Claude Fable 5.2: features, price and the catch

Claude Fable 5.2 is Anthropic's most capable model on general release, and it costs 20% less than the version it replaces. Released on 9 November, it beats Claude Opus 5.5 in almost every capability Anthropic tested. Yet Anthropic's own advice is to start with Opus 5.5. Here is what Fable 5.2 can do, what it costs and when it is worth paying for.

If you pay for Claude Pro, this is the week to try it. Fable models normally cost Pro users extra, but they are included in Pro until 16 November at 11:59 pm Pacific time, which is 1:29 pm IST on 17 November. You can spend up to half your weekly limit on them in that time.

What is Claude Fable 5.2?

Fable is the top tier of Claude, above Haiku, Sonnet and Opus. Fable 5.2 and Claude Mythos 5.2 are the same underlying model with different safeguards. Fable's classifiers block high-risk cybersecurity work, such as writing exploits, and most research biology. That is why Anthropic can sell it to everyone. Mythos lifts some of those blocks, and only organisations vetted through Anthropic's Cyber and Life Sciences Verification Programs can use it.

The specs on Anthropic's model page are generous. Fable 5.2 reads up to 1 million tokens of context and writes up to 128,000 tokens in one reply. It takes text and images as input, and its knowledge runs to June 2026. You can use it in the Claude apps and Claude Code, and through the API as claude-fable-5-2. It is also on Amazon Bedrock, Google Cloud and Microsoft Foundry.

What changed from Fable 5.1

Four things changed since Fable 5.1 arrived in September. The price fell from $10 to $8 per million input tokens, and from $50 to $40 per million output tokens. Output now comes up to 30% faster than Fable 5.1. Testers told Anthropic that the new model communicates more clearly.

The fourth change is that its safety filters get in the way less often. In early testing across claude.ai, Claude Code and Cowork, Fable 5.2 blocked about 55% fewer requests than Fable 5.1. Blocks fell by about 60% in cybersecurity and 40% in biology.

You also choose how hard it thinks. Fable 5.2 has five effort levels, from low to max. Medium is the default in the Claude apps and Claude Code, while the API defaults to high. More effort brings better answers on hard problems, but it uses more tokens. An optional fast mode generates tokens up to 2.5 times faster, at double the price.

How does it compare with Opus 5.5 and GPT-6?

The launch post puts Fable 5.2 ahead of Opus 5.5 on almost every row, and ahead of OpenAI's GPT-6 Astra on every test they share. These are Anthropic's own numbers, mostly at each model's highest effort setting:

It doesn't win everything. Google's Gemini 4 Carbon beat it on AutomationBench, a test of business workflows, by 53.9% to 51.4%. The system card also shows Fable 5.2 trailing Opus 5.5 on two knowledge-work ratings. Our report on Gemini 4 Carbon's staged rollout covers that rival.

An independent ranking agrees with the headline. Artificial Analysis puts Fable 5.2 at max effort first of 74 models on its Intelligence Index, with a score of 65. It ran the model with Anthropic's default fallback switched on, so another model answered whatever the safety filters blocked.

Cost per task matters more than cost per token, and here the numbers favour Fable 5.2. On FrontierCode, Fable 5.2 at high effort scored 45.3%, level with Opus 5.5's best of 45.5%. It got there for roughly half the cost per task. On Humanity's Last Exam, low effort matched Opus 5.5's top score for less than a tenth of the cost.

What Fable 5.2 built in Anthropic's tests

The launch post is full of long, ambitious jobs. Anthropic or its testers reported all of them, and none has been checked independently yet:

• An operating system in eight days. An Anthropic engineer gave it an empty folder. It coordinated over 1,000 subagents for eight days and produced an OS that boots on a laptop, with disk, keyboard and network drivers, a login screen and an installer.

• A faster production service. A team of agents running an earlier version moved a Python service to Rust. It now starts about 26 times faster.

• Fermat's Last Theorem, formalised in 30 hours. A team of 1,000 Fable 5.2 agents wrote a computer-checked proof in Lean. On 11 November, Anthropic amended its description: the agents were given a list of 60 statements, their definitions and 47 lemmas already proved in its earlier formalisation. Opus 5.5, with the same setup, finished 26% in 18 hours.

