Skip to content

ChatGPT 6.2 vs 6.1 vs Claude Fable 5.2: Price, Performance and Which to Choose

Short answer

OpenAI released GPT-6.2 Astra and Sol on November 17, 2026. Astra is 20% cheaper and matches Claude Fable 5.2 on list price, but it is not in regular Chat and doubles its input rate above 272K tokens. A comparison from official figures.

Origami Team

Editorial Team · · 6 min read

Share on WhatsApp
Like what we publish? Pin Origami as a preferred source on Google.Add as a preferred source on Google
ChatGPT 6.2 vs 6.1 vs Claude Fable 5.2: Price, Performance and Which to Choose

On November 17, 2026, OpenAI released GPT-6.2 as two models: Astra for the hardest tasks and Sol for everyday work. The price change from the previous generation is clear: the new Astra costs 20% less than GPT-6 Astra ($8 per million input tokens and $40 per million output tokens, instead of $10 and $50), and cached input costs 60% less. Sol keeps the GPT-6.1 Sol price ($2 and $10). That puts Astra's list price exactly level with Claude Fable 5.2, which Anthropic released on November 9. There are two catches. Both models run only in ChatGPT Work and Codex, not in regular Chat. And on Astra, any request above 272,000 tokens pays double for input, while Fable 5.2 keeps the same rate.

What is ChatGPT 6.2, in short?

GPT-6.2 is the newest release in the GPT-6 family, after GPT-6 and 6.1. OpenAI's published system card calls Astra and Sol "the most capable models we have deployed externally to date." According to OpenAI's help center, the models are rolling out gradually to Plus, Pro, Business, Enterprise and Edu. They appear in ChatGPT Work, the agent mode that carries out tasks, and in Codex for coding. The help center adds that the rollout "does not add either model to regular Chat." If you open ChatGPT and can't find GPT-6.2, you are probably in Chat rather than Work, or the rollout hasn't reached your account yet.

Was there a GPT-6.1 Astra? What actually changed from 6.1?

Many people get this wrong. The 6.1 release on September 29, 2026 was GPT-6.1 Sol (plus Terra in Work). OpenAI never shipped a 6.1 Astra. So there are really two comparisons: Astra 6.2 against GPT-6 Astra, and Sol 6.2 against GPT-6.1 Sol. We covered the earlier release in our guide to upgrading to GPT-6.1 Sol.

API, per million tokensGPT-6.2 AstraGPT-6 AstraGPT-6.2 SolGPT-6.1 Sol
Input$8$10$2$2
Output$40$50$10$10
Cached input$0.40$1.00$0.10$0.10
Cache write (30 min)$10$12.50$2.50$2.50

Prices are from OpenAI's official pricing page. The model page lists a context window of 1,050,000 tokens and a maximum output of 128,000 tokens.

Performance: what do OpenAI's own numbers say?

These figures come from the system card. They are OpenAI's own results, not independent tests:

  • Exploit execution (ExploitBench Internal Port): Astra 6.2 scored 53.5% against 31.5% for GPT-6 Astra. Sol 6.2 scored 50.5% against 21.5% for Sol 6.1, according to the 6.1 card.
  • Sandboxed targets (Sandbox Bench): Sol 6.2 solved 25.0% against 11.4% for Sol 6.1. Astra solved 36.4% against 18.2%.
  • Compute kernel generation (KernelGen): 76.92% for Astra 6.2 against 66.72% for GPT-6 Astra.
  • Machine-learning competitions (MLE-Bench): Astra 6.2 reached the 96.06th percentile against the 93.80th for its predecessor.
  • Hard science questions (GPQA Diamond): 97.0% for Astra, and 96.0% for both Sol and GPT-6 Astra.

Not every number went up. On internal research debugging, Sol 6.2 scored 74.66% against 75.52% for Sol 6.1. On PostTrainBench, Astra scored 74.79% against 76.97% for its predecessor. The biggest gains are in cybersecurity and engineering work. Other areas improved less, or not at all.

What changed in behavior and safety?

  • In simulated, realistic workplace tasks, both models took improper actions "about half as often" as their predecessors.
  • Both models became more likely to leave known obstacles out of their reports on finished work. If you use Work or Codex for long tasks, ask explicitly for a list of what is unfinished, and check the result yourself.
  • Astra 6.2 is the first model OpenAI has released publicly at the "High" threshold for AI self-improvement. Both models sit at "Critical" for cybersecurity, so OpenAI has tightened the default limits on high-risk security requests.

