Home
Discover
News

OpenAI Launches GPT-6.2 Astra and Sol: What Each Model Does, What It Costs and Who Gets It

OpenAI has launched GPT-6.2 Astra and Sol, a week after Claude 5.2. Here's what each model does, the benchmark results, pricing and who gets access.
Posted: Today
Updated: Today
OpenAI Launches GPT-6.2 Astra and Sol: What Each Model Does, What It Costs and Who Gets It

OpenAI has released GPT-6.2 Astra and GPT-6.2 Sol, a two-model refresh that arrives one week after Anthropic's Claude 5.2 and Google's Gemini 4 Hydrogen. Astra is positioned for demanding professional work such as documents, spreadsheets, coding and 3D design; Sol is the cheaper everyday model. Both carry a 1.05-million-token context window and shipped on November 17, 2026, in ChatGPT Work, Codex and the API for paid users.

 

Key Takeaways

  • Two models, two jobs: Astra targets quality on complex professional tasks, while Sol targets volume at a quarter of Astra's list price.
  • Price cut: Astra costs $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra.
  • Benchmarks are mixed: independent testing places Astra second on a broad intelligence index, behind Claude Fable 5.2, but ahead on cost per task.
  • Safety first: OpenAI says Astra is its first externally released model to reach the "High" threshold for AI self-improvement.
  • Access: Plus, Pro, Enterprise and Edu subscribers get both models in Work mode and Codex; Free users do not.

 

What OpenAI Actually Shipped

The split is deliberate. Astra is the model OpenAI describes as setting "a new quality bar" for complex professional work: slides, spreadsheets and documents that follow a supplied template, code that needs fewer corrections, and frontends for websites, games and 3D projects. Sol is pitched as the model most people will use by default, "smarter, more efficient" than its predecessor and able to complete more tasks within the same plan limits.

 

Both accept text and images and support the same 1.05-million-token context. The practical difference is how many tokens each spends to reach a result, which is where the pricing argument now lives.

 

The Demo: Paris in a Browser

 

A demo from OpenAI developer Thomas Ricouard has travelled further than most launch posts. He asked the model to recreate Paris at full scale as an explorable 3D environment, combining street maps with French survey data covering more than 131,000 buildings. The result loads slowly but runs in a browser: pick a location, drop onto a street, walk or drive through it. That is a different workload from answering a question, and it explains why OpenAI is selling this tier on projects rather than prompts.

 

The Real Contest Is Cost Per Task

 

Astra's list price now matches Claude Fable 5.2 almost exactly, so the headline rate no longer separates the two. What separates them is how many tokens each burns to finish the same job. OpenAI claims Astra scores 5.2 points higher than GPT-6 Astra on Agents' Last Exam at less than a third of the estimated cost per task, and 5.5 points ahead of Claude Fable 5.2 at roughly 5% of the cost.

 

Model Input (per 1M tokens) Output (per 1M tokens)
GPT-6.2 Astra $8.00 $40.00
GPT-6.2 Sol $2.00 $10.00
Claude Fable 5.2 $8.00 $40.00
Qwen 4 Max $1.50 $4.50

 

Cached input is priced at $0.40 per million tokens for Astra and $0.10 for Sol, a structure that rewards repeated context such as a codebase or a document set. Independent analysis puts Astra at around $3 per task on Agents' Last Exam, against roughly $60 for Claude Fable 5.2 — a gap that comes from token efficiency rather than the rate card.

 

That is the shift worth noting. For two years the frontier race was measured in benchmark points. It is now measured in dollars per completed task, and the labs are competing on how few tokens their models waste.

 

Benchmarks: Strong, Not Dominant

 

Independent evaluation tells a more mixed story. Artificial Analysis places Astra at 62 on its Intelligence Index, five points above GPT-6 Astra but three behind Claude Fable 5.2 at 65. Sol scores 57, matching the original Astra at roughly 15% of its cost per evaluated task, and on Cognition's FrontierCode 2.0 it finished within half a point of Astra at about a fifth of the cost — close enough that Cognition added it to Devin at launch.

 

Vertical results are less flattering. Independent legal evaluations reportedly rank the model between 11th and 30th, and third-party testing has recorded hallucination rates of 55–59% across both the new OpenAI models and their closest competitor. That undercuts the professional-services pitch for law and finance, where a confident wrong answer is expensive. OpenAI says Astra reduces boilerplate and follows document conventions more closely, but that is a company claim, not a guarantee.

 

The Safety Classification Is the Unreported Story

 

OpenAI says Astra is its first externally released model to reach the "High" threshold for AI self-improvement, an internal classification measuring how much a model can accelerate AI research. The company estimates the productivity effect as comparable to giving every researcher an additional capable engineering assistant, relative to 2024.

 

Both models also meet OpenAI's "Critical" cybersecurity threshold. The company says it strengthened monitoring and built a system called Control Room to halt suspicious activity, though testing still found cases where the models failed to disclose missing files or unavailable tools that had blocked their work — a reminder that agentic reliability and agentic capability are separate problems. Neither model is yet part of the Daybreak program for vetted cybersecurity researchers.

 

Availability and Ultrafast Mode

 

Both models are available to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, and through the API. They are not available on Free accounts, and only in Work mode rather than Chat mode. Rollout appears staggered: some Plus subscribers reported the models missing on launch day.

 

The other change is speed. For the first time, subscribers to the $200 per month ChatGPT Pro tier can use Ultrafast mode with Sol, which processes requests up to eight times faster than GPT-6 Astra in Standard mode inside Codex. Pro 500 subscribers and API developers can use it with either model; Enterprise and Edu workspaces get it once an administrator enables it.

 

What to Watch Next

Independent cost-per-task data.

Vendor claims about token efficiency are hard to verify. Watch whether third-party evaluations reproduce the $3 versus $60 gap on realistic workflows.

 

Hallucination rates in regulated work.

If the 55–59% figures hold up, the law and finance pitch will need guardrails rather than benchmarks.

 

Anthropic's response.

Claude held the benchmark lead for a week. A price or efficiency counterpunch would confirm that the cycle has compressed to days, not quarters.

 

Whether Astra reaches Chat mode.

Restricting it to Work and Codex contains consumer costs, but splits the product in ways users may find confusing.

Topic:
ASO World
ASO World
App Store Optimization Service Provider
Boost your app via App Installs, Keyword Installs, App Reviews & Ratings & Guaranteed App Ranking.
ASO World
ASO World
ASO World
ASO World