable docs and the separate Mythos overview. The lineage explains the split: Mythos appeared in April 2026 for select organizations, and Fable arrived about two months later with extra guardrails attached. One release, two sets of doors, one benchmark table.

The 30% claim meets a stopwatch

Anthropic says Fable 5.2 produces output "up to 30% faster" than Fable 5.1. Independent measurement of the same model came back with a different headline. Artificial Analysis ranks Fable 5.2 first of 74 models on its Intelligence Index, at a score of 65, so the capability claim holds up under outside review. Then the timing columns: 44.4 minutes to complete one Intelligence Index task, 71st of 72 models measured, on output of 103 tokens per second and a time to first answer token of 235.06 seconds. The token count explains part of it. The model emitted 275 million output tokens across the Index against a median of 72.8 million, which the firm calls "very verbose".

So the percentage is real and nearly irrelevant to the experience: it compares this model to its own predecessor. Thirty per cent faster than a slow model is still slow against the field, and always-on adaptive thinking is much of the bill. Cost lands in the same place: $26.54 per Intelligence Index task, 73rd of 73 on that column, at list prices of $8 and $40 against market medians of $2 and $10. The fast mode is the intended escape hatch, lifting generation up to 2.5 times for $16 input and $80 output per million tokens, at which point the discount is gone twice over.

Where the "step above" stops

The announcement puts Fable 5.2 a step above Claude Opus 5.5 in almost all evaluations, and prints the table where that phrase meets its exceptions.

Benchmark	Fable 5.2	Fable 5.1	Opus 5.5	Gemini 4 Carbon
FrontierCode 2.0	49.1	32.4	45.5	41.1
FrontierSWE v2	69.9	56.3	62.3	61.2
AutomationBench	51.4	31.4	42.5	53.9
AA-Briefcase v1.1	66.1	61.5	61.2	61.8
Chartography, no tools	73.4	44.8	64.4	72.5
Humanity's Last Exam, tools	74.2	65.6	67.7	57.2

Two rows matter. On AutomationBench, Gemini 4 Carbon scores 53.9 against Fable's 51.4, so the newest Anthropic model is not the best model on agentic automation, and the vendor's own table says so. On AA-Briefcase the raw score favours Fable, 66.1 against 61.2, while the Elo conversion puts Opus 5.5 ahead at 1800 to 1785, a reminder that a fixed-task percentage and a preference rating measure different things. Anthropic's efficiency argument is more persuasive than the speed one: at High effort Fable 5.2 reaches 45.3 on FrontierCode, about level with Opus 5.5 at Max effort and its 45.5, for roughly half the cost per task. At Xhigh it goes to 49.1, the highest Anthropic has recorded for a released model.

Cybersecurity is where the restricted version earns its name. Mythos 5.2 scores 76.3 on CyScenarioBench against 61.7 for Mythos 5.1 and 67.6 for Opus 5.5 "without safeguards", but the announcement notes each lab ran a unique task set, so the column is a within-family comparison, not a cross-vendor one. RoboDojo Sim, the footnote adds, was run internally.

A prospectus with a discount attached

The business frame is harder to read than the benchmarks. Amodei argued in September for slowing the pace of development; two months later his company cut the price of its frontier model during an IPO run-up. CNBC's amended prospectus figures give the pressure a shape: $33.24 billion of revenue for the nine months ended 30 September, up roughly 1,200% year on year, against a valuation target that has already been trimmed from $2 trillion. Penn told CNBC that Fable 5.2 is aimed at "power users or power use cases" and that most daily work should start on Opus 5.5. That is a company asking customers to pay frontier prices only where the work justifies it, sensible product advice and a hedge against the cost column.

The spec sheet and its conditions

The model documentation lists a 1M-token context window, 128K maximum output, text and image input with text output, a June 2026 knowledge cutoff, and a retirement date no sooner than 9 November 2027. Two conditions deserve more attention than usual. Adaptive thinking is always on and cannot be turned off, with default effort set to high. And the model requires 30-day data retention, so it is unavailable under zero data retention unless a customer is specifically authorized. Buyers with hard privacy constraints therefore have a simpler answer than the table implies.

Safety claims follow the pattern of the system card: new evaluations for sandbox escapes, what the company calls its strongest-ever protections against distillation attacks, and cyber and biology safeguards described as robust as Fable 5.1's while producing fewer unnecessary blocks. Prerelease anecdotes include software migrations, three electronic devices designed from scratch and two open problems in black-hole physics, all unverifiable from outside.

What the launch actually offers is a modest capability gain, a real price cut, and a distribution strategy that splits one model into a public door and a verified one. Against the 5.1 generation in September, the honest summary is that the gains are concentrated in coding and science benchmarks, the automation and preference columns have holes the announcement prints itself, and per-task cost remains the worst on the independent index. Pacing or not, the frontier got cheaper on 9 November. It did not get faster in any sense a stopwatch would confirm.