The short answer
● Cheapest route to a frontier answer — Opus 5.5 at $4 / $20. Anthropic itself says start here and move up only when a task needs it.
● Best scores, and now cheaper than GPT-6 — Fable 5.2 at $8 / $40. Wins nine of eleven published benchmarks; Astra is $10 / $50.
● The one to stop quoting — Gemini 4 Argon at $2 / $10 — it becomes $4 / $20 on 1 December 2026.
● The one you probably cannot buy — Gemini 4 Carbon. It is in the benchmark table and not on Google’s public price list.
Fable 5.2’s launch reset the price ladder, so the useful question is no longer which model is strongest. It is which one you should be paying for. Anthropic publishes head-to-head figures against Opus 5.5, GPT-6 Astra and Gemini 4 Carbon, so for once the table can be built from one source rather than stitched together.
Price first, because it decides most of this
| Model |
Input / MTok |
Output / MTok |
Context |
Notes |
| Claude Fable 5.2 |
$8 |
$40 |
1M |
Fast mode $16 / $80, up to 2.5× faster |
| Claude Fable 5.1 |
$10 |
$50 |
1M |
Previous generation — the 20% cut is from here |
| Claude Opus 5.5 |
$4 |
$20 |
1M |
Anthropic’s recommended starting point |
| Claude Sonnet 5.5 |
$2 |
$10 |
1M |
Fast latency |
| Claude Haiku 5.5 |
from $0.10 |
from $0.50 |
1M |
Cheapest Claude |
| GPT-6 Astra |
$10 |
$50 |
1.05M |
25% more expensive than Fable 5.2 |
| Gemini 4 Argon |
$2 → $4 |
$10 → $20 |
— |
Price doubles 1 December 2026 |
| Gemini 4 Carbon |
not published |
not published |
— |
Not on Google’s public price list |
Claude prices from the Claude Platform model pages. GPT-6 Astra from OpenAI’s model page. Gemini from Google AI Studio’s pricing page.
Two things fall out. Fable 5.2 undercuts GPT-6 Astra by 20% on both sides while beating it on every benchmark Anthropic reported — a position Anthropic did not hold a generation ago. And Gemini 4 Argon is mid-price-change: $2 / $10 through 30 November, $4 / $20 from 1 December. Any cost model built on Argon this month is built on a number with twenty days left on it.
On Gemini 4 Carbon. It appears in Anthropic’s comparison and wins a row. It is not listed on Google AI Studio’s pricing page, so we cannot quote a price and will not invent one. Treat its benchmark results as a capability signal rather than a purchasing option until Google publishes terms.
The benchmarks
| Benchmark |
Fable 5.2 |
Opus 5.5 |
GPT-6 Astra |
Gemini 4 Carbon |
Fable 5.1 |
| FrontierCode 2.0 agentic coding |
49.1% |
45.5% |
42.8% |
41.1% |
32.4% |
| FrontierSWE v2 agentic coding |
69.9% |
62.3% |
65.5% |
61.2% |
56.3% |
| Terminal-Bench-Science 0.1 research |
74.6% |
58.7% |
68.1% |
64.8% |
52.6% |
| OSWorld 2.1 computer use |
89.3% |
87.2% |
79.5% |
81.5% |
83.5% |
| RoboDojo Sim robotics |
64.9% |
39.4% |
29.8% |
— |
27.6% |
| Remote Labor Index knowledge work |
41.3% |
21.3% |
20.8% |
— |
17.9% |
| Humanity’s Last Exam with tools |
74.2% |
67.7% |
57.2% |
— |
65.6% |
| AA-Briefcase v1.1 rubric |
66.1% |
61.2% |
52.0% |
61.8% |
61.5% |
| AA-Briefcase Elo |
1785 |
1800 |
1569 |
1611 |
1669 |
| Chartography no tools |
73.4% |
64.4% |
71.0% |
72.5% |
44.8% |
| AutomationBench business workflows |
51.4% |
42.5% |
41.4% |
53.9% |
31.4% |
Figures as published by Anthropic. FrontierCode is at Xhigh effort for both Fable models and Medium for Gemini 4 Carbon; Remote Labor Index is at High effort. Bold marks the two rows Fable 5.2 does not win. Dashes are results Anthropic did not report, not zeros.
Nine of eleven to Fable 5.2. The two it loses are worth naming precisely: Gemini 4 Carbon takes AutomationBench on business workflows, 53.9% to 51.4%, and Opus 5.5 takes the AA-Briefcase Elo row, 1800 to 1785 — while losing the rubric score on that same benchmark, 61.2% to 66.1%. Against GPT-6 Astra there is no row where Astra leads.
