GPT-6.2 Astra vs Claude Fable 5.2 vs Gemini 4 Carbon
GPT-6.2 Astra vs Claude Fable 5.2 and Gemini 4 Carbon, plus the upgrade from GPT-6 Astra: launch benchmarks, cost per task, prices and which to pick.
OpenAI released GPT-6.2 Astra on November 17, 2026. It's OpenAI's most capable model, it costs 20% less per token than GPT-6 Astra, and it lands at exactly the price of Anthropic's Claude Fable 5.2: $8 per million input tokens and $40 per million output tokens. Google's Gemini 4 Carbon, the third frontier model teams are weighing this month, costs a quarter of that until November 30.
This guide compares the three on the numbers available in launch week, explains what each number can and can't tell you, and answers the question many OpenAI customers have now: is it worth moving off GPT-6 Astra? Each pair links to a full side-by-side page in our AI model comparison tool.
Disclosure and scope: Van Data Team builds AI systems for clients on models from OpenAI, Anthropic and Google, and we have no paid ties to any of them. We drafted this with help from Anthropic's Claude, and Anthropic makes one of the three models compared here. Most scores below come from OpenAI's launch post, which sets its new models against their rivals; we say so each time and add independent results where they exist.
Key Takeaways
- Same price, split results: GPT-6.2 Astra and Fable 5.2 both cost $8 and $40 per million tokens. In OpenAI's charts, Astra leads on long professional tasks and Fable on code that maintainers would merge.
- Astra used far less per task: at max effort, OpenAI measured $4.47 per Agents' Last Exam task for Astra against $60.05 for Fable 5.2.
- Carbon is the budget pick: $2 and $10 until November 30, then $4 and $20, and level with Astra on OpenAI's own agent test.
- GPT-6 Astra users should test the upgrade: cheaper tokens and higher scores on all five public benchmarks OpenAI shows.
- Long prompts cost more on Astra: past 272,000 input tokens, OpenAI bills the whole request at $16 and $60.
- No independent scores yet: Arena and Epoch AI had not rated GPT-6.2 Astra when we wrote this.
GPT-6.2 Astra, Fable 5.2 and Gemini 4 Carbon at a Glance
| Model | Released | Price per 1M tokens (in / out) | Cached input | Context / max output | Where you can use it |
|---|---|---|---|---|---|
| GPT-6.2 Astra | Nov 17, 2026 | $8 / $40 | $0.40 | 1.05M / 128K | OpenAI API, ChatGPT, Codex, GitHub Copilot |
| Claude Fable 5.2 | Nov 9, 2026 | $8 / $40 | $0.20 | 1M / 128K | Claude API and apps, GitHub Copilot |
| Gemini 4 Carbon | Oct 28, 2026 | $2 / $10 to Nov 30, then $4 / $20 | $0.10, then $0.20 | 1.05M / 64K by default | Gemini API (paid tier), GitHub Copilot |
| GPT-6 Astra | Sep 3, 2026 | $10 / $50 | $1 | 1.05M / 128K | OpenAI API, GitHub Copilot |
All prices are list prices per million tokens. OpenAI bills a GPT-6.2 Astra request with more than 272,000 input tokens at $16 and $60 for the whole request, while Anthropic bills Fable 5.2 at one rate across its 1M window. Carbon has been generally available on the paid Gemini API since November 10, and Google's $2 and $10 are introductory rates that end on November 30, 2026.
What's New in GPT-6.2 Astra
OpenAI's launch post and model page list these changes from GPT-6 Astra:
- Lower prices: $8 and $40 per million tokens, down from $10 and $50, with cached input down 60%, from $1 to $0.40
- Fewer tokens per task: OpenAI credits better token efficiency, along with the lower prices, for Astra's lower cost per task
- Office work: slides, spreadsheets and documents that follow a template and a profession's conventions
- Science in a terminal: 77.1% on Terminal-Bench-Science 0.1, up from 68.1% for GPT-6 Astra
- Same limits: a 1,050,000-token context with up to 922,000 input tokens, 128,000 output tokens, and text and image input
It's in the API (Responses, Chat Completions and Batch), in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in GitHub Copilot on the Pro+, Max, Business and Enterprise plans.
