GPT-6.2 Astra and Sol in the API: prices, the Ultrafast change in March, and what to fix before you switch
OpenAI's gpt-6.2-astra costs $8 input and $40 output per million tokens, 20% less than GPT-6 Astra; gpt-6.2-sol keeps $2/$10. Here are the full rate tiers, the Ultrafast price that doubles on March 1, 2027, and the API changes OpenAI asks you to make first.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. Astra is its frontier model for the hardest work; Sol is the cheaper everyday model. Both are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol.
For developers, three things matter more than the benchmark charts: what each tier costs, a scheduled Ultrafast price change, and a new API rule that throws away the model's earlier reasoning when you edit conversation history. This piece covers all three.
API prices per million tokens
Standard rates from OpenAI's pricing page:
- gpt-6.2-astra: $8 input, $0.40 cached input, $40 output. Cache writes: $10 (30-minute) or $20 (12-hour).
- gpt-6.2-sol: $2 input, $0.10 cached input, $10 output. Cache writes: $2.50 (30-minute) or $5 (12-hour).
- For comparison, gpt-6-astra: $10 input, $1 cached input, $50 output. Astra's input and output prices fall 20% and cached input 60%. Sol's rates are the same as GPT-6.1 Sol's.
- These Standard rates apply up to 272K input tokens; above that, GitHub's Copilot price list shows Astra at $16 / $0.80 / $60 and Sol at $4 / $0.20 / $15 (input / cached / output). Data-residency (regional processing) endpoints add 10% for models released on or after March 5, 2026.
The other tiers scale from Standard. Batch is half price (Astra $4 / $20, Sol $1 / $5). Fast is double (Astra $16 / $80, Sol $4 / $20). Ultrafast is three times Standard through February 28, 2027 (Astra $24 / $120, Sol $6 / $30) and six times Standard from March 1, 2027, Pacific time. From March 1 that means $48 / $240 for Astra and $12 / $60 for Sol, so if you build on Ultrafast now, plan for the increase. There is no Flex tier for GPT-6.2. Eligible background requests are 30% below Standard with a 2-hour completion window and 40% below with an 8-hour window.
Where you can use it
- ChatGPT Work and Codex: Plus, Pro, Enterprise and Edu. Eligible Enterprise and Edu users get it once a workspace admin enables it.
- Ultrafast in ChatGPT: Pro 500 subscribers with either model; Pro 200 subscribers get it for the first time, starting with GPT-6.2 Sol.
- API: gpt-6.2-astra and gpt-6.2-sol, with Ultrafast for both at launch. Standard, Fast and Batch support US and EU data residency; Ultrafast is US only.
- GitHub Copilot: Pro+, Max, Business and Enterprise, rolling out gradually, billed at provider list price. On Business and Enterprise, new models are on by default unless an admin has turned off the global default.
The rule that can raise your bill: don't edit the prefix
Starting with GPT-6.2, the API drops the model's earlier reasoning items when it detects a mid-conversation change to the conversation history, a model downgrade, or a change of end user or organization. OpenAI's docs say the model may then redo work, which costs more tokens and time, and answer quality can drop if it can't recover. A changed prefix also stops input after the changed point from being served from the prompt cache.
Common ways integrations trigger this, per OpenAI's guide: changing reasoning.effort, text.format or tools between requests; rewriting the developer message; summarizing or truncating old turns; shortening past tool outputs or downsizing past images. The fixes OpenAI lists:
- To change effort or the output schema mid-conversation, append a configuration_update item instead of changing the request parameter.
- To add tools, append an additional_tools item; to restrict them on a turn, use tool_choice with allowed_tools. Keep existing tool definitions in order.
- To change instructions, append a new developer message rather than editing the original.
- To shrink context, use OpenAI compaction (/responses/compact or context_management), after which reasoning items are kept. Truncate long tool outputs before you first send them, not afterwards.
- To find out what happened, check input_transformations in the response for reasoning items the API removed, and Prompt Cache Diagnostics for cache misses.
What a broken prefix costs, using Standard Astra rates and leaving out cache-write charges: a turn with 200K tokens of context, 180K of it cached, and 8K output costs about $0.55 ($0.16 new input, $0.07 cached input, $0.32 output). If an edit to an early message means none of the 200K is read from cache, the same turn costs about $1.92, roughly 3.5 times as much, before any extra reasoning the model has to redo. That is our arithmetic from OpenAI's listed prices.
