Claude Fable 5.2 vs GPT-6 Astra and Gemini 4 Carbon
Anthropic’s benchmarks put Fable 5.2 ahead of GPT-6 Astra in most tests. Here’s where it falls short, how API prices compare, and what the results leave out.
👁 114 reads
In September, I wrote that the world’s two most expensive models had identical API prices: Claude Fable 5.1 and GPT-6 Astra both cost $10 per million input tokens and $50 per million output tokens. That tie ended on November 9, 2026. Anthropic released Claude Fable 5.2 at $8 and $40, then placed it alongside OpenAI’s GPT-6 Astra and Google’s Gemini 4 Carbon in its announcement table. Fable leads in almost every row.
Almost: Google takes one row, and another is effectively a tie within the margin of error. There’s also a caveat worth putting up front: these are Anthropic’s figures. The company chose the tests and described its rivals’ results. It commissioned some measurements from others: Cognition ran FrontierCode 2.0, with each model using its own tool (Claude in Claude Code, GPT-6 Astra in Codex CLI, and Gemini 4 Carbon in Antigravity CLI), as described in the system card on page 210. I went through the results row by row and compared them with prices from OpenAI and Google’s sources.
Anthropic’s benchmark table
Results are at each model’s highest available effort level unless otherwise noted. Fable’s results include safeguards. A dash means Anthropic did not report a result.
| Test | Fable 5.2 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|
| FrontierCode 2.0 (agentic coding) | 49.1% | 42.8% | 41.1% (Medium) |
| FrontierSWE v2 | 69.9% | 65.5% | 61.2% |
| AutomationBench (business workflows) | 51.4% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 68.1% | 64.8% |
| OSWorld 2.1 (computer use) | 89.3% | 79.5% | 81.5% |
| RoboDojo Sim (robotics) | 64.9% | 29.8% | - |
| AA-Briefcase v1.1 (office work) | 66.1% | 52.0% | 61.8% |
| Remote Labor Index | 41.3% | 20.8% | - |
| Humanity’s Last Exam with tools | 74.2% | 57.2% | - |
| Chartography (charts, no tools) | 73.4% | 71.0% | 72.5% |
Source: Anthropic’s announcement, November 9, 2026.
Where Fable leads - and where it doesn’t
The biggest gaps appear in tasks that take a long time and require planning. In the Remote Labor Index, which uses real freelance-market projects, Fable 5.2 completes twice as many projects at a human level as GPT-6 Astra (41.3% versus 20.8%). The gap is even larger in the robotics simulation: 64.9% versus 29.8%. On Humanity’s Last Exam, a set of expert-level questions, Fable leads by 17 percentage points. These aren’t differences likely to disappear just by changing settings.
The gap is smaller in coding. Fable scores 49.1% on FrontierCode 2.0, but only at the highest effort level, Xhigh. GPT-6 Astra scores 42.8%. Gemini 4 Carbon’s score is at Medium effort, making this the least fair comparison in the table. It’s unclear what Carbon would score at maximum effort.
Two rows need a closer look. Gemini 4 Carbon leads on AutomationBench, a business-process automation test (53.9% versus 51.4%). On Chartography, which tests chart reading, Fable scores 73.4%, Carbon 72.5%, and Astra 71.0%. A gap of less than one percentage point is effectively a tie. What’s more, the efficiency chart in Anthropic’s announcement shows Fable scoring 66.4% on Medium effort - well below both rivals at their maximum settings. On charts and tables, Fable only pulls ahead when it’s allowed to think for longer.
My own test
In my test, Fable 5.2’s competitors were GPT-6.1 Sol and Gemini 4 Argon, not GPT-6 Astra and Gemini 4 Carbon from the manufacturers’ table. Across 27 tasks, all the models, including Opus 5.5, scored 100% on coding, puzzles, and image questions - a tie. Differences emerged in cost and Polish-language text: Fable cost $0.75, Argon $0.14, and Sol $0.11. In the Polish-language texts, the judge ranked Fable third on average out of four, behind Sol (1.33) and ahead of Argon (3.83). The test was conducted in Polish; these results are not an English-language evaluation. GPT-6.1 Sol was the judge and also evaluated itself, so I’d treat the ranking cautiously.