There's a safety change too. OpenAI's system card rates GPT-6.2 Astra Critical for cybersecurity and High for biological and chemical capabilities, like GPT-6 Astra, and adds High for AI self-improvement, a first for a model OpenAI has released. It says it applied its strongest safeguards yet. See the full GPT-6.2 Astra model page for every spec and source.
GPT-6.2 Astra vs Claude Fable 5.2: Same Price, Split Results
The rate cards match: $8 and $40 per million tokens. The details differ:
- Cache reads: Fable 5.2 charges $0.20 per million cached tokens, half of Astra's $0.40
- Long prompts: Astra's rates jump to $16 and $60 past 272,000 input tokens; Fable 5.2 keeps one rate across its 1M window
- Speed: Fable 5.2 has an optional fast mode at $16 and $80; Astra's Fast mode costs twice its standard rate, and Ultrafast three times (six times from March 2027)
On scores, OpenAI's charts split the four tests both models appear in, at max effort:
- Astra leads on Agents' Last Exam, which tests long professional workflows: 77.6% against 72.8%
- Astra leads narrowly on AutomationBench, which tests business automation across apps: 52.4% against 51.4%
- Fable leads on FrontierCode 2.0, which asks whether code changes meet maintainers' standards for merging: 45.6% against 44.2%, and 49.1% at xhigh effort
- Fable leads on BenchCAD 2.1 Core, a test of writing CAD programs: 83.5% against 82.5%
OpenAI notes that the Agents' Last Exam leaderboard result for Fable 5.2 leaves out nine tasks that used a fallback model. On Terminal-Bench-Science, Astra scores 77.1%, and Anthropic reports 74.6% for Fable 5.2 in its own harness, so that one is close too.
The bigger gap is cost per task. In OpenAI's charts, Astra at max effort spent $4.47 per Agents' Last Exam task against $60.05 for Fable 5.2, and $5.97 per FrontierCode task against $44.70. Same price per token, very different bills, because Fable 5.2 at max effort uses far more tokens. Those are OpenAI's measurements on its own setup, so check them on your work before you rely on them.
Independent tests can't settle it yet. Fable 5.2 has the highest Epoch AI scores on FrontierMath and SimpleQA Verified among the models in our catalog and ranks first on Arena's WebDev board at 1826, while GPT-6.2 Astra had no independent rating when we wrote this. In the Artificial Analysis Intelligence Index figures OpenAI charts, Fable 5.2 still leads, 64.6 against 62.1.
Our read: if your work is long agent runs and documents, start with Astra; if it's careful code changes, Fable 5.2 is still the stronger bet, at a higher cost per task. Compare them on our Fable 5.2 vs GPT-6.2 Astra page, and see our Fable 5.2 guide for how it compares with Opus 5.5.
GPT-6.2 Astra vs Gemini 4 Carbon: The Price Gap
Gemini 4 Carbon costs $2 and $10 per million tokens through November 30, a quarter of Astra's rates. From December 1, Google charges $4 and $20, still half. Our Gemini 4 Carbon and Argon pricing guide covers that change.
On OpenAI's charts the two are close, with Carbon at high effort, its highest setting there:
- Agents' Last Exam: 77.6% for Astra against 77.4% for Carbon, a tie in practice
- AutomationBench: Carbon leads, 53.9% against 52.4%
- FrontierCode 2.0: Astra leads, 44.2% against 38.4% (Carbon scores 41.1% at medium effort)
Per task, the price gap mostly disappears in OpenAI's numbers. Carbon spent $4.60 per Agents' Last Exam task against Astra's $4.47, and $2.12 per AutomationBench task against $0.83. A cheaper token doesn't help if a model uses more of them, which is why we measure cost per task, not cost per token.
