GPT-6.2 Astra vs Claude Fable 5.2 vs Gemini 4 Carbon vs Opus 5.5: The November 2026 Frontier Compared
Nine days, three frontier launches. Gemini 4 Carbon reached general availability on November 10 after a limited October 28 debut, Claude Fable 5.2 arrived on November 9 at a lower price than Fable 5.1, and GPT-6.2 Astra landed on November 17 with a 20% price cut. Add Claude Opus 5.5, still the value pick at half Fable’s price, and you have four models that all claim the top spot somewhere.
This is the comparison for people who will put one of these into a product or a daily workflow: what each one wins, what the independent tests say, and what it really costs per finished task.

- AA Index #1
- Fable 5.2
- AA Index #2
- GPT-6.2 Astra
- LMArena #1 (prelim.)
- Carbon 1526
- Cheapest of the four
- Opus 5.5 $4/$20
Quick verdict
| If your main job is… | Pick | Headline number |
|---|---|---|
| Long autonomous agent runs, large codebases | Claude Fable 5.2 | FrontierCode 2.0 49.1%, SWE-bench Pro 93.0%, Remote Labor Index 41.3% |
| Professional deliverables: slides, spreadsheets, legal, science | GPT-6.2 Astra | AA Index #2; leads Terminal-Bench-Science and Harvey LAB-AA per OpenAI |
| Business-process automation, security work | Gemini 4 Carbon | AutomationBench 53.9% (best in Anthropic’s own table) |
| Hard work at a sensible price | Claude Opus 5.5 | AA Index 59 at $4/$20; best AA-Briefcase Elo (1800) |
| High volume | None of these: use Sol, Sonnet or Hydrogen | $1.50–$2 input per 1M tokens |
Spec sheet
| GPT-6.2 Astra | Claude Fable 5.2 | Gemini 4 Carbon | Claude Opus 5.5 | |
|---|---|---|---|---|
| Maker | OpenAI | Anthropic | Google DeepMind | Anthropic |
| Released | Nov 17, 2026 | Nov 9, 2026 | Oct 28 (limited); GA Nov 10 | Sept 22, 2026 |
| Price per 1M (in / out) | $8 / $40 | $8 / $40 | $2 / $10 intro; $4 / $20 from Dec 1 | $4 / $20 |
| Cached input per 1M | $0.40 | $0.20 | Implicit caching only | — |
| Context window | Not stated at launch (GPT-6: 1.05M) | 1M | 1,048,576 | 1M |
| Fast mode | Ultrafast (API, Pro plans) | $16 / $80, up to 2.5x faster | — | — |
| Model ID | gpt-6.2-astra | claude-fable-5-2 | gemini-4-carbon | claude-opus-5-5 |
The headline prices hide a twist: GPT-6.2 Astra and Fable 5.2 now cost exactly the same per token. Gemini 4 Carbon is the cheapest flagship until November 30, then doubles to Opus 5.5’s price.
Independent scores
Launch tables are chosen by the maker. These are the numbers from third parties that ran every model the same way.
| Independent test | GPT-6.2 Astra | Claude Fable 5.2 | Gemini 4 Carbon | Claude Opus 5.5 |
|---|---|---|---|---|
| AA Intelligence Index | #2 | #1 | 60 | 59 |
| LMArena (blind votes, Nov 9) | Pending | Not yet listed | 1526 (#1, prelim.) | 1506 (#4) |
| ARC-AGI-2 | — (GPT-6 Astra: 95.0%) | 93.8% | — | 93.3% |
| AA-Omniscience hallucination rate | 59% | — | — | — |
| AA-3D (Elo) | 1318 | 1271 | — | — |
| BenchLM overall | 89.4 (est.) | 92.8 (est.) | 85.0 (est.) | 85.8 |
The picture is consistent across sources: Fable 5.2 is first on capability, Astra a close second, Carbon and Opus 5.5 a step behind and nearly tied. Carbon’s LMArena lead shows people prefer how it answers in chat, which is not the same as getting hard tasks right.
Launch-table benchmarks
Anthropic’s Fable 5.2 table includes Carbon and Opus 5.5 (and GPT-6 Astra, since GPT-6.2 came later):
| Benchmark | Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | 69.9% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 58.7% | 68.1% | 64.8% |
| OSWorld 2.1 (partial) | 89.3% | 87.2% | 79.5% | 81.5% |
| AA-Briefcase (Elo) | 1785 | 1800 | 1569 | 1611 |
| Remote Labor Index | 41.3% | 21.3% | 20.8% | — |
| Humanity’s Last Exam (tools) | 74.2% | 67.7% | 57.2% | — |
OpenAI’s own claims for GPT-6.2 Astra, against Fable 5.2:
- Agents’ Last Exam: 5.5 points above Fable 5.2 at about 5% of its cost (Fable’s score excludes nine fallback tasks).
- Harvey LAB-AA v1.1 (legal): more than double Fable 5.2’s pass rate at about one-sixth of the cost.
- Terminal-Bench-Science 0.1: nine points above GPT-6 Astra; OpenAI leads this test per Artificial Analysis.
- GDPval v2: 6.2 points above GPT-6 Astra with 22% fewer output tokens.
Each maker picks tests that flatter its model. Read these tables as “where each model is strong”, not as a ranking.
