metir
metir
Docs
Download on App StoreGet it on Google PlayLog inSign up
Back to Blog
GPT-6.2 Astra
GPT-6.2 Sol
OpenAI
AI Model Pricing
AI Benchmarks 2026

GPT-6.2 Astra and Sol: Release, Benchmarks and Pricing

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. Astra's API price is 20% lower; Sol's is unchanged. Benchmarks, pricing and migration notes.

Metir AI TeamNovember 17, 20266 min read
GPT-6.2 Astra and Sol: Release, Benchmarks and Pricing

On November 17, 2026, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, two models that update its flagship and its everyday workhorse together. OpenAI's community announcement describes Astra as "a frontier model built for your hardest projects" and Sol as an efficient model for everyday work and multistep agentic workflows. The headline for GPT-6.2 Astra pricing is a 20% cut in standard API rates, while Sol keeps the price it had on GPT-6.1 Sol.

The OpenAI claims below come from OpenAI's launch post as relayed in its community announcement, its developer documentation and press coverage of OpenAI's launch charts. Figures labelled "OpenAI says" are OpenAI's own reported numbers, not independent tests.

$8 / $40GPT-6.2 Astra per million input / output tokens
$2 / $10GPT-6.2 Sol per million input / output tokens
1.05MToken context window on both models
20%Astra's standard price cut vs GPT-6 Astra
OpenAI logoOpenAI
Anthropic logoAnthropic
Google logoGoogle
GPT-6.2 is an OpenAI release; launch coverage compares it with Anthropic's Claude Fable 5.2 and notes Google's Gemini 4 Hydrogen shipped the same week.

GPT-6.2 Astra and Sol pricing

Both models are available in the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI's model pages list the same shape for each: a 1,050,000-token context window (922,000 tokens maximum input), up to 128,000 output tokens, text and image input and text output. The knowledge cutoff is May 18, 2026 for Astra and June 1, 2026 for Sol.

ModelInputCached inputCache write (30 min)Output
GPT-6.2 Astra$8.00$0.40$10.00$40.00
GPT-6 Astra (previous)$10.00$1.00$12.50$50.00
GPT-6.2 Sol$2.00$0.10$2.50$10.00

USD per million tokens, standard rates. Sources: OpenAI's GPT-6.2 Astra and GPT-6.2 Sol model pages, and 9to5Mac, which reports Astra's standard input and output prices are 20% lower than GPT-6 Astra's and that Sol keeps its GPT-6.1 pricing. The GPT-6 Astra row is from the earlier pricing we covered in our GPT-6.1 Sol post.

Several modifiers apply to both models, per OpenAI's model pages:

  • Long prompts: above 272K input tokens, input and cache rates double and output costs 1.5x, for the whole request.
  • Batch and speed tiers: Batch is 50% of Standard, Fast mode is 2x, and Ultrafast is 3x through February 28, 2027, then 6x from March 1, 2027.
  • Background windows: OpenAI's API changelog says the Responses API public beta added two-hour and eight-hour background completion windows on November 17, with token rates 30% and 40% lower than Standard.

A rough worked example, using only the list prices above: a request with 50,000 input tokens and 5,000 output tokens costs about $0.15 on Sol and about $0.60 on Astra, the same arithmetic Digital Applied uses. Real bills depend on caching and how many reasoning tokens a model spends.

“

Measure cost per accepted result, including failed runs and reviewer time, rather than comparing token prices alone.

Digital Applied, on comparing GPT-6.2 with rival models

What OpenAI says about GPT-6.2 benchmarks

The numbers below are OpenAI-reported. They come from its community announcement, 9to5Mac and Digital Applied's summaries of OpenAI's launch charts.

BenchmarkOpenAI-reported result
FrontierCode 2.0 (Cognition; whether changes to real codebases are good enough for maintainers to merge), max effortSol within half a point of Astra at about one-fifth the cost
Agents' Last ExamAstra at xhigh scores 5.2 points above GPT-6 Astra at max, at less than half the cost per task
Agents' Last Exam, vs Anthropic's Claude FableAstra ahead by 5.5 points at roughly 5% of Fable's cost; nine tasks where Fable used a fallback model were excluded
ExploitBench Internal PortAstra up 22 points over GPT-6 Astra
LifeValNew Sol beat GPT-6 Astra at about one-seventh the evaluation cost
AA-AnalystAgentAstra 62.5%, counting only questions answered correctly on all five attempts
AA-3D (Elo, vote-based)Astra 1,318, Fable 1,271, Sol 1,259

Sources: OpenAI community announcement, 9to5Mac, Digital Applied and its Fable 5.2 comparison.

Two details deserve care. The AA-AnalystAgent figure is a strict all-five-attempts score, so it says little about average single-attempt accuracy. The AA-3D numbers are Elo ratings from votes, not percentages of quality. 9to5Mac also relays OpenAI's account of early testers reporting "cleaner code style and more focused tests, with fewer rounds of revision."