Carbon also has independent results Astra lacks: Arena Text rates it 1526, the top score in our catalog, and Epoch AI measures 96.5% on GPQA Diamond. One limit to plan for: Carbon returns at most 65,536 output tokens per request by default. See the Gemini 4 Carbon vs GPT-6.2 Astra page.
Should You Upgrade from GPT-6 Astra?
For most teams, yes, after a test run. GPT-6.2 Astra is cheaper per token and scores higher on every public benchmark OpenAI charts for both:
| At max effort, OpenAI's charts | GPT-6.2 Astra | GPT-6 Astra |
|---|---|---|
| Agents' Last Exam v1.1 | 77.6% ($4.47 per task) | 73.1% ($8.99 per task) |
| Terminal-Bench-Science 0.1 | 77.1% ($7.15) | 68.1% ($23.80) |
| FrontierCode 2.0 | 44.2% ($5.97) | 42.8% ($6.68) |
| BenchCAD 2.1 Core | 82.5% ($5.43) | 62.6% ($9.71) |
| AutomationBench 1.0.6 | 52.4% ($0.83) | 41.4% ($1.73) |
The headline claim in OpenAI's post, 5.2 points higher than GPT-6 Astra on Agents' Last Exam at under a third of the cost, compares Astra at xhigh effort (78.3%) with GPT-6 Astra at max. Either way, the new model is ahead and cheaper per task on every row.
Before you switch, rerun your own test set. A new model can follow the same prompt differently, and a 30-minute cache write costs $10 per million tokens, more than regular input, so check your caching pattern too. Compare them line by line on the GPT-6.2 Astra vs GPT-6 Astra page.
If you don't need Astra's top scores, look at GPT-6.2 Sol, released the same day at $2 and $10. On FrontierCode 2.0 it came within half a point of Astra, 43.8% against 44.2%, at about a fifth of the cost per task. We compare it with Claude Sonnet 5.5 and Gemini 4 Argon in our GPT-6.2 Sol guide, and the two OpenAI models side by side on the GPT-6.2 Astra vs GPT-6.2 Sol page.
Benchmark Scores Side by Side
Here are the public benchmarks in OpenAI's launch charts, each model at its highest effort there: max for the OpenAI and Anthropic models, high for Gemini 4 Carbon.
| Benchmark | GPT-6.2 Astra | Claude Fable 5.2 | Gemini 4 Carbon | GPT-6 Astra |
|---|---|---|---|---|
| Agents' Last Exam v1.1 | 77.6% | 72.8% | 77.4% | 73.1% |
| Terminal-Bench-Science 0.1 | 77.1% | 74.6% (Anthropic's run) | Not listed | 68.1% |
| FrontierCode 2.0 | 44.2% | 45.6% | 38.4% | 42.8% |
| BenchCAD 2.1 Core | 82.5% | 83.5% | Not listed | 62.6% |
| AutomationBench 1.0.6 | 52.4% | 51.4% | 53.9% | 41.4% |
OpenAI's Terminal-Bench-Science figure for Fable 5.2, 73.3%, comes from Artificial Analysis's run in a different harness, so we show Anthropic's own number instead. We left out OpenAI's internal tests and the evaluators' indexes, which you'll find on the GPT-6.2 Astra model page with notes on each.
Worked Example: What a Task Costs
Here's a simple cost model, using our own math and list prices. Take an agent task that reads 200,000 tokens and writes 20,000:
- GPT-6.2 Astra: $1.60 for input plus $0.80 for output, or $2.40
- Claude Fable 5.2: the same $2.40
- Gemini 4 Carbon: $0.60 through November 30, and $1.20 from December 1
- GPT-6 Astra: $2.00 plus $1.00, or $3.00
Now double the prompt to 400,000 tokens, past Astra's 272,000-token line:
- GPT-6.2 Astra: the whole request moves to $16 and $60, so $6.40 plus $1.20, or $7.60
- Claude Fable 5.2: $3.20 plus $0.80, or $4.00
- GPT-6 Astra: $20 and $75 past its threshold, so $8.00 plus $1.50, or $9.50
So for very long prompts, Fable 5.2 costs about half as much as Astra at the same base price. But these examples assume each model reads and writes the same number of tokens, and OpenAI's charts show they don't: at max effort Fable 5.2 used many more tokens per task. Measure on your own work, at the effort level you'd actually run.
