Tech • AI • Robotics • Game
Anthropic Just Dropped Fable 5.2 (Full Breakdown)
9.4/10TL;DR
Anthropic has launched Claude Fable 5.2 with lower pricing, faster responses, and stronger benchmark results than Opus 5.5 in several advanced tasks, while still positioning Opus 5.5 as the default model for most everyday work.
KEY POINTS
Launch and model position
Fable 5.2 is presented as Anthropic’s most capable generally available model, above Opus, Sonnet, and Haiku in its product lineup. A related Mythos 5.2 variant was also released, but it is restricted to vetted organizations and includes tighter safeguards around areas such as cybersecurity and biology.
Price cut and speed gains
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the prior version, a reduction of about 20%. Anthropic also says answers are produced up to 30% faster than Fable 5.1, and a new fast mode in Claude Code can run up to 2.5 times faster, though at double the cost.
Still pricier than Opus
Even after the price reduction, Fable 5.2 remains roughly twice the price of Opus 5.5 on token rates. Anthropic’s argument is that token pricing alone does not capture total task cost: on an internal coding test, Fable 5.2 at high effort reportedly matched Opus 5.5 Max at about half the cost per task. A new medium effort setting is now the default in Claude apps and Claude Code, with claims that it uses fewer tokens on hard tasks than Fable 5.1.
Benchmark lead in coding and knowledge work
On agentic coding, a benchmark measuring whether generated code is acceptable to developers, Opus 5.5 scored 45.5% while Fable 5.2 reached 49.1%, which Anthropic describes as its best result yet for a released model. The largest practical jump for nontechnical users appeared in knowledge work, where Opus 5.5 scored 21.3% and Fable 5.2 scored 41.3%, nearly doubling the result.
Smaller edge in computer use
In computer-use tasks, which evaluate how well a model can operate software interfaces and navigate a machine, Fable 5.2 posted 89.3% versus 87.2% for Opus 5.5. The narrow gap suggests only a modest improvement in one of the most closely watched categories for agent-style automation.
Major jump in robotics
The biggest increase came in robotics. Fable 5.1 scored 27.6%, Opus 5.5 scored 39.4%, and Fable 5.2 reached 64.9%, outperforming OpenAI’s GPT-6 Astra on that chart. Anthropic showcased robots using the model to solve a puzzle cube, unscrew a bottle cap, and draw a robot, and said the system generalized to robotics tasks it had not specifically been trained on.
Not a clean sweep
The new model did not lead every category. In business workflow evaluations, Google Gemini 4 Carbon scored slightly higher, indicating that Fable 5.2 is not dominant across all enterprise use cases despite its broader gains.
Long-run task example and early tester feedback
In one internal test, an engineer reportedly gave the model an empty folder and asked for a prototype operating system. The system then ran more than 1,000 helper agents continuously for eight days and produced a working operating system that could boot on a laptop and included a login screen and installer. Early testers also highlighted stronger intent understanding and a greater willingness to admit uncertainty rather than hallucinate.
Looser safety filtering and subscriber perk
Anthropic says Fable 5.2 is blocked by safety filters about 55% less than Fable 5.1, a potentially important change for users frustrated by overly strict refusals. For paid users, the company is temporarily adding extra Fable access: Pro subscribers can use the model within their plan for up to 50% of weekly usage limits through November 16, with one free reset, while Max and Ultra subscribers receive resets based on how long they have been subscribed.
CONCLUSION
Claude Fable 5.2 brings measurable gains in advanced reasoning, coding, and robotics while becoming cheaper and faster than its predecessor. But with Opus 5.5 still framed as the stronger daily default for most coding and knowledge tasks, the upgrade case appears strongest for users tackling harder, high-value workloads.
Ask a question