Three frontier models now fight for the top spot: Claude Fable 5.2 (November 9), Gemini 4 Carbon (November 10) and GPT-6 Astra (September 3). Each claims to be the best.
This comparison uses only published prices and published benchmark numbers, with sources. It covers cost, context, speed, coding and which model fits which job.
- Price (input / output per MTok): Fable 5.2 $8 / $40, GPT-6 Astra $10 / $50, Gemini 4 Carbon $2 / $10 until November 30, then $4 / $20.
- Context: all three offer about 1M tokens. Astra lists 1.05M; Carbon’s 1M is shared between input and output.
- Benchmarks: in Anthropic’s published table, Fable 5.2 leads most rows, such as 69.9% on FrontierSWE v2. Carbon leads AutomationBench.
- Independent index: Artificial Analysis lists Fable 5.2 at 65, Carbon at 60 and Astra at 57.
- Pick: Fable 5.2 for the hardest agentic and coding work, Carbon for value, Astra if you live in ChatGPT and Codex.
The Three Models at a Glance
| Spec | Claude Fable 5.2 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|
| Maker | Anthropic | OpenAI | |
| Released | November 9, 2026 | September 3, 2026 | November 10, 2026 (broad access) |
| API model ID | claude-fable-5-2 |
gpt-6-astra |
gemini-4-carbon |
| Input / output per MTok | $8 / $40 | $10 / $50 | $2 / $10 intro, then $4 / $20 |
| Context window | 1M | 1.05M | Up to 1M combined input and output |
| Max output | 128K | 128K | Shares the 1M budget |
| Knowledge cutoff | June 2026 | April 30, 2026 | March 2026 |
| Input types | Text, images | Text, images | Text, images, audio, video |
| Fast mode | Yes, $16 / $80, up to 2.5x output speed | Yes, 2x the standard rates | None announced |
API Pricing Compared
All prices below are per million tokens and come from each company’s official pages. Fable 5.2 cut the price from Fable 5.1’s $10 / $50.
| Price line | Fable 5.2 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|
| Input | $8 | $10 | $2 (to Nov 30), $4 after |
| Output | $40 | $50 | $10 (to Nov 30), $20 after |
| Cached input | $0.20 (cache hit) | $1.00 | Not in the launch post |
| Batch | 50% off ($4 / $20) | 50% off (Batch and Flex) | Not in the launch post |
| Long prompts | Same rate across 1M | Over 272K input: 2x input, 1.5x output | Not in the launch post |
| Fast mode | $16 / $80 | 2x applicable rates | None announced |
What one big request costs
Take a request with 100,000 input tokens and 10,000 output tokens, with no caching or batch discount:
- Fable 5.2: $0.80 input + $0.40 output = $1.20.
- GPT-6 Astra: $1.00 input + $0.50 output = $1.50.
- Gemini 4 Carbon (intro): $0.20 input + $0.10 output = $0.30.
- Gemini 4 Carbon (from December): $0.40 + $0.20 = $0.60.
Token prices are not the whole story. Thinking models spend different amounts of output on reasoning, so cost per finished task can differ from cost per token.
Note: Artificial Analysis lists a cost to run its Intelligence Index of $26.54 for Fable 5.2 (Max) and $6.24 for GPT-6 Astra (Max). Google says Carbon ranked first on several agent benchmarks at more than 2x lower cost per task than the next-best external model.
For every model’s full price list, see our AI API pricing comparison.
Context Window and Output Limits
All three handle roughly a million tokens, which is enough for a large codebase or several long books. The details differ.
- Fable 5.2: 1M context, up to 128K output, and the same per-token price across the whole window.
- GPT-6 Astra: 1.05M context and 128K output, but prompts over 272K input tokens cost 2x on input and 1.5x on output.
- Gemini 4 Carbon: a combined input and output window of up to 1M. It is the only one that also takes audio and video.
On long-context retrieval, Google’s model card reports Carbon at 96.2% on GraphWalks (256K to 1M, BFS F1), up from 84.2% for Gemini 4 Argon.
