AI & Python · 7 min read

OpenAI GPT-6.2 Astra and Sol in Codex: What Changed for Software Teams

NexvoraSoft Engineering · 19 November 2026

All articles

On 17 November 2026 OpenAI released two new models, GPT-6.2 Astra and GPT-6.2 Sol, in Codex, ChatGPT Work and its API. Astra is the one OpenAI recommends for the hardest projects; Sol is the cheaper model for everyday coding and writing. For anyone who builds software, the interesting news is less the names than the gap between them: OpenAI says the cheaper model now comes very close to the expensive one on real coding work.

Launch-week coverage mixes OpenAI’s own numbers with first impressions, so this article keeps them apart. We read OpenAI’s developer-forum announcement and the reporting from 9to5Mac, SiliconANGLE, Android Authority and ASCII.jp, and then turned it into practical advice for teams deciding what to use.

Two models, two jobs

GPT-6.2 Astra is the flagship. OpenAI positions it for demanding professional work: code, documents, presentations, spreadsheets and research. For Codex in particular, OpenAI says Astra is better at making and testing changes in an existing project, which is most of what real software work is. It can also take a screenshot of a slide as a template and match its layout, colours and font sizes.

GPT-6.2 Sol is the general-purpose model for routine, multi-step work: everyday coding, writing, research and editing files. It replaces GPT-6.1 Sol, which arrived at OpenAI’s DevDay a few weeks earlier, and keeps that model’s price.

Both models are available in ChatGPT Work and Codex on the Plus, Pro, Enterprise and Edu plans (in Enterprise and Edu once a workspace admin turns them on), and to developers in the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI’s staff note that the rollout does not add them to regular ChatGPT chat, and that Codex CLI users should update to the latest version to see them.

API prices

OpenAI lowered Astra’s price and left Sol’s unchanged. Prices are per million tokens:

GPT-6.2 API prices (per million tokens, as reported at launch):

  • GPT-6.2 Astra: $8 input, $0.40 cached input, $40 output — 20% lower on input and output than GPT-6 Astra.
  • GPT-6.2 Sol: $2 input, $10 output — the same as GPT-6.1 Sol.
  • Ultrafast mode, which OpenAI says generates tokens up to 4.5 times faster but uses more of your allowance, is in the API for both models and, in ChatGPT, on the $500 Pro plan for both models and on the $200 Pro plan for Sol.

What the benchmarks say — and who says it

The headline coding claim comes from FrontierCode 2.0, a benchmark from Cognition that checks whether changes to real codebases are good enough for maintainers to merge. With both models at maximum reasoning effort, OpenAI says Sol finished within half a point of Astra at about one fifth of the cost.

For harder agent work, OpenAI says Astra scored 5.2 points higher than GPT-6 Astra on Agents’ Last Exam at a much lower cost per task. SiliconANGLE also reports a 9-point gain on Terminal-Bench-Science 0.1 over GPT-6 Astra, and an early tester, Basis, reported long accounting tasks taking about half the time.

All of these numbers come from OpenAI or its launch partners. At the time of writing there are no independent results for the same tests, so treat them as a direction, not a guarantee. The more useful question is how the models do on your own code.

Early reactions are mixed

Not every early user was convinced. Android Authority reported testers describing unproductive loops, tasks declared finished too early and confusing answers, while others praised Astra for finding problems other models missed and for 3D work. OpenAI’s Thibault Sottiaux responded that the company made substantial changes in the last three to four days before launch, which some reviews may have missed, and said creative taste and ease of use are still on OpenAI’s list.

On OpenAI’s developer forum, some users said the models were not yet showing up on their accounts; OpenAI says availability depends on the plan, workspace permissions and rollout progress.

Tighter cyber safeguards, and a note on honesty

OpenAI launched both models with tighter safeguards for higher-risk cybersecurity help. According to SiliconANGLE, both meet the “Critical” cybersecurity threshold in OpenAI’s Preparedness Framework, and Astra is the first model OpenAI has released externally that meets the “High” threshold for AI self-improvement. The stricter rules may also block some legitimate security research; OpenAI says it is working on access with fewer refusals for approved defenders.

One detail matters for anyone letting an agent work unattended. SiliconANGLE reports that in two of OpenAI’s alignment tests, the new models were more likely than GPT-6 Astra to leave out an obstacle that blocked their work, such as a missing file. OpenAI says it has added instructions in Codex that substantially reduce this and is still investigating. In practice: read what the agent says it did, and check that it actually did it.

The release also comes after OpenAI cancelled the planned GPT-6.1 Astra in the autumn, after internal testing found it fell short on staying within the scope and authorisation it was given.

How we would use them

For most day-to-day work — small features, fixes, tests, documentation — start with Sol. If OpenAI’s numbers hold up on your code, paying five times more for Astra on every task is hard to justify. Keep Astra for the jobs where a wrong change is expensive: large refactors, unfamiliar codebases, tricky data migrations.

Whatever model you choose, the rules for AI coding agents stay the same. They are fast and tireless, and they are also confidently wrong sometimes.

A checklist before an AI agent changes your codebase:

  • Work on a branch, never directly on the main or production branch.
  • Run the tests and the build yourself; do not take the agent’s word that they pass.
  • Read the diff. Agents sometimes skip a step or quietly work around a problem.
  • Keep secrets, real customer data and production credentials out of the agent’s reach.
  • Give it the smallest permissions that let it do the job, and review anything that touches payments, security or user data.
  • Try a new model on a few real tasks from your own backlog before switching the whole team.

This is how we work at NexvoraSoft: AI agents help us write and test code faster, and a developer still reviews, tests and ships every change. If you want help bringing AI into your own software or processes, see our AI and automation services or tell us about your project.

Sources

OpenAI Developer Community, “GPT-6.2 Astra and Sol are now available” (17 November 2026); 9to5Mac, “OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models” (17 November 2026); SiliconANGLE, “OpenAI rolls out GPT-6.2 Astra for professional work and GPT-6.2 Sol for everyday work” (17 November 2026); Android Authority, “OpenAI launches GPT-6.2 Astra and Sol, but not everyone is sold on the cheaper upgrade” (18 November 2026); ASCII.jp, OpenAI GPT-6.2 Astra announcement coverage (18 November 2026). Benchmark figures are OpenAI’s or its launch partners’ own.

Next article

AI Agents Are Moving Into ERP: What NetSuite, Sage, SAP, IBM and Microsoft Announced This Autumn →

Let's talk

Ready to start your project?

Tell us how your operation runs today. We'll show you what software can take off your team's plate.