Claude Fable 5.2 Explained: Anthropic’s New No. 1 AI Model — and the Price Cut That Isn’t

Checked against primary sources on November 11, 2026.

1. The short answer
2. One model, two doors: Fable vs Mythos
3. The benchmarks, in plain English
4. The 20% price cut that isn’t
5. Who should actually use it
6. The IPO backdrop
7. What to watch next
8. Frequently asked questions
9. Sources

The short answer

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, and within a day the new model took the top spot on the Artificial Analysis Intelligence Index — 65 points, five clear of Google’s Gemini 4 Carbon and six clear of Anthropic’s own Opus 5.5. The headline price fell 20% from the previous generation, to $8 per million input tokens and $40 per million output tokens. But the price cut is not the whole story: at maximum effort, Fable 5.2 uses so many tokens that a finished task costs about $26.50 on average — roughly 2.6 times Opus 5.5. The smart way to use it is at lower effort settings, where it beats Opus 5.5 on quality for less money. If you use Claude for ordinary coding and writing, Anthropic still recommends Opus 5.5 as the default; Fable 5.2 is for the hardest, most open-ended work.

One model, two doors: Fable vs Mythos

Fable 5.2 and Mythos 5.2 are the same underlying model. The difference is the safety boundary around it, and who is allowed through each door.

Claude Fable 5.2 is the version everyone can use. It carries strong safeguards that block most cybersecurity attack work and biology-related requests — penetration testing, exploit development, vulnerability analysis and similar tasks are refused. It is available now in the Claude apps, Claude Code, the Claude Platform, and through Amazon Web Services, Google Cloud and Microsoft Azure.

Claude Mythos 5.2 is the same model with some of those restrictions loosened. It is available only to organisations that pass Anthropic’s Cyber Verification Program or its Life Sciences Verification Program — vetted security firms and biomedical research organisations, not the general public.

The split is the same pattern Anthropic used for the 5.1 generation, but the safeguards are now less clumsy. Anthropic says Fable 5.2’s guardrails falsely block about 55% fewer legitimate requests than Fable 5.1 — about 60% fewer in cybersecurity and 40% fewer in biology — at the same level of protection. That matters because the most common complaint about the Fable line was not that it refused dangerous requests, but that it refused harmless ones that merely sounded technical.

Under the hood, the model has a 1-million-token context window and can produce up to 128,000 tokens of output in a single response. Anthropic says its outputs arrive up to 30% faster than Fable 5.1, and there is a new optional fast mode that speeds up token generation by up to 2.5 times — at double the price. The knowledge cutoff is June 2026.

The benchmarks, in plain English

Benchmarks are the part of a model launch most likely to mislead, so here is the honest version: the table below is Anthropic’s own scorecard, and Anthropic chose the tests and the settings. The one independent check available so far — Artificial Analysis, which runs its own nine-test battery — agrees with the headline, but the vendor numbers below have not been independently reproduced yet.

Test (what it measures)	Fable 5.2	Fable 5.1	Opus 5.5	GPT-6 Astra	Gemini 4 Carbon
FrontierCode 2.0 (agentic coding)	49.1%	32.4%	45.5%	42.8%	41.1%
FrontierSWE v2 (software engineering)	69.9%	56.3%	62.3%	65.5%	61.2%
AutomationBench (business workflows)	51.4%	31.4%	42.5%	41.4%	53.9%
Terminal-Bench-Science (agentic science)	74.6%	52.6%	58.7%	68.1%	64.8%
Humanity’s Last Exam, with tools	74.2%	65.6%	67.7%	57.2%	—
Remote Labor Index (real work tasks)	41.3%	17.9%	21.3%	20.8%	—

Three things stand out. First, the jumps over Fable 5.1 are large: FrontierCode up 17 points, Terminal-Bench-Science up 22, the Remote Labor Index more than doubled. Second, Fable 5.2 leads in almost every row — but not every one. Gemini 4 Carbon still wins the business-automation test, and on Artificial Analysis’s human-to-human knowledge-work ranking, Opus 5.5 stays ahead on Elo (1,800 vs 1,785). Third, Mythos 5.2’s specialist numbers are striking in its own lane: 76.3% on CyScenarioBench and 89.6% on CVE Rediscovery, and in Anthropic’s October test run of 1,829 known vulnerabilities, Mythos 5.2 found 10.2% more than Opus 5.5 and 14.4% more than Mythos 5.1.

The independent check: Artificial Analysis put Fable 5.2 through its Intelligence Index v5.0.1 at all five effort settings. At maximum effort it scores 65 points — first of the 117 models on the board, ahead of Carbon (60), Opus 5.5 (59) and Fable 5.1 (50). At its High and Xhigh settings it scores 62 each, which would still be first overall. That is a stronger result than the vendor table alone, because Artificial Analysis runs the same tests on every model and publishes its methods.

Anthropic also published a set of internal experiments that are impressive but should be read as demonstrations, not benchmarks: a team of more than 1,000 Fable 5.2 sub-agents built a prototype operating system that boots on a laptop in eight days; another 1,000-agent team formalised the proof of Fermat’s Last Theorem in 30 hours, where the first attempt with an earlier Claude took 11 days and Opus 5.5 managed only 26% of the same job in 18 hours; and the model solved two open problems in black-hole physics and designed three working electronic devices from a one-paragraph prompt. These show what the model can do with enormous compute and careful scaffolding — not what it will do for you on a Tuesday afternoon.