Benchmarks: Only Published Numbers
The most useful single table comes from Anthropic’s Fable 5.2 announcement, which lists all three models. These are vendor-published numbers, so treat them as claims, not independent tests.
| Benchmark | Fable 5.2 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|
| FrontierSWE v2 | 69.9% | 65.5% | 61.2% |
| FrontierCode 2.0 | 49.1% (xhigh) | 42.8% | 41.1% (medium) |
| Terminal-Bench-Science 0.1 | 74.6% | 68.1% | 64.8% |
| OSWorld 2.1 (offline, partial) | 89.3% | 79.5% | 81.5% |
| AutomationBench | 51.4% | 41.4% | 53.9% |
| Chartography (no tools) | 73.4% | 71.0% | 72.5% |
| AA-Briefcase v1.1 (rubric) | 66.1% | 52.0% | 61.8% |
| AA-Briefcase v1.1 (Elo) | 1785 | 1569 | 1611 |
| Remote Labor Index | 41.3% | 20.8% | Not listed |
| Humanity’s Last Exam (with tools) | 74.2% | 57.2% | Not listed |
| RoboDojo Sim | 64.9% | 29.8% | Not listed |
The Carbon numbers in Anthropic’s table match Google’s own model card for FrontierSWE (61.2%), AutomationBench (53.9%), OSWorld 2.1 (81.5%), Terminal-Bench-Science (64.8%) and Chartography (72.5%). That cross-check is a good sign.
Independent ranking: Artificial Analysis
Artificial Analysis runs its own test suite. At the time of writing, its model pages show:
- Fable 5.2 (Max): Intelligence Index 65, ranked #1 of 74.
- Gemini 4 Carbon (High): Intelligence Index 60, ranked #4.
- GPT-6 Astra (Max): Intelligence Index 57, ranked #8.
These figures update often, so check the live pages before you decide. Our November 2026 AI leaderboard tracks more models.
Warning: OpenAI’s GPT-6 Astra post claims state-of-the-art results on benchmarks such as Agents’ Last Exam and TerminalBench-4.0 at launch in September. Those claims predate Fable 5.2 and Carbon, and the scores were only shown as images, so they are not repeated here.
Speed and Latency
Speed has two parts: how long until the answer starts, and how fast tokens stream after that. Thinking models can pause a long time before answering on hard prompts.
| Measure (Artificial Analysis) | Fable 5.2 (Max) | GPT-6 Astra (Max) | Gemini 4 Carbon (High) |
|---|---|---|---|
| Output speed | 101 tokens/s | 37 tokens/s | 80 tokens/s |
| Time to first answer token | About 234 s | About 238 s | About 67 s |
Those first-token times reflect heavy reasoning on benchmark prompts. Simple prompts at lower effort return much sooner.
- Fable 5.2: Anthropic says output is up to 30% faster than Fable 5.1, and fast mode adds up to 2.5x more output speed for double the price. See our Claude fast mode guide.
- GPT-6 Astra: OpenAI offers a fast mode at 2x the standard rates.
- Gemini 4 Carbon: Google made no speed claims, and notes Carbon uses more tokens than Argon to explore thoroughly.
Coding and Agents
For coding agents, the most relevant published rows are FrontierSWE v2, FrontierCode 2.0 and Terminal-Bench-Science. Fable 5.2 leads all three in Anthropic’s table.
- Fable 5.2: 69.9% on FrontierSWE v2, up from 56.3% on Fable 5.1. Anthropic still suggests starting most coding work on Opus 5.5 and moving up when needed.
- GPT-6 Astra: second on FrontierSWE v2 at 65.5%. It is built into Codex, with hosted shell, apply patch and computer use tools in the API.
- Gemini 4 Carbon: 61.2% on FrontierSWE v2, but Google reports first place on CursorBench 4.0 and Zapier’s AutomationBench.
For computer use, Fable 5.2 posts 89.3% on OSWorld 2.1 (offline, partial), against 81.5% for Carbon and 79.5% for Astra. For business automation, Carbon’s 53.9% on AutomationBench edges Fable’s 51.4%.
Tip: Benchmarks rarely match your codebase. Run the same five real tickets through each model and compare pass rates and total cost per ticket, not just scores.
How to Get Access
Claude Fable 5.2
- API:
claude-fable-5-2on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. - Apps: eligible paid Claude plans. Anthropic says Fable is included for up to 50% of weekly usage limits on Pro and standard Team seats.
- Data rule: requires 30-day data retention, so zero-retention organizations need Anthropic’s approval.
GPT-6 Astra
- API:
gpt-6-astraon the OpenAI API, plus Amazon Bedrock and a Google Cloud private preview. - Apps: ChatGPT Pro, Enterprise and Business Premium, and Codex. OpenAI said Plus and Business would follow within days of launch.
Gemini 4 Carbon
- API: Google AI Studio and Gemini for business, for paying developers and enterprises.
- Apps: rolling out to Google AI Ultra subscribers in the Gemini app, Antigravity and AI Mode in Search.
- Special programs: free 6-month access for approved scientists, and “Carbon for cyber” through the Fairwind Program.
