Claude Fable 5.2: what changes for companies

On November 9, 2026, Anthropic released Claude Fable 5.2, the new version of its most advanced model. It comes with a lower price, faster responses and improvements in coding, reasoning and vision.

For companies that use artificial intelligence, the news is more than a version number: it changes the cost of the most complex work you can hand to a model, and how long an agent can carry it forward on its own.

01What was released

Alongside Fable 5.2, Anthropic introduced Mythos 5.2. According to the company they are the same underlying model, with one difference: Fable 5.2 includes safeguards against misuse in cybersecurity and biology, which is why it is available to everyone; Mythos 5.2 is reserved for verified organizations working in those fields.

Fable 5.2 can be used on Claude's Pro, Max, Ultra, Team and Enterprise plans, on the developer platform with the identifier claude-fable-5-2, and on the Amazon, Google and Microsoft clouds.

02Faster and less expensive

per million input tokens

$8

per million output tokens

$40

faster than Fable 5.1, at most

30%

the speed with fast mode, at most

2.5×

The price drops from $10 and $50 to $8 and $40 per million tokens. A fast mode is also available, at $16 and $80, which according to Anthropic speeds up generation by up to 2.5 times.

Anthropic adds that, at the default effort level, Fable 5.2 completes difficult tasks using fewer tokens than its predecessor: the real saving can therefore be greater than the list-price cut alone.

03What it does better

Anthropic points to improvements in coding, reasoning and vision, and reports that those who tried it in preview find it clearer in its communication than previous versions.

The most significant change concerns long-running work. According to the company, Fable 5.2 can carry forward engineering work that is not fully defined for days and coordinate hundreds of sub-agents. Among the examples cited is the migration of a production service from Python to Rust, completed by a group of agents: the resulting service starts about 26 times faster.

FrontierCode 2.0, agentic coding

49.1%

FrontierSWE v2, software engineering

69.9%

AutomationBench, business workflows

51.4%

OSWorld 2.1, computer use

89.3%

These are results stated by Anthropic. On FrontierCode 2.0 the previous model, Fable 5.1, stopped at 32.4% and Opus 5.5 at 45.5%. Fable 5.2 is not first in every test, however: on AutomationBench a competing model scores 53.9%.

04Safety and data

Anthropic states that Fable 5.2's safeguards in cybersecurity and biology are as robust as those of Fable 5.1, but are triggered less often when they shouldn't be: in early tests, blocks were about 55% fewer. In terms of behavior, the company rates the risk of unsafe actions in critical situations as low.

There is one point companies need to know before adopting it: for Fable models, Anthropic applies 30-day data retention by default for safety monitoring purposes. A program that makes the model compatible with zero data retention is in private preview. Anyone handling sensitive data must take this into account when choosing.

05Which model to choose

Anthropic's guidance is clear: for most coding and office-work tasks you start with Opus 5.5; you move up to Fable for the most open-ended and ambitious technical work, and down to Sonnet 5.5 or Haiku 5.5 for simpler tasks.

Fable 5.2 can also work together with smaller models, in two ways:

• Advisor: a smaller model does the work and consults Fable on the difficult steps.
• Coordinator: Fable breaks the work down and assigns the simple parts to faster, cheaper models.

06What it means for companies

From our point of view, there are three practical consequences.

1. 01Complex work costs less. Some use cases that were set aside because they were too expensive to run deserve a fresh calculation.
2. 02More autonomy requires more control. An agent that works for hours or days needs tracked phases, independent checks and human approvals at the steps that matter.
3. 03The model is not a choice for life. A better one comes out every few months: it pays to have an architecture in which changing it does not mean redoing everything.

The best model matters less than how you put it to work: on which processes, with which controls.

This is the approach we took with Sinapse Agent, our AI agent platform: the system chooses the most suitable models, task by task, and nothing is written without a person's approval.

Source: Anthropic's announcement of November 9, 2026. Prices and results are those stated by the company at the time of release.

Cover image: Anthropic.