OpenAI GPT-6.2 Astra and Sol: Prices, Limits, Background Windows
OpenAI added two models to its API on 17 November 2026. The changelog says to "use GPT-6.2 Astra for demanding reasoning and coding tasks, or GPT-6.2 Sol for everyday coding and agent workflows." On the same day it opened a background tier in the Responses API that trades latency for a lower token price, and GitHub put both models into Copilot.
What happened
Both models are on the Responses API and Chat Completions, but "tool calling requires the Responses API", so agent code still on Chat Completions cannot use tools with them. Here are the published numbers from the Astra and Sol model pages:
gpt-6.2-astra |
gpt-6.2-sol |
|
|---|---|---|
| Input / cached input / output (per 1M) | $8.00 / $0.40 / $40.00 | $2.00 / $0.10 / $10.00 |
| Cache write (per 1M) | $10.00 | $2.50 |
| Over 272K input tokens | input 2x, output 1.5x | input 2x, output 1.5x |
| Context / max input / max output | 1,050,000 / 922,000 / 128,000 | 1,050,000 / 922,000 / 128,000 |
| Knowledge cutoff | 18 May 2026 | 1 Jun 2026 |
| Reasoning effort | low to max, none rejected |
low to max, none rejected |
Compared with what they replace on the pricing tracker: Astra is cheaper than gpt-6-astra ($10 input, $1 cached, $50 output), and cached input falls by more than half. Sol costs exactly what gpt-6.1-sol costs. OpenAI's changelog does not deprecate either older model.
Background completion windows are in public beta in the Responses API. You choose a two-hour or eight-hour window, and token rates are 30% or 40% lower than Standard. They apply to gpt-6.2-astra, gpt-6.2-sol, gpt-6.1-terra and gpt-6-luna. The model pages also list Batch at 50% of Standard, so this new tier sits between Standard and Batch.
In GitHub Copilot, both models are generally available on Pro+, Max, Business and Enterprise, though not on Free or Pro. They are billed "at provider list pricing under usage-based billing". New models are switched on automatically unless an admin has turned off the global default. GitHub's framing differs slightly from OpenAI's: it positions Sol for "complex agentic coding and in-depth code review". The same day GitHub added Microsoft's MAI Code 2 and MAI Code 2 Flash and set 18 December 2026 as the date MAI-Code-1.1-Flash is removed from every Copilot experience.
What changes for engineers
- Astra-class work got cheaper, but only if you switch. Nothing moves automatically. Calls naming
gpt-6-astrakeep the old price. - The migration is not just a model string. OpenAI links a migration guide covering changes to how reasoning is preserved when you edit a conversation prefix, and to the top-level
instructionsfield. Agents that rewrite history or inject system text are the ones to test. reasoning_effort: "none"is rejected. Code that sets it to save latency will error on both models.- Long prompts still double. Past 272K input tokens the input rate doubles and output goes up 1.5x, so a 500K-token context on Astra is billed at $16 input and $60 output.
- Non-interactive jobs have a new lever. Evals, nightly summaries and code-review backlogs can move to the eight-hour window for 40% off, with a stated upper bound on how long they wait.
A cost example
This is arithmetic from list prices, not a measurement. Take one agent turn with 200K input tokens, half of them cached, and 20K output tokens:
| Model | Uncached input | Cached input | Output | Total |
|---|---|---|---|---|
gpt-6-astra |
$1.00 | $0.10 | $1.00 | $2.10 |
gpt-6.2-astra |
$0.80 | $0.04 | $0.80 | $1.64 |
gpt-6.2-astra, 8-hour window |
$0.98 | |||
gpt-6.2-sol |
$0.20 | $0.01 | $0.20 | $0.41 |
The background line assumes the 40% discount covers all three token types. OpenAI's changelog says "token rates" without splitting them out, so check your first invoice.
What I ran
Nothing. I have no OpenAI API key on this machine, so there are no latency or quality numbers here. When I run these models on my fixed task set I will publish the results as a separate post.
What to do this week
- Run your eval set on
gpt-6.2-astrabefore switching anything fromgpt-6-astra. Pay attention to tool calls and to conversations where you edit earlier turns. - Search your code for
reasoning_effortset tononeand for tool calls on Chat Completions. - Move one batch-like job to a background window and compare cost and completion time against Batch.
- Copilot admins: decide whether the auto-enable default is what you want, and enable the newer MAI Code Flash model before 18 December if people use MAI-Code-1.1-Flash.
- Update your cost model. The LLM cost calculator now has both models.
The AI model release tracker and the LLM API pricing tracker are updated with today's rows.
FAQ
How much does GPT-6.2 Astra cost?
$8 per million input tokens, $0.40 cached input and $40 output on the Standard tier, doubling on input and rising 1.5x on output above 272K input tokens. Background windows take 30% or 40% off, and Batch takes 50% off.
What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?
OpenAI positions Astra for demanding reasoning and coding, and Sol for everyday coding and agent workflows. Both have a 1,050,000-token context and 128,000 max output. Sol costs a quarter of Astra's price.
Is GPT-6.2 in GitHub Copilot?
Yes. Both are generally available on Copilot Pro+, Max, Business and Enterprise from 17 November 2026, billed at provider list price under usage-based billing.
Related: OpenAI GPT-6.1 Terra, How to Cut LLM API Costs With Caching and Batching and GitHub Copilot model deprecations in December.
Spot an error or something out of date? Email me and I will fix it and log it on the corrections page.
I'm Ajay Kumar — I build and operate PandaStack, an open-source Firecracker microVM cloud for AI agents, and write about the infrastructure and tooling around it.
Need this kind of infrastructure work? See what I do or email hello@ajayk.sh.
Related
Gemini 4 Neon Is GA at $3/$15, and Lands in GitHub Copilot Same Day
Google released gemini-4-neon on 18 November 2026 at $3 input and $15 output per 1M tokens, with a 1M context window and four thinking levels. GitHub put it in Copilot the same day. Where it sits against Carbon, Hydrogen and the rest.
5 minNov 13, 2026NEAR AI Breached Through an SGLang RCE: No CVE, No Upstream Fix Yet
NEAR AI says an attacker used an unpatched remote code execution flaw in SGLang on 8 November 2026 to reach ~100 users' account data, 18 legacy agents and requests running in its TEE. No CVE or SGLang advisory exists. What SGLang operators should do now.
6 minNov 10, 2026Alphabet Q3 2026: $51.4B Capex, Cloud Up 103%, Free Cash Flow at -$9.1B
Alphabet's Q3 2026 results: $51.4 billion of quarterly capex, Google Cloud revenue up 103% to $30.7 billion, negative free cash flow of $9.1 billion, an $802 billion commitments book and new debt and equity to pay for it.
7 min