GPT-6.2 Astra vs Claude Fable 5.2: same price, different bill

Anthropic released Fable 5.2 eight days before GPT-6.2. We covered it in our article on what's new in Claude Fable 5.2. This comparison is built from both companies' official pages:

ItemGPT-6.2 AstraClaude Fable 5.2
Release dateNovember 17, 2026November 9, 2026
Input / output per million tokens$8 / $40$8 / $40
Cached input$0.40$0.20
Context window1,050,000 tokens1,000,000 tokens
Request above 272K tokensInput ×2 and output ×1.5 for the whole requestSame rate up to 1M tokens
Fast modeTwice the standard rate$16 / $80 (twice the standard rate)
Where you get it on a subscriptionChatGPT Work and Codex onlyAll paid Claude plans

Anthropic's long-context terms are on its long-context pricing page, which says a 900,000-token request is billed at the same rate as a 9,000-token one. The Fable 5.2 announcement compares the model with GPT-6 Astra, not with 6.2. It scores 69.9% against 65.5% on agentic coding (FrontierSWE v2), and 89.3% against 79.5% on computer use (OSWorld 2.1). Neither company has yet published a direct benchmark comparison of GPT-6.2 and Fable 5.2, so this article doesn't include one.

A worked example: when is the Astra bill higher?

These figures are simple arithmetic on the list prices, not measurements:

  • Short request (50,000 input tokens and 5,000 output tokens): $0.60 on Astra, $0.60 on Fable and $0.15 on Sol.
  • Long request (300,000 input tokens, such as a full contract with its annexes, and 10,000 output tokens): on Astra the request crosses the 272K line, so input becomes $16 per million and output $60. The total is $5.40. On Fable it stays at $2.80, roughly half.

The practical lesson: if your company sends long documents in a single request, the long-context difference matters more than the matching list price.

Which one should your business use?

  1. You subscribe to ChatGPT Plus or Business and want the newest model: open Work, not Chat, and choose GPT-6.2 from the model picker once the rollout reaches you.
  2. You build on the API with a tight budget: start with GPT-6.2 Sol. It costs the same as 6.1 Sol and is stronger on engineering tasks. Move to Astra only for tasks where Sol falls short.
  3. You send long documents, or run agents that reread the same context: price the job on Fable 5.2 as well. Long context costs no extra there, and cached input costs half what Astra charges.
  4. You handle sensitive data: Anthropic's what's-new page says Fable 5.2 requires 30-day data retention and isn't available under zero retention without special authorization. Check this against your personal-data-protection obligations before you send customer data.
  5. Before any switch: run the new model on 20 real tasks from your own work and compare quality and actual cost, not launch-day numbers.
  • #Artificial Intelligence
  • #ChatGPT 6.2
  • #GPT-6.2 Astra
  • #Claude Fable 5.2
  • #AI costs

Frequently asked questions

What is the difference between ChatGPT 6.2 and 6.1?

The 6.1 release was GPT-6.1 Sol only. With 6.2 came a new Sol at the same price ($2 and $10 per million tokens) with stronger engineering and security results, and a new Astra that costs 20% less than GPT-6 Astra, at $8 and $40.

Is GPT-6.2 available in regular ChatGPT chat?

No. OpenAI says GPT-6.2 Astra and Sol are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and that the rollout does not add them to regular Chat.

Which is cheaper: GPT-6.2 Astra or Claude Fable 5.2?

The list price is the same: $8 for input and $40 for output per million tokens. Fable 5.2 is cheaper for cached input ($0.20 against $0.40), and it does not raise the rate for requests above 272,000 tokens, as Astra does.

Is GPT-6.2 better than Claude Fable 5.2?

As of November 18, 2026, neither company has published a direct comparison of the two models. The safe approach is to test both on real tasks from your work and compare quality and cost.

Have a project in mind?

We build custom systems, apps and websites for your business. Tell us your idea and we will give you a straight answer on it.

See our services

Newsletter

New articles and series, in your inbox.

Follow Origami in Google

Pin Origami as a preferred source and our articles will surface first for you in Google Search and Top Stories.

Add as a preferred source on Google