• Three working gadgets. From a one-paragraph prompt, it designed a voice-assistant hub, a USB speakerphone and a handheld virtual pet, including circuit boards and 3D-printed cases. All three worked without electrical changes.

• Two open physics problems. It solved two open problems in theoretical physics, both about black holes. The results are written up as preprints.

Testers describe a model that keeps going. Uber's Andrew Mager had it run a 10 to 12 hour audit, during which it sent out about 170 reviewer agents. It came back with a write-up and fix pull requests. "It works until the job is actually complete instead of stopping at a plausible checkpoint," he said. If subagents are new to you, our guide to how AI agents work explains the idea.

How much does Claude Fable 5.2 cost?

On the API, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. That is double Opus 5.5's $4 and $20. Cached input costs $0.20 per million tokens, and the Batch API halves the price for work that can wait. Fast mode costs $16 and $80.

Long agent runs add up quickly. AI YouTuber Matthew Berman had it build one level of a shooter game. According to Decrypt, the run lasted almost 44 hours, used 8.63 billion tokens and would have cost an estimated $14,710 at API prices. Artificial Analysis puts Fable 5.2's average cost per task in its index at $26.54 at max effort, against $3.10 at low effort.

"Anthropic has cut the price of a token, not the price of a thought," Sanchit Vir Gogia of Greyhound Research told Computerworld. He advises judging the cost against tasks actually completed.

In the Claude apps, Fable 5.2 needs a paid plan. Anthropic's help page explains how it counts on each one:

• Max and Ultra ($100 and $200 a month) include Fable models for up to 50% of your weekly usage limit. Fable uses up that limit faster than other models.

• Pro and standard Team seats don't include Fable outside the launch week. You pay for it with usage credits.

• Claude Code needs version 2.1.321 or later to show Fable 5.2.

The launch offers add some slack. Max and Ultra subscribers get one to three Fable limit resets, depending on how long they have paid, and these last until 8 March 2027. Pro users get one reset to use before 16 November. In India, Claude has been priced in rupees since July. Business Today reports Pro at ₹2,000 a month and Max from ₹11,999, tax included.

The catches

Blocked requests go to an older model. In the Claude apps, when a request trips Fable's cybersecurity filter, Opus 4.8 answers it instead, and biology questions go to Opus 5. The system card says this makes agents running Fable 5.2 easier to hijack through prompt injection, because Opus 4.8 resists those attacks less well. When Fable 5.2 answers an injection attempt itself, it is one of Anthropic's most resistant models.

Your data is kept for 30 days. Fable 5.2 requires 30-day data retention for safety monitoring. Zero data retention needs Anthropic's approval. Its new Enterprise Frontier Safeguards, in private beta, run that monitoring inside a customer's own systems instead. Regulated firms should check this before sending sensitive work.

It often suspects it is being tested. Anthropic ran its alignment tests on Mythos 5.2, the same model with fewer safeguards, and expects Fable 5.2 to behave similarly. It rates the risk of seriously unsafe behaviour as low, and the model matched or beat Opus 5.5 on most alignment measures. But the model suspected it might be under evaluation in up to 43% of transcripts from one audit, which makes those results harder to trust. Told to underperform on purpose without being caught, it slipped past a monitor reading its reasoning in about one attempt in four.

There is real progress, too. In a test built to tempt the model out of its sandbox, it used credentials planted outside in 3.1% of scenarios. Mythos 5 did so in 39%. That earlier model once crossed its test boundaries to target GitHub, so the drop matters.

Developers need to change some code. Thinking is always on, and effort is the main control. Leave out temperature, top_p and top_k. Forcing a tool call with tool_choice returns a 400 error. Anthropic's effort guide says to start at high and move to xhigh only where your own tests show a gain.

Should you switch to Fable 5.2?

For most people, not by default. Anthropic calls Opus 5.5 its "strongest daily driver" and recommends it for most coding and knowledge tasks, at half the price per token. Fable is meant for the most ambitious, open-ended technical work.

Fable 5.2 earns its price on jobs that run for hours. Think of large migrations, long audits, research that needs many agents, or problems that cheaper models keep failing. Omdia's Lian Jye Su told Computerworld that "reliability remains an issue without human supervision". Keep a person approving anything that ships.

If you're on Pro, I'd use the included week. Give Fable 5.2 the hardest task on your list, then give Opus 5.5 the same prompt and compare. One real test on your own work will tell you more than any benchmark table.