The 20% price cut that isn’t

The list price fell: $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. That is a real 20% cut in the per-token rate. But you do not buy tokens — you buy finished tasks, and Fable 5.2’s best work uses far more tokens to get there.

Worked example 1: one long coding session. Suppose a session uses 200,000 input tokens and 50,000 output tokens. On Fable 5.2 that costs 200,000 × $8 ÷ 1,000,000 = $1.60 for input, plus 50,000 × $40 ÷ 1,000,000 = $2.00 for output — $3.60 in total. The same session on Fable 5.1 cost $4.50, so the 20% saving is real. But on Opus 5.5, priced at $4 and $20 per million, the same session costs $0.80 + $1.00 = $1.80 — exactly half. Fable 5.2 is still twice the price of Opus per token. In fast mode, the session costs $7.20 and finishes up to 2.5 times faster.

Worked example 2: the effort ladder. Artificial Analysis publishes the average cost of one index task at each effort setting:

Fable 5.2 effort	Index score	Average cost per task
Low	54	$3.10
High	62	$9.08
Xhigh	62	$14.83
Max	65	$26.54

Opus 5.5 at its maximum scores 59 at about $10 a task. So Fable 5.2 at High effort beats Opus 5.5 at maximum on both score (62 vs 59) and cost ($9.08 vs ~$10) — that is the sweet spot, and it is the number Anthropic itself highlights: on FrontierCode, Fable 5.2 at High effort roughly matches Opus 5.5 at Max for about half the cost per task. But the last three index points, from 62 to 65, cost an extra $17.46 per task — nearly triple the High-effort price. Maximum effort is for work where the answer is worth $26, not for everyday jobs. Our cost-per-finished-task comparison explains the same trap across all four frontier models: per-token prices tell you almost nothing about what a job costs.

Worked example 3: a monthly budget. A developer running 20 sessions like example 1 each month would spend about $72 on Fable 5.2, $36 on Opus 5.5 and $90 on Fable 5.1 at list prices — before prompt caching, which cuts input costs sharply on repeated work, and before any plan allowance. The practical rule: route ordinary work to Opus, and reserve Fable for the jobs where Opus gets stuck.

Who should actually use it

Anthropic’s own guidance is unusually blunt: for most coding and knowledge tasks, start with Opus 5.5, and move up to Fable or Mythos — or down to Sonnet 5.5 or Haiku 5.5 — depending on the scope and complexity of the work. Fable 5.2 can also act as an advisor that smaller models consult, or as an orchestrator that delegates simpler tasks to cheaper models.

For regular Claude users, the change is quieter than the launch headlines suggest. Medium effort is the default in the Claude apps and Claude Code, and at Medium effort Fable 5.2 does challenging tasks with fewer tokens than Fable 5.1 did — so the same plan allowance should stretch further on hard jobs, even though the model is more capable. What changes in practice: fewer false refusals on technical questions, faster responses, and better results on long, multi-step work such as migrating a codebase or researching a dense topic. If your work is short emails, summaries and everyday questions, you will notice almost nothing — and that is fine, because Fable is not built for that. Our Mistral Large 4 licence explainer covers the other open-weights option if you need a model you can run yourself.

The IPO backdrop

The launch landed one day before Anthropic started its IPO roadshow on November 10, 2026, with a reported price range of $70 to $78 a share — a detail our roadshow article covers in full, alongside the filing numbers from the S-1. The timing is not a coincidence: a new No. 1 ranking the week you ask investors to value the company is the best marketing a model lab can buy, and it is free. It also raises the stakes. If Fable 5.2’s lead holds up under independent testing, it supports the growth story in the filing; if rivals close the gap quickly — Google’s Gemini 4 Carbon is five points behind and still improving — the ranking is a snapshot, not a moat. None of this is investment advice, and Anthropic has not confirmed a listing date.

The launch also lands in a week when the whole industry is repricing itself. OpenAI cut its fundraise from $30 billion to $20 billion on what investors called tepid demand, and Anthropic’s own valuation has been marked down from about $2 trillion to the $1.2–$1.6 trillion range in the same conversation. Against that background, a model launch is also a message to investors: the product pipeline is still moving.

What to watch next

First, independent reproduction. Anthropic’s benchmark table is self-reported, and the history of model launches says the first independent tests — from Artificial Analysis, METR and others — will qualify the headline numbers within weeks. Watch whether Fable 5.2 keeps its lead at the lower effort settings, where the money is.

Second, the response. Google has not answered Fable 5.2 yet, and its Gemini 4 Carbon is the closest rival on the index. A Carbon update or a price move would change the maths in the tables above, and we will update this article if one lands.

Third, fast mode in practice. A 2.5-times speedup at double the price is a good deal only if your bottleneck is waiting, not money. Early reports from developers on whether fast mode is worth it for real coding sessions will be more useful than any benchmark.

Frequently asked questions

Sources

• Anthropic, “Introducing Claude Fable 5.2 and Claude Mythos 5.2” (November 9, 2026) — https://www.anthropic.com/claude-fable-and-mythos-5-2
• Artificial Analysis, Intelligence Index / model leaderboard — https://artificialanalysis.ai/models
• Anthropic, API pricing — https://www.anthropic.com/pricing

Figures checked against primary sources on November 11, 2026. Anthropic’s benchmark scores are vendor-reported unless stated; Artificial Analysis scores are independently measured.