API prices: Fable is no longer the most expensive
| Model (API, per 1 million tokens) | Input | Output |
|---|---|---|
| Claude Fable 5.2 | $8 | $40 |
| GPT-6 Astra | $10 | $50 |
| Gemini 4 Carbon (introductory price) | $2 | $10 |
| Gemini 4 Carbon (after introductory period) | $4 | $20 |
Fable’s prices come from Anthropic’s pricing page, GPT-6 Astra’s from its model page in OpenAI’s documentation, and Gemini 4 Carbon’s from Google’s announcement on October 28, 2026. I checked all three on November 9. GPT-6 Astra also charges more for long prompts: above 272,000 input tokens, the entire prompt costs twice as much on input and 1.5 times as much on output. Fable 5.2 charges the standard rate for its entire 1-million-token context. That can make a real difference when working with large repositories.
Gemini 4 Carbon is the cheapest of the three, but there’s a catch: it’s not yet broadly available. Google is making it available first to selected cybersecurity defenders, then to trusted partners through the Fairwind Program. A wider release for developers and businesses is described as “coming soon.” I covered this in my article on Gemini 4 Carbon. For anyone choosing a model today, the practical options in this table are Fable 5.2 and GPT-6 Astra. Fable is 20% cheaper and, according to Anthropic’s data, scores higher in all 10 rows where both models have results. Against Gemini 4 Carbon, it leads in six of the seven rows where both have results.
Which models are missing from the comparison?
Anthropic compared Fable only with GPT-6 Astra and Gemini 4 Carbon. The table leaves out cheaper rival models: OpenAI’s GPT-6.1 Sol and Google’s Gemini 4 Argon. Both are available through OpenRouter for $2 and $10 per million tokens, respectively (model list checked on November 9), making them four times cheaper than Fable. For many use cases, they’re the real competition.
One comparison with GPT-6.1 Sol is tucked away in the system card, in the description of Terminal-Bench 4.0, which has 66 terminal tasks covering computational biology, physics simulations, CAD, and formal proofs. Fable 5.2 scores 67.6% with safeguards on, Opus 5.5 scores 66.4%, and Mythos 5.2 without safeguards scores 70.5%. The public leaderboard gives GPT-6.1 Sol a score of 58.2%, and GPT-6 Astra 58.2% (OpenAI reported 57.9% itself). Anthropic notes that OpenAI did not publish its own Sol results for this test (system card, page 211). A 10-point lead over Sol on terminal tasks is substantial, but Sol costs four times less.
Cybersecurity is a separate case. Here, Anthropic compares Mythos 5.2, the version without safeguards, with GPT-6 Cyber: 76.3% versus 61% on CyScenarioBench. But a footnote in the announcement says each lab gets different tasks in this test, so the figures are difficult to treat as a direct comparison. Read more in the Claude Mythos 5.2 hub.
My takeaway: according to Anthropic, Fable 5.2 clearly pulls ahead of GPT-6 Astra on long agentic tasks and multi-step work, while costing less. It has no clear lead in business automation or chart reading. For a full guide to the model, see the Claude Fable 5.2 hub.
Frequently asked questions
Is Claude Fable 5.2 better than GPT-6 Astra?
According to Anthropic’s table, yes: Fable scores higher in all 10 tests where both models have results. Its lead on Chartography is only 2.4 points, though. These are the model maker’s figures, not an independent measurement.
Which model has cheaper API pricing: Fable 5.2 or GPT-6 Astra?
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, compared with $10 and $50 for GPT-6 Astra. Astra also charges more for prompts exceeding 272,000 input tokens.
Can I use Gemini 4 Carbon?
It’s not broadly available yet. Google is providing access to selected cybersecurity specialists and Fairwind Program partners, and has announced a wider release without giving a date. Check with Google for current availability.
Why aren’t GPT-6.1 Sol and Gemini 4 Argon in the comparison?
Anthropic left them out of the announcement table. The system card includes just one GPT-6.1 Sol result: 58.2% on Terminal-Bench 4.0 according to the public leaderboard, compared with 67.6% for Fable 5.2. It doesn’t give a Gemini 4 Argon result alongside Fable.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (November 9, 2026)
- Claude documentation: Pricing (checked November 9, 2026)
- OpenAI: GPT-6 Astra in the API documentation (checked November 9, 2026)
- Google: Introducing Gemini 4 Carbon (October 28, 2026)
- Anthropic: Claude Fable 5.2 and Claude Mythos 5.2 System Card, pages 210-211
- OpenRouter model list (API /api/v1/models), checked on the evening of November 9, 2026
