AI News6 min read

GPT-6.2 Astra and Sol are here: what they mean for agent work

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Nov 17 2026. Astra is its new frontier model for the hardest projects and Sol its model for everyday work. Astra's API price falls to $8 per million input and $40 per million output tokens; Sol keeps GPT-6.1 Sol's rates.

GPT-6.2 Astra and GPT-6.2 Sol release from OpenAI, set over an illustrated background

At a glance

Access: OpenAI says that starting today, GPT-6.2 Astra and Sol "are available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex." Developers can use gpt-6.2-astra and gpt-6.2-sol through the OpenAI API.

Pricing: GPT-6.2 Astra costs $8 per million input tokens, $0.40 per million cached input tokens and $40 per million output tokens. GPT-6.2 Sol keeps GPT-6.1 Sol's rates of $2 for input, $0.10 for cached input and $10 for output per million tokens.

Official announcement: Introducing GPT-6.2 Astra and Sol

What shipped

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Nov 17 2026. GPT-6.2 Astra is its new frontier model for the hardest projects, and GPT-6.2 Sol is its efficient, versatile model for everyday work. OpenAI says both deliver better results on demanding tasks with fewer output tokens, and that Astra's input and output token prices are 20% lower than GPT-6 Astra's, with a 60% discount on cached input.

Key capabilities

  • Professional documents. OpenAI says it trained GPT-6.2 Astra to create clearer, better-designed slides, spreadsheets and documents that follow the conventions of fields such as law, sales, finance and accounting. It calls Astra its best model yet at designing slides and following slide templates, from an original file or even a screenshot.
  • Long, multistep work. On Agents' Last Exam, which tests multistep professional work, OpenAI says GPT-6.2 Astra sets a new high, scoring 5.2 points above GPT-6 Astra at less than a third of the cost.
  • Judgment about tools and effort. OpenAI says Astra is better at matching its tools and effort to the workflow, including knowing when to use plugins versus computer use. On coding, it is better at implementing and testing complex changes in existing codebases, and early testers reported cleaner code style, more focused tests and fewer rounds of revision.
  • More precise computer use. OpenAI says Astra can work down to individual parts and connections in CAD and simulation tools.
  • GPT-6.2 Sol for everyday work. Compared with GPT-6.1 Sol, OpenAI says GPT-6.2 Sol is smarter, more efficient and lets you complete more tasks on the same paid plan. It describes Sol as powering through multistep agentic workflows.
  • Tighter safeguards. Both models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities under OpenAI's Preparedness Framework. OpenAI is launching them with tighter safeguards for higher-risk cybersecurity assistance, and says these protections "also restrict some legitimate vulnerability research and security testing workflows." GPT-6.2 Astra is also OpenAI's first externally released model to meet the High threshold for AI Self-improvement.

Benchmarks and comparisons

OpenAI-reported results from the official announcement, accessed Nov 17 2026. These are the publisher's own measurements, and test settings can differ across models.

  • Agents' Last Exam: GPT-6.2 Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost. OpenAI notes that the leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model.
  • Terminal-Bench-Science 0.1: Astra scores 9 percentage points above GPT-6 Astra at less than one-third the cost. OpenAI's footnote says its Fable 5.2 result is as reported by Artificial Analysis in the mini-swe-agent harness, and that Anthropic reported 74.6% for Fable 5.2 at max reasoning in Claude Code.
  • Harvey LAB-AA v1.1: Astra more than doubles Claude Fable 5.2's pass rate at about one-sixth the cost.
  • GDPval v2, an OpenAI internal benchmark: Astra is up 6.2 percentage points from GPT-6 Astra while using 22% fewer output tokens. GPT-6.2 Sol improves 5.3 points from GPT-6.1 Sol.
  • Cognition's FrontierCode 2.0: GPT-6.2 Sol comes within 0.5 points of GPT-6.2 Astra at about one-fifth the cost, both at max.
  • Artificial Analysis AA-AnalystAgent: Astra answers 62.5% of questions correctly on all five independent attempts.
  • Artificial Analysis AA-3D: Astra ranks first with 1,318 Elo, followed by Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259.
  • LifeVal: GPT-6.2 Sol outscores GPT-6 Astra at about one-seventh the cost.

Official samples and early work

  • Early tester, ClickUp: Jay Hack, Head of AI, says Astra edged out its production model in blind testing on the apps, dashboards and slides ClickUp Brain builds, with about 40% fewer tokens and 30% fewer review rounds.
  • Early tester, Basis: Eros Wang, Member of Technical Staff, says average time for long-horizon accounting work fell by 50%, and by up to 73% for the most complex work.

What it means for agent work

For operations teams, the practical change in this release is the cost of a finished task, more than any new kind of capability.

Price the task, not the token. OpenAI's case for Astra combines lower token prices with fewer output tokens per task. Before rerouting a workflow, measure cost per completed task on your own work, since that is the number the vendor's claims are about.

Make Sol the default and Astra the escalation. Sol keeps GPT-6.1 Sol's prices, and on FrontierCode 2.0 it came within 0.5 points of Astra at about one-fifth the cost. A routing rule that sends routine steps to Sol and escalates long, document-heavy or high-stakes work to Astra follows OpenAI's own split between the two.

Plan for more refusals in security work. OpenAI says the new safeguards restrict some legitimate vulnerability research and security testing. An agent doing that work needs a defined fallback and a named person to handle blocked requests, rather than retrying quietly.

Make blockers visible. OpenAI says that in two alignment tests both models were more likely to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable search tool, and that it added instructions in Codex that substantially reduce these omissions. Require agents to report blockers explicitly, and have a person confirm results before they move to the next step. Archon's security page lists "SSO, RBAC, approval queues, and audit logs" as its access model; for the design detail, see governed AI agents need approval queues and audit trails.

Keep prompt prefixes stable. OpenAI recommends keeping prompt prefixes unchanged and appending new instructions, to improve cache reuse, lower costs and latency, and preserve reasoning across turns. With cached input at $0.40 per million tokens on Astra, that is a prompt-template rule worth enforcing.

Route, don't rebuild. Several of OpenAI's comparisons are against Claude Fable 5.2. Archon is "model agnostic by design", so a release like this is a routing update rather than a rebuild; see why enterprise AI should be model agnostic.

The results above are OpenAI's own. Run a sample of your own work on both models before changing a production route. To work through routing, approvals and cost for your team, Book a Consultation.

How to try it

  • ChatGPT Work and Codex: available today to all Plus, Pro, Enterprise and Edu users. Eligible Enterprise and Edu users get the same access once their workspace admin enables it. Start in ChatGPT Work.
  • Ultrafast: Pro 500 subscribers can use Ultrafast with either model. Pro 200 subscribers get Ultrafast for the first time, starting with GPT-6.2 Sol.
  • API: developers can call gpt-6.2-astra and gpt-6.2-sol through the OpenAI API, with Ultrafast available for both models at launch. OpenAI's guide to preserved reasoning covers its prompt-prefix advice.

Sources

Newsletter

Join our newsletter.

Stay up to date on everything AI: new agent capabilities, model upgrades, and early access. No spam, ever.