At the end of September, Anthropic launched Claude Opus 5.5 At the time, most tasks had already reached Fable 5.1’s level, but the price was much lower. In addition, Fable had weekly limits, so many people switched their main model to Opus 5.5 (its quota usage is really economical). After about two months, Anthropic officially upgraded both Claude Fable and Claude Mythos to version 5.2. This time, Fable 5.2 surpassed Opus 5.5 again in most official tests. More importantly, the Token price also dropped by 20%.
Anthropic also announced several rather striking real-world cases this time, such as having Fable 5.2 start from an empty folder and take 8 days to build an operating system that can boot on a laptop. It could even solve two unresolved problems in black hole physics. Fable 5.2 is now available in Claude, and to let users fully experience it, Anthropic has also sent out Fable reset vouchers.
Claude Fable 5.2 and Mythos 5.2 officially launch: outperforming Opus 5.5 and 20% cheaper, building an operating system in 8 days and solving two black hole physics problems.
Anthropic earlier officially announced Claude Fable 5.2 and Claude Mythos 5.2, replacing the 5.1 versions released in September. The two are the same underlying model, differing only in safety protections and who can access them: Fable 5.2 is available to all paid plans and API users, while Mythos 5.2 is only available to verified cybersecurity and life sciences organizations.
This release mainly enhances programming, reasoning, scientific research, and visual understanding. According to officials, Fable 5.2 and Mythos 5.2 are a tier above Opus 5.5 in almost all capabilities.
The key points regarding specifications and speed are as follows:
- Faster generation: output up to 30% faster than Fable 5.1 and Mythos 5.1.
- Using fewer tokens: With Medium reasoning effort set on difficult tasks, Fable 5.2 uses fewer tokens than Fable 5.1.
- Ultra-long context: a 1-million-token context window is provided by default, with a maximum output of 128K tokens in a single response—the same as Opus 5.5.
- Fast Mode: an optional fast mode that can boost token generation speed by up to 2.5x, but the price also doubles.
- Clearer communication: Anthropic noted that early testers generally found 5.2’s responses easier to understand than the previous Fable and Mythos.
Below is the Fable 5.2 test data published by Anthropic:
| Test item | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 (Programming) | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 (Software Development) | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench (Enterprise Workflow) | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science (Scientific Research) | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% |
| WANDR (Agent Search) | 82.1% | 68.7% | 72.3% | — | — |
| Humanity’s Last Exam (cross-disciplinary reasoning, tools available) | 74.2% | 65.6% | 67.7% | 57.2% | — |
| OSWorld 2.1 (Computer Operation) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim (Robot Control) | 64.9% | 27.6% | 39.4% | 29.8% | — |
| Chartography (chart recognition, without tools) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
| Remote Labor Index (remote knowledge work) | 41.3% | 17.9% | 21.3% | 20.8% | — |
| AA-Briefcase v1.1 (Knowledge Work Scoring) | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| AA-Briefcase v1.1 (knowledge-work Elo rating, not a percentage) | 1785 | 1669 | 1800 | 1569 | 1611 |
Based on the results of FrontierCode 2.0 tests for AI agents on coding, Fable 5.2 scored 49.1%, while the previous generation, Fable 5.1, scored only 32.4%—an improvement of nearly 17 percentage points. It also leads Opus 5.5 at 45.5% and OpenAI GPT-6 Astra at 42.8%. The gap is even more pronounced in the scientific research category, Terminal-Bench-Science: Fable 5.2 scored 74.6%, while Opus 5.5 scored only 58.7%:
The largest gap is in two relatively new test items, namely “Remote Labor Index” and “RoboDojo.”
In the Remote Labor Index, Fable 5.2 scored 41.3%, nearly double Opus 5.5’s 21.3%. RoboDojo lets AI control a simulated dual-arm robot using only camera footage; Fable 5.2 scored a very high 64.9%, Opus 5.5 scored 39.4%, and Fable 5.1 only 27.6%. Anthropic also specifically noted that Fable 5.2 made considerable progress in controlling physical robots, even on tasks for which it had not been specifically trained.
However, it doesn’t beat Opus 5.5 on every test.
On the AA-Briefcase knowledge work test’s rubric score, Fable 5.2 scored 66.1%, higher than Opus 5.5’s 61.2%, but when converted to Elo ranking, Opus 5.5’s 1800 is instead slightly higher than Fable 5.2’s 1785.
In addition, in the AutomationBench enterprise workflow test, Fable 5.2’s 51.4% clearly beat Opus 5.5 but lost to Google Gemini’s 53.9%. Chartography chart recognition is also close: Fable 5.2, using the highest reasoning intensity, scored 73.4%, only slightly higher than Gemini 4 Carbon’s 72.5% and GPT-6 Astra’s 71.0%.
Below are the API prices for Claude Fable 5.2 and other models. Although they are cheaper than Fable 5.1, they are still twice as high as Opus 5.5:
| Model | Input cost | Output cost |
|---|---|---|
| Fable 5.1 | $10 | $50 |
| Fable 5.2 | $8 | $40 |
| Fable 5.2（Fast Mode） | $16 | $80 |
| Mythos 5.2 | $8 | $40 |
| Opus 5.5 | $4 | $20 |
| GPT-6 Astra | $10 | $50 |
Besides benchmark scores and API pricing, Anthropic also shared several real-world case studies this time, which better illustrate the scale of work Fable 5.2 can handle.
Case 1: Building an operating system prototype in 8 days.
An engineer threw “designing an operating system prototype” as an experiment at Fable 5.2. Starting from an empty folder, Fable 5.2 coordinated more than 1,000 subagents working continuously for 8 days, and ultimately produced an operating system that can boot on a laptop, including a kernel, hard drive/keyboard/network drivers, file storage, windows, a login screen, an installer, and built-in Claude features. However, the official side did not disclose the number of lines of code or the cost.
Case 2: Solving two theoretical physics problems involving black holes.
Both problems involve quantum gravity, with results based on the theory of “holographic duality.” One of them completes a calculation first proposed as early as 2003 but never finished, proving that in a very small universe, the process of transitioning into a black hole happens suddenly; the other identifies the exact limits of applicability of a technique that has been used for more than 15 years to rule out candidate theories of quantum gravity. Anthropic said Fable 5.2 completed most of the work autonomously, and the relevant preprint papers have also been made public.
Claude Fable 5.2 is now available starting today across all paid plans, Claude Code, Claude API, and major cloud platforms including AWS, Google Cloud, and Microsoft Azure. The API model name is claude-fable-5-2. Fast Mode supports Claude Code, Claude Platform, Google Cloud, and Microsoft Foundry.
To mark the recent launch, Anthropic is also offering a limited-time promotion this week, from November 9 to November 16 (Pacific Time). Pro plans and standard Team seats can use Fable directly with the quota included in their plan, up to 50% of the weekly usage cap, plus 1 Fable quota reset. Max, Ultra, and premium Team seats will receive 1 to 3 Fable quota resets depending on subscription length.
Note: You must start a paid subscription before 10:00 a.m. Pacific Time on November 9, 2026, to be eligible for the reset voucher.
Source: KOCPC Chinese