Skip to content
What is GPT-6.2? OpenAI's Astra and Sol, and which to use
Back to Blog
#AI models#GPT-6.2#GPT-6.2 Astra#GPT-6.2 Sol#GPT-6.2 pricing#gpt-6.2-astra API

What is GPT-6.2? OpenAI's Astra and Sol, and which to use

Amir Khan

GPT-6.2 is OpenAI's new pair of models, released on 17 November 2026: GPT-6.2 Astra for the hardest coding and professional work, and GPT-6.2 Sol for everyday work and agents that run at volume. Both are in the API now as gpt-6.2-astra and gpt-6.2-sol, and in ChatGPT Work and Codex. Astra costs $8 per million input tokens and $40 per million output tokens; Sol costs $2 and $10, a quarter of Astra at every rate.

If you are choosing for a product, the practical question is which of the two each job needs, and what changes in your code when you switch.

Astra or Sol: which one does your job need?

OpenAI splits them by difficulty. Its release notes describe Astra as built "for complex coding and professional work", such as tracing a bug across services and testing the fix. Sol "handles tasks that combine text, images and tools such as web search". Both read text and images and write text, and both work through the Responses and Chat Completions APIs, with Batch and Flex processing.

GPT-6.2 AstraGPT-6.2 Sol
OpenAI's descriptionMost capable model, for your hardest projectsGo-to model for everyday work, coding and agents at scale
Input, per 1M tokens$8.00$2.00
Cached input, per 1M tokens$0.40$0.10
Output, per 1M tokens$40.00$10.00
API model idgpt-6.2-astragpt-6.2-sol

Prices are OpenAI's standard rates for requests under 272K tokens of context, read on 18 November 2026. Batch processing costs half; US or EU data residency adds 10%.

Sol's model page gives the rest of its specification: a 1,050,000-token context window, up to 128,000 output tokens, and a knowledge cutoff of 1 June 2026. It supports web search, file search, code interpreter, computer use and MCP tools in the Responses API. Above 272K input tokens, Sol's input and cache rates double and output costs 1.5 times as much, for the whole request. Check Astra's own model page for its context window and output limit before you plan long-document jobs.

Our view: run your own test set on Sol first, and move to Astra only the tasks where Sol's answers are not good enough. At four times the price, Astra has to save real rework to pay for itself.

What does a job cost?

Take a code review: you send 40,000 tokens of diff and surrounding code, and the model returns 4,000 tokens of reasoning and comments.

  • Sol: 0.04 × $2 + 0.004 × $10 = $0.08 + $0.04 = $0.12
  • Astra: 0.04 × $8 + 0.004 × $40 = $0.32 + $0.16 = $0.48

A thousand reviews a month is $120 on Sol or $480 on Astra. Output length varies with the reasoning effort you set, so measure token counts on your own jobs before you quote a price.

What changes when you switch your code?

More than the model name. OpenAI's "Using GPT-6.2" guide lists what to check:

  • Tool calling needs the Responses API. Both models accept Chat Completions requests, but tool calls only work through Responses.
  • Drop sampling parameters. Remove temperature, top_p, logprobs and top_logprobs.
  • Reasoning effort starts at medium. If you used none or minimal, use low; Sol's API rejects none.
  • Move fixed instructions into a developer message at the start of input. From GPT-6.2, top-level instructions come after the input, so they no longer count toward prompt caching.
  • Do not edit earlier turns. Changing earlier messages or tool calls mid-conversation makes the Responses API drop the model's reasoning. The guide warns that answers can get worse and use more tokens as a result.

Both models support Zero Data Retention for approved organisations, and US and EU data residency. In ChatGPT, you pick Astra or Sol in the Work model picker; OpenAI's release notes say neither has been added to regular Chat.

"GPT-6.2 Astra follows your template more closely when turning source documents into slides." OpenAI, ChatGPT release notes

Want GPT-6.2 built into your product or workflow, with Sol and Astra tested on your own jobs? See what we build or tell us about the job. For OpenAI's cheaper model for images and screens, read What is GPT-6.1 Terra?

Sources