Parameter changes when you migrate
- Set model to gpt-6.2-astra or gpt-6.2-sol.
- Reasoning effort: OpenAI suggests starting at medium; if you used none or minimal, use low.
- Tool calling needs the Responses API. Chat Completions still works for plain requests.
- Remove temperature, top_p, logprobs and top_logprobs from Chat Completions requests; in Responses, remove temperature, top_p and top_logprobs, and drop message.output_text.logprobs from include.
- Coming from GPT-5.5 or earlier: replace prompt_cache_retention with prompt_cache_options.ttl set to "30m", and review cache-write billing.
- Zero Data Retention: for approved organizations, store is treated as false, so pass prior output items back in order, including encrypted reasoning items.
OpenAI also notes behavior changes: GPT-6.2 Astra writes more concisely by default, so revisit 'be concise' instructions, and when intent is unclear it may propose a next step and wait for confirmation instead of carrying on.
What OpenAI claims, and what's self-reported
OpenAI says Astra scores 5.2 points above GPT-6 Astra on Agents' Last Exam at under a third of the cost, sets a new high on Terminal-Bench-Science 0.1, leads the models OpenAI tested on its internal GDPval v2 (6.2 points above GPT-6 Astra), and that Sol comes within 0.5 points of Astra on Cognition's FrontierCode 2.0 at about one-fifth of the cost (both at max effort on that benchmark). These are OpenAI's numbers and include each model's token efficiency, so per-task cost gaps are not the same as the per-token price gaps above. Like GPT-6 Astra, both models meet the Critical threshold for cybersecurity under OpenAI's Preparedness Framework.
Try it with your AI. Give this to your coding agent to audit an existing OpenAI integration before switching it to GPT-6.2:
Audit this codebase's OpenAI API calls before we switch to gpt-6.2-astra or gpt-6.2-sol.
Do not edit anything yet; report first.
1. Find every place that calls the OpenAI API and list: endpoint (Responses or Chat
Completions), model, and parameters.
2. Flag parameters GPT-6.2 does not accept: temperature, top_p, logprobs, top_logprobs
(Chat Completions); temperature, top_p, top_logprobs and message.output_text.logprobs
in include (Responses).
3. Flag tool calling that goes through Chat Completions; it needs the Responses API.
4. Flag anything that changes the conversation prefix between turns: changing
reasoning.effort, instructions, text.format or tools mid-conversation, editing
earlier input items, editing earlier developer messages, truncating or summarizing
earlier turns or tool outputs.
5. If migrating from GPT-5.5 or earlier, flag prompt_cache_retention (replace with
prompt_cache_options.ttl set to "30m").
For each finding give file, line, and the change you propose, then wait for my approval.
Do not print or move API keys.
Project context: [LANGUAGE / SDK VERSION / WHERE THE API CALLS LIVE]FAQ
How much does gpt-6.2-astra cost? $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens at Standard. gpt-6.2-sol costs $2, $0.10 and $10.
Is GPT-6.2 Sol cheaper than GPT-6.1 Sol? No. Its API rates are the same as GPT-6.1 Sol's; OpenAI says it gets more done per token.
Does GPT-6.2 support temperature? No. OpenAI's migration guide says to omit temperature and top_p for both models.
When does Ultrafast get more expensive? It is three times the Standard rate through February 28, 2027, and six times from March 1, 2027 (Pacific time).
Can I use GPT-6.2 in Codex and GitHub Copilot? Yes. Codex has it for Plus, Pro, Enterprise and Edu users. GitHub Copilot has it on Pro+, Max, Business and Enterprise, rolling out gradually.
Related on claudefox
- tstack: nine agent skills that take a change to a merged PR in Claude Code and Codex
- DeepSeek opens its app's search, file, voice and safety features to the API
- Claude Code product lead Cat Wu is joining OpenAI
- Claude gets a Personal Finance beta in the US
Source: OpenAI: Introducing GPT-6.2 Astra and Sol, OpenAI API docs: Pricing, OpenAI API docs: Using GPT-6.2, OpenAI API docs: Keep the prefix unchanged, GitHub changelog: GPT-6.2 Astra and GPT-6.2 Sol now available, GitHub Docs: Models and pricing for GitHub Copilot
This story first appeared in the claudefox daily briefing for November 18, 2026.