The part that changes your bill: effort
Fable 5.2 takes an effort parameter — Low, Medium, High, Xhigh, Max — and the same model spans roughly a tenfold cost range across it. Medium is the default in the Claude apps and Claude Code; the API defaults to High. Anthropic’s own figures make the trade-off unusually concrete:
| Task |
Fable 5.2 effort |
Result |
Cost vs Opus 5.5 top score |
| FrontierCode |
High |
45.3% vs Opus’s 45.5% |
about half |
| FrontierCode |
Xhigh |
49.1% — best published |
higher |
| WANDR agentic search |
Medium |
matched Opus’s top score |
about 40% |
| Humanity’s Last Exam |
Low |
matched Opus’s top score |
under one-tenth |
| BenchCAD 2.1 hardware |
Max |
83.5 vs Opus’s 65.3 |
higher |
Read that table before you pick a model. A Low-effort Fable call matching Opus 5.5 at under a tenth of the cost per task is a stronger argument than any benchmark row here. The corollary: if you deploy Fable at Xhigh by default, you will pay for headroom most of your traffic does not use.
When each one is the right answer
Claude Opus 5.5 — $4 / $20
Anthropic’s own recommendation is to start here for most coding and knowledge tasks, then move up or down. It wins the AA-Briefcase Elo row outright. At half Fable’s price with scores in the same band on several benchmarks, this is the default unless you have measured a reason to leave.
Claude Fable 5.2 — $8 / $40
For long-horizon agentic work: sessions that run for hours, multistep research, analysis carried through to a finished document. The standout margins are robotics (64.9% against Opus’s 39.4%), knowledge work on the Remote Labor Index (41.3% against 21.3%) and scientific research (74.6% against 58.7%). It can also act as an Advisor consulted by Haiku, Sonnet or Opus, or as an Orchestrator delegating to cheaper models — which is how you get the capability without paying Fable rates for every token.
GPT-6 Astra — $10 / $50
1.05M context and a 30 April 2026 cutoff. It leads no row in Anthropic’s table and costs 25% more than Fable 5.2. It is closest on FrontierSWE v2 at 65.5% against 69.9%. If you are on Astra, the honest read is that the price gap now runs the wrong way.
Gemini 4 Carbon — price not published
Wins AutomationBench at 53.9% and is competitive on Chartography at 72.5%. But it is absent from Google’s public pricing, and the Gemini 4 model Google does list is Argon, whose price doubles on 1 December. Strong on paper, hard to plan around.
Decision framework
- You have not measured anything yet → Opus 5.5 at $4 / $20. Measure, then move.
- Agents running for hours unattended → Fable 5.2. The Remote Labor Index gap is the closest proxy published for this and it is nearly double.
- High-volume, latency-sensitive, simple → Haiku 5.5 from $0.10 / $0.50, or Fable at Low effort if quality matters more than latency.
- Robotics or hardware design → Fable 5.2, and not close: 64.9% on RoboDojo against Opus’s 39.4%, and 83.5 on BenchCAD against 65.3.
- Business process automation → benchmark Gemini 4 Carbon yourself if you can get access; it is the one row Fable loses to it.
- Currently on GPT-6 Astra → price out Fable 5.2 at Medium or High effort. Lower list price and better published scores.
- Costing a build on Gemini 4 Argon → redo the model with $4 / $20, not $2 / $10.
- Security or biology work that gets declined → that is the Fable vs Mythos question.
What none of these fixes
Fable 5.2 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it. Forced tool use returns a 400 error, and you must omit temperature, top_p and top_k. Thinking is adaptive and cannot be turned off. If any of those is load-bearing in your stack, the benchmark table is beside the point.
Frequently asked questions
Is Claude Fable 5.2 better than GPT-6 Astra?
On the eleven benchmarks Anthropic published, Fable 5.2 leads every row where both were measured, including FrontierSWE v2 at 69.9% against 65.5% and OSWorld 2.1 at 89.3% against 79.5%. Fable 5.2 is also $8 / $40 per million against Astra’s $10 / $50. Note the benchmarks come from Anthropic, so run your own evaluation on your own tasks before migrating.
Should I pay double for Fable 5.2 over Opus 5.5?
Not by default. Anthropic recommends starting on Opus 5.5. The case for Fable is long-horizon agentic work, robotics, hardware and scientific research, where the margins are large. And Fable at High effort scored 45.3% on FrontierCode against Opus’s 45.5% at roughly half the cost per task, so effort setting can make Fable the cheaper choice rather than the dearer one.
What is the cheapest way to use Fable 5.2?
Lower the effort parameter, cache aggressively and batch. Cache reads are $0.20 per million against $8 for fresh input, and the Batch API is 50% off. You can also run Fable as an Orchestrator that delegates to Haiku or Sonnet, so Fable rates apply only to the reasoning.
How much is Gemini 4 Carbon?
Google has not published a price for it. It is not listed on Google AI Studio’s pricing page. Gemini 4 Argon is listed at $2 / $10 per million through 30 November 2026, rising to $4 / $20 on 1 December 2026.
Which has the largest context window?
GPT-6 Astra at 1,050,000 tokens, narrowly. Fable 5.2, Mythos 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 are all 1M with 128K max output. In practice the difference is not decisive.
Which model has the most recent knowledge?
Fable 5.2 and Mythos 5.2, with a June 2026 cutoff. GPT-6 Astra’s is 30 April 2026.