What each model is best at
Claude Fable 5.2
The strongest model for long, unsupervised work: multi-hour coding runs, research agents, computer use. It nearly doubled Fable 5.1 on the Remote Labor Index (41.3% vs 17.9%) and leads every agentic coding row. Watch the bill: one hands-on test found it wrote twice the tokens of Fable 5.1 and cost 56% more per task, even at a lower per-token price. Anthropic says Fable 5.2 at High effort matches Opus 5.5’s best FrontierCode score at about half the cost, so don’t default to max effort.
GPT-6.2 Astra
OpenAI’s pitch is professional deliverables: slides, spreadsheets, legal memos and scientific terminal work. It leads AA-3D and AA-Robotics, and it is cheaper than GPT-6 Astra ($8/$40 vs $10/$50, cached input $0.40 vs $1.00). The weak spot is factual restraint: a 59% AA-Omniscience hallucination rate means it often guesses rather than admits uncertainty. Ground it in your own documents.
Gemini 4 Carbon
Google’s deepest-thinking Gemini 4 model, tuned first for cybersecurity (finding, validating and patching vulnerabilities) and strongest in Anthropic’s table on AutomationBench. It is the cheapest flagship at $2/$10 until November 30. Pair it with Gemini 4 Hydrogen: Google says using Carbon to plan and Hydrogen to execute came close to Carbon’s Terminal-Bench score while finishing 19% faster. More in our Gemini 4 lineup guide.
Claude Opus 5.5
Still the best price-to-capability ratio at the top: AA Index 59, the top AA-Briefcase Elo, #1 on WebDev Arena (1810), and SWE-bench Pro 89.9%, all at $4/$20. For most teams it is the default, with Fable 5.2 as the escalation.
Price per token vs price per task
Per-token prices mislead because models use very different numbers of tokens to finish the same job.
| Model | Price per 1M (in / out) | Cost per AA Index task | Notes |
|---|---|---|---|
| Claude Fable 5.2 | $8 / $40 | $26.54 | Highest score, highest cost per task |
| GPT-6.2 Astra | $8 / $40 | Lower than Fable (OpenAI holds most of AA’s cost frontier) | Fewer output tokens per task |
| Gemini 4 Carbon | $2 / $10 › $4 / $20 | — | Double the price after Nov 30 |
| Claude Opus 5.5 | $4 / $20 | — | Half Fable’s token price |
| GPT-6.2 Sol (for reference) | $2 / $10 | About 15% of GPT-6 Astra’s | Matches GPT-6 Astra’s score |
The practical rule: send routine requests to a mid-tier model and escalate only the hard ones. Our model routing guide shows how to do that without an extra classifier call.
Confirm with three tasks from your own work before committing
Test before you switch
- Pick 10–20 real tasks from last month, with known good answers.
- Run each model at two effort levels (for example High and max). Record pass rate, tokens and time.
- Compute cost per successful task, not cost per token.
- Re-run monthly. At this release pace, today’s winner has a shelf life of weeks.
Our guide to AI agent testing and evaluation covers how to build that test set.
FAQ
Is GPT-6.2 Astra better than Claude Fable 5.2?
On the independent Artificial Analysis Intelligence Index, no: Fable 5.2 is #1 and Astra #2. OpenAI reports Astra ahead on Agents’ Last Exam, Harvey LAB-AA and AA-3D, often at a fraction of Fable’s cost per task. Fable 5.2 leads agentic coding and long-horizon work.
How much does GPT-6.2 Astra cost?
$8 per million input tokens and $40 per million output tokens, with cached input at $0.40. That is 20% below GPT-6 Astra and the same per-token price as Claude Fable 5.2.
Is Gemini 4 Carbon available?
Yes. Google announced it on October 28, 2026 for cyber defenders first, and it became generally available on November 10. Introductory pricing is $2/$10 per million tokens through November 30, then $4/$20.
Should I use Claude Fable 5.2 or Opus 5.5?
Start with Opus 5.5 for most work: AA Index 59 at $4/$20. Escalate to Fable 5.2 ($8/$40) for long autonomous agent runs and the hardest coding tasks. Anthropic says Fable 5.2 at High effort matches Opus 5.5’s best FrontierCode score at about half the cost per task.
Which AI model is the smartest in November 2026?
Claude Fable 5.2 leads the Artificial Analysis Intelligence Index and BenchLM’s overall ranking. GPT-6.2 Astra is second on AA. Gemini 4 Carbon leads LMArena’s blind preference votes.
Sources
Scores as published November 9–18, 2026. Launch-table numbers come from each maker and may change as benchmarks are re-run.
- Anthropic: Claude Fable 5.2
- Artificial Analysis: OpenAI has the #2 model with GPT-6.2 Astra and Sol
- explainx: GPT-6.2 Astra and Sol launch, pricing and benchmarks
- OfficeChai: GPT-6.2 Astra and Sol benchmarks
- BenchLM: Frontier AI models, November 2026
- GrandLinux: AI models compared, November 2026
- TeamoRouter: Gemini 4 Hydrogen vs Carbon vs Argon
- VentureBeat: Google launches Gemini 4 Hydrogen
- eesel: Claude Fable 5.2 pricing, benchmarks and changes from 5.1
- LogRocket: AI dev tool power rankings, November 2026
