GPT-6.2 Astra got cheaper and Sol kept its price. For coding agents, the coding results are mixed, so test before you switch

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on November 17 in ChatGPT, Codex, the API, and GitHub Copilot. Astra costs 20% less than GPT-6 Astra, and Sol keeps GPT-6.1 Sol's price. Move Sol-based agents over and compare results. Watch the reasoning-effort change and the Ultrafast price jump in March.

OpenAI's launch post positions Astra for the hardest projects and Sol for everyday coding and multistep work. The API IDs are gpt-6.2-astra and gpt-6.2-sol. Copilot added both the same day, per GitHub's changelog.

We haven't run either model on our own tasks yet. The figures below are OpenAI's, GitHub's, or independent evaluations as reported, plus our own cost arithmetic, and we say which is which.

Where you can use them

SurfaceWho gets itNotes
ChatGPT Work and CodexPlus, Pro, Enterprise, EduUltrafast for both on Pro $500; Sol only on Pro $200; Enterprise and Edu once an admin enables it
OpenAI APIAll developersUltrafast available for both at launch
GitHub CopilotPro+, Max, Business, EnterpriseGradual rollout; billed at provider list price under usage-based billing

Copilot's changelog doesn't list Pro or Free. On Business and Enterprise, the models turn on automatically under default model enablement unless an admin disabled the global default or these models.

What they cost

Per million tokens, from the Astra and Sol model pages, with older rows from OpenAI's pricing page and Anthropic's Opus 5.5 page:

ModelInputCached inputOutput
GPT-6.2 Astra$8.00$0.40$40.00
GPT-6 Astra$10.00$1.00$50.00
GPT-6.2 Sol$2.00$0.10$10.00
GPT-6.1 Sol$2.00$0.10$10.00
GPT-6.1 Terra$1.00$0.05$5.00
Claude Opus 5.5$4.00$0.20$20.00

Cached input on Astra dropped from $1.00 to $0.40, a 60% cut. Agent loops reread the same context constantly, so that matters more than the headline 20%.

Our arithmetic for one uncached agent step with 200,000 input tokens and 20,000 output tokens:

  • GPT-6.2 Astra: $1.60 + $0.80 = $2.40, down from $3.00 on GPT-6 Astra
  • GPT-6.2 Sol: $0.40 + $0.20 = $0.60, unchanged
  • Claude Opus 5.5: $0.80 + $0.40 = $1.20

OpenAI also says both models use fewer output tokens per task. On GDPval v2, Astra used 22% fewer output tokens than GPT-6 Astra. That isn't a coding benchmark, so treat it as a hint and measure your own runs.

What the coding results say

The coding signal is weaker than the launch framing suggests. Here is what's available:

ResultSourceWhat it shows
Sol within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the costOpenAIFor merge-quality changes, Sol is close to Astra
Astra 9 points above GPT-6 Astra on Terminal-Bench-Science 0.1OpenAIA gain on terminal tasks, no absolute score given
Sol caught nearly twice as many real issues as GPT-6.1 Sol in code reviewGitHub, early testingReview passes may improve at the same price
Astra flat against GPT-6 Astra; Sol down 3 points from GPT-6.1 Sol on the Coding Agent IndexArtificial Analysis, as reported by The DecoderNo clear coding-agent gain

OpenAI's post doesn't report SWE-bench. Its Terminal-Bench footnote also compares Anthropic's own Fable 5.2 figure of 74.6% in Claude Code with a different number measured in another harness, so don't read the cross-vendor comparison literally. Our take: this release is a price cut on Astra and a possible review upgrade on Sol, not a clear jump in agent coding.

Catches before you switch

  • No none reasoning effort. Both model pages say the API rejects requests that set reasoning effort to none. If your agent uses that setting for quick tool-routing calls, those calls will fail. Route them to a cheaper model or set a low effort level.
  • The 272K cliff. Above 272,000 input tokens, input and cache rates double and output costs 1.5x, for the whole request. The context window is 1,050,000 tokens, with 922,000 max input, but long requests cost much more.
  • Ultrafast doubles in March. Ultrafast costs 3x standard through February 28, 2027, then 6x from March 1. An Astra step from our example would go from $7.20 to $14.40 in Ultrafast. Budget for it now if you turn it on.
  • Tighter cyber safeguards. OpenAI says the launch includes stricter safeguards on higher-risk security help, which can block some legitimate vulnerability research. The GPT-6.2 models aren't on the Daybreak programs for vetted defenders yet.
  • Knowledge cutoffs. Astra's is May 18, 2026 and Sol's is June 1, 2026. Neither knows about libraries released since.

What we'd do

Sol users: switch your Codex or API default to gpt-6.2-sol and rerun a week of review and agent tasks. It's the same price, and GitHub's review result is the most relevant signal for coding. Keep GPT-6.1 Sol pinned in one job as a control until you see your own numbers.

Astra users: move to 6.2 for the lower price, especially on cache-heavy loops. If you chose Astra for hard multi-file changes, check whether 6.2 Sol is close enough, since FrontierCode puts it within half a point at a fifth of the cost.

Everyone else: Opus 5.5 is still half the price of Astra per token. Our GPT-6.1 Terra guide covers the cheaper tier for triage and subagents.

Sources

Researched and drafted by an AI agent from the linked sources. How posts are made.