Which Model Should You Use?
| If you need | Start with | Why |
|---|---|---|
| Long agent runs and office documents | GPT-6.2 Astra | Leads Agents' Last Exam in OpenAI's charts at a low cost per task |
| Code changes that maintainers will merge | Claude Fable 5.2 | Leads FrontierCode 2.0 in both OpenAI's and Anthropic's tables |
| The lowest token bill | Gemini 4 Carbon | A quarter of the price until November 30, half after |
| Prompts over 272,000 tokens | Claude Fable 5.2 | One rate across the 1M window |
| An upgrade for GPT-6 Astra workloads | GPT-6.2 Astra | Cheaper and higher on every public benchmark OpenAI shows |
Most teams will end up with two of these: a default for routine work and a stronger model for the hard cases. Log which model handled each task, so the next release becomes a quick test rather than a migration.
What the Benchmarks Don't Tell You
- Vendor charts: almost every score here comes from OpenAI, which chose the tests and ran most of them
- Effort settings: each model can run at several levels, and both scores and costs change with them
- Missing rows: Carbon is absent from two of the five tests, and Fable 5.2's Terminal-Bench-Science score comes from Anthropic
- Launch week: GPT-6.2 Astra came out on November 17, and independent results from Arena and Epoch AI will follow
Use these numbers to build a short list, then run your own tasks on two or three models and compare success rate, cost and time per finished task.
GPT-6.2 Astra: Our Take
GPT-6.2 Astra is a clear upgrade on GPT-6 Astra: cheaper per token, cheaper per task in OpenAI's charts and better on every public test OpenAI shows. Against Claude Fable 5.2, at the same price, it trades wins rather than dominating, and OpenAI's own numbers put Fable ahead on merge-quality code. Gemini 4 Carbon stays the value pick, at least until its price doubles on December 1.
Our advice: if you run GPT-6 Astra today, test the switch this week. If you're choosing between labs, run GPT-6.2 Astra and Fable 5.2 on the same twenty tasks before you commit, and keep Carbon in the test while its launch price lasts.
Van Data Team in practice
How do I choose between GPT-6.2 Astra, Claude Fable 5.2 and Gemini 4 Carbon for my agents?
Choose on your own tasks rather than on launch charts. We collect a set of real tasks from your workflow, run each model at the effort level you would pay for, and measure how many tasks it finishes, what each finished task costs and how long it takes. Each kind of task then goes to the cheapest model that clears your bar, with a stronger model as the fallback. We have built routing across Claude, OpenAI and Gemini models for a client platform.
AI Agents
Zeno Fitness and Habit PlatformMulti-provider routing across Claude 3 Haiku, OpenAI, and Gemini
AI Model Brief
The AI models worth knowing, once a week
What shipped, how it scores on independent benchmarks, what it costs per million tokens and what was deprecated, from the data behind our AI model benchmarks.
We use your address only to send the brief. See our privacy policy.
See this week’s brief
Founder of Van Data Team. Builds Vanaxity, the AI agent system for SEO, GEO, and AEO, and leads data engineering delivery for B2B teams.
About the authorAI model benchmarks
Models in this article
Current pricing, context window and independent scores for each model.
Article FAQ
Questions readers usually ask next.
These short answers clarify the practical follow-up questions that often come after the main article.
Need a similar system?
If this article maps to a workflow your team already operates, the next step is usually a scoped review of the system, constraints, and rollout path.
Consultation
Scope your next AI workflow.
Share the system and blockers. We will map fit, gaps, and the next step.