More on each model: our Fable 5.2 guide and our Gemini 4 Carbon benchmarks and access guide.
Which One Should You Pick?
| If you need… | Pick | Why |
|---|---|---|
| The hardest long-running agent or coding tasks | Fable 5.2 | Top published scores on FrontierSWE, Terminal-Bench-Science, OSWorld and Remote Labor Index |
| The lowest bill at frontier quality | Gemini 4 Carbon | $2 / $10 until November 30, still half of Fable’s price after |
| Audio or video input | Gemini 4 Carbon | The only one of the three that takes audio and video |
| Business workflow automation | Gemini 4 Carbon or Fable 5.2 | Carbon leads AutomationBench; Fable is close behind |
| Deep ChatGPT and Codex integration | GPT-6 Astra | Native in ChatGPT, Codex and OpenAI’s tool stack |
| Fast streaming for live users | Fable 5.2 (fast mode) | Highest measured output speed, plus up to 2.5x more with fast mode |
Many teams will mix models: Carbon or a cheaper model for routine work, and Fable 5.2 for the tasks that keep failing. That beats paying top rates for everything.
How to Test All Three Yourself
- Pick 10 real tasks. Use actual tickets, documents or prompts from your work, not toy questions.
- Write a pass rule for each. For example, “tests pass” or “the summary lists all five risks”.
- Get API keys. Create keys in the Claude Console, the OpenAI platform and Google AI Studio.
- Match the settings. Use similar effort levels (for example high on all three) and the same system prompt.
- Log tokens and time. Record input tokens, output tokens, wall time and pass or fail for every run.
- Work out cost per pass. Divide total cost by the number of passed tasks for each model.
- Re-run after price changes. Carbon’s price doubles on December 1, so check your maths again then.
Troubleshooting: Common Problems and Fixes
Fable 5.2 returns a 400 when I set thinking
Thinking is always on for Fable 5.2. Remove the thinking setting or send adaptive, and control depth with the effort setting instead.
Fable 5.2 says my organization is not allowed
Fable 5.2 needs 30-day data retention. Organizations set to zero data retention get an error unless Anthropic has authorized them.
My Astra bill jumped on long prompts
Requests over 272K input tokens cost 2x on input and 1.5x on output for the whole request. Trim context or split the job.
Carbon is not in the Gemini app
App access is rolling out to Google AI Ultra subscribers first. Developers can use it now through AI Studio with a paid account.
Carbon costs more than I budgeted in December
The $2 / $10 rate is introductory and ends November 30, 2026. From December the rate is $4 / $20.
The first token takes minutes
All three think before answering at high effort. Lower the effort for simple tasks, and stream responses so users see progress.
Frequently Asked Questions
Is Claude Fable 5.2 better than GPT-6 Astra?
In Anthropic’s published table, Fable 5.2 beats Astra on every shared benchmark, and Artificial Analysis ranks it higher (65 vs 57). It is also cheaper per token at $8 / $40 versus $10 / $50.
Is Gemini 4 Carbon better than Fable 5.2?
Fable 5.2 leads most published rows, but Carbon leads AutomationBench at 53.9% versus 51.4%. Carbon costs a quarter of Fable’s price until November 30 and half after.
Which is the cheapest of the three?
Gemini 4 Carbon, at $2 / $10 per million tokens through November 30, 2026, and $4 / $20 after. GPT-6 Astra is the most expensive at $10 / $50.
Which has the biggest context window?
GPT-6 Astra lists 1.05M tokens. Fable 5.2 has 1M, and Carbon has up to 1M shared between input and output.
Which is fastest?
Artificial Analysis measures Fable 5.2 at about 101 output tokens per second, Carbon at 80 and Astra at 37. Fable 5.2 also offers a paid fast mode with up to 2.5x more output speed.
Are these benchmark numbers independent?
The big table comes from Anthropic, though Google’s model card confirms the Carbon figures. The Artificial Analysis index is independent but changes often.
Which model is best for coding?
On published coding benchmarks, Fable 5.2 leads with 69.9% on FrontierSWE v2. Google reports Carbon first on CursorBench 4.0, so test both on your own code.
Sources
- Anthropic: Claude Fable 5.2 and Mythos 5.2
- Claude Docs: Pricing
- OpenAI Developers: GPT-6 Astra model page
- OpenAI Developer Community: Introducing GPT-6 Astra
- Google: Gemini 4 Carbon is now available
- Google DeepMind: Gemini 4 Carbon model card
- Artificial Analysis: Claude Fable 5.2
- Artificial Analysis: Gemini 4 Carbon
- Artificial Analysis: GPT-6 Astra