~1/5Cost of Sol vs Astra on FrontierCode 2.0 (OpenAI-reported)
0.5 ptSol's gap to Astra on that test at max effort
+5.2 ptsAstra xhigh over GPT-6 Astra max on Agents' Last Exam
62.5%Astra on AA-AnalystAgent, all five attempts correct

Where GPT-6.2 is available

Per OpenAI's community post and 9to5Mac, both models are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API. Ultrafast, which 9to5Mac says can generate tokens at up to 4.5 times the standard rate while using more of a plan's allowance, comes to the $200 Pro plan for the first time with Sol. The $500 Pro plan gets it on both models, and eligible Enterprise and Edu workspaces get it once an admin enables it. Our Ultrafast explainer covers how that tier works.

Screenshot of OpenAI's Codex desktop app beside a browser window
OpenAI's Codex desktop app, one of the places GPT-6.2 is available. This July 2026 screenshot shows a GPT-5.6-generation model in the composer, not GPT-6.2. Image: Wikideas1, Wikimedia Commons, CC0.

Safeguards and a release that follows a crowded week

OpenAI's community post says that, because of the models' higher capabilities, they ship with tighter safeguards for higher-risk cybersecurity help, and that OpenAI is working to bring them to Daybreak so approved defenders can get reduced refusals. Some secondary summaries claimed more specific Preparedness-threshold language. We could not confirm that wording in any source we read, so we do not repeat it. Digital Applied separately reports OpenAI found both models more likely to skip an obstacle such as a missing file, and that new Codex instructions reduced this.

9to5Google places the release about a week after Anthropic's Claude 5.2 and Google's Gemini 4 Hydrogen, and says OpenAI describes the models as roughly on par with Anthropic's latest at often much lower cost. Anthropic, for its part, claims its Fable 5.2 beats rival models across most benchmarks. These are competing vendor claims. Our posts on Claude Fable 5.2 and Gemini 4 Hydrogen cover the other launches.

Migration notes for developers

OpenAI's Using GPT-6 guide and model pages list the practical changes:

  • Reasoning effort: the guide suggests starting new integrations at medium and moving from none or minimal to low.
  • APIs: tool calling requires the Responses API. Chat Completions works without tools.
  • Reasoning items: starting with GPT-6.2, the API drops reasoning items when it detects prefix edits to the conversation history, a model downgrade, or a change of end user or organization. That can reduce answer quality and raise token use, so keep the prefix stable.
  • Style: Astra is more concise and direct by default, so revisit "be concise" instructions.
  • Unsupported parameters: temperature, top_p and log-probability options are removed, and fine-tuning and predicted outputs are not supported.
  • Data: Zero Data Retention is available to approved organizations, and US and EU data residency covers Standard, Fast and Batch. Ultrafast is US-only.

What the evidence does and does not show

  • Cost per task is a harness result. OpenAI's cost-per-task claims depend on how many tokens each model spends, so list price alone does not predict them.
  • Sol versus Astra is task-dependent. A half-point gap on FrontierCode 2.0 does not imply the same gap on every workload, and the 4x list-price difference matters most where a mistake is expensive to find later.
  • Independent checks are pending. Every figure above is OpenAI-reported on launch day, and the official launch post was not readable to us.

GPT-6.2 on Metir

GPT-6.2 Sol and GPT-6.2 Astra are now the models behind Metir's OpenAI high-tier picks, so you can compare them against Claude and Gemini models in the same chat.

Sources:

  • GPT-6.2 Astra and Sol are now available - OpenAI Developer Community
  • GPT-6.2 Sol model page - OpenAI API docs
  • GPT-6.2 Astra model page - OpenAI API docs
  • Using GPT-6 guide - OpenAI API docs
  • API changelog - OpenAI API docs
  • OpenAI's launch post, introducing GPT-6.2 Astra and Sol (primary source)
  • OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models - 9to5Mac
  • OpenAI releases GPT-6.2 Astra and Sol for complex projects - 9to5Google
  • GPT-6.2 Astra and Sol release, pricing and access - Digital Applied
  • GPT-6.2 Astra and Sol vs Claude Fable 5.2 and Opus 5.5 - Digital Applied

Image credits

Header image: the OpenAI website seen through a magnifying glass, a 2023 stock-style photo by Jernej Furman, Wikimedia Commons, CC BY 2.0 (source); it is not an image of the GPT-6.2 announcement. In-body image: screenshot of the OpenAI Codex desktop app by Wikideas1, Wikimedia Commons, CC0 (source); it shows a GPT-5.6-generation model, not GPT-6.2.

Ready to experience AI that adapts to you?

metir brings together the world's best AI models in one seamless experience. Start for free today.

Get Started Free
metir

metir is where work gets done. Hire your team of Orbs, AI teammates with their own computer, alongside your mail, meetings and workspace.

© 2026 metir. All rights reserved.

Product

  • Features
  • Pricing
  • Research
  • Docs
  • Blog
  • Enterprise

Company

  • Docs
  • Support
  • Careers

Legal

  • Terms of service
  • Privacy policy

Personalisation is powerful. Privacy is non-negotiable.

Status: All systems operational