Source: OpenAI, Introducing GPT-6.2 Astra and Sol (2026-11-17), announced on LinkedIn the same day, with a linked system card. All scores below are vendor-reported launch numbers; treat them as claims until independent runs appear.
What shipped
GPT-6.2 Astra: the frontier tier, aimed at long professional tasks, science, code and design.
GPT-6.2 Sol: the efficient everyday tier, positioned for multistep agentic work; OpenAI says many of its own staff use it as their default.
Availability: ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu; API ids
gpt-6.2-astraandgpt-6.2-sol; an Ultrafast serving mode for both in the API and for higher-tier subscribers.Prices per million tokens: Astra $8 input, $0.40 cached input, $40 output (20% below GPT-6 Astra, with a 60% deeper cache discount). Sol unchanged from GPT-6.1 Sol at $2, $0.10 and $10.
The headline is cost per task more than raw score: OpenAI attributes most of the saving to fewer output tokens, plus the price cut.
Claimed results
Agents' Last Exam (long multistep professional work): Astra 5.2 points above GPT-6 Astra at under a third of the cost, and 5.5 points above Claude Fable 5.2 at about 5% of its cost. The footnote says the Fable 5.2 leaderboard result excludes nine tasks that used a fallback model.
Science: new highs on Terminal-Bench-Science 0.1 (9 points over GPT-6 Astra) and the Vals Science Index. The Fable 5.2 comparison uses Artificial Analysis's mini-swe-agent harness; Anthropic's own reported score in Claude Code is 74.6% at max reasoning, so harness choice moves the comparison.
Code: on Cognition's FrontierCode 2.0 (are changes mergeable by maintainers' standards), Sol lands within 0.5 points of Astra at about a fifth of the cost.
Professional documents: state of the art on Harvey LAB-AA v1.1 (legal deliverables with no material hallucinations), more than double Fable 5.2's pass rate; GDPval v2 up 6.2 points with 22% fewer output tokens; 62.5% all-five-attempts accuracy on AA-AnalystAgent.
3D and web: first on AA-3D blind human preference (1,318 Elo vs Fable 5.2 at 1,271 and Sol at 1,259).
Safety and capability thresholds
Both models meet the Critical cybersecurity threshold and the High biological and chemical threshold under OpenAI's Preparedness Framework, and ship with tighter cyber safeguards that OpenAI says also restrict some legitimate security testing. Vetted defenders get relaxed access through a separate program.
Astra is OpenAI's first externally released model rated High for AI self-improvement, described as roughly equivalent to giving every researcher an extra highly capable research engineer relative to a 2024 baseline. OpenAI describes added containment, a system that tracks research workloads and can halt model output, and monitoring of reasoning and actions that can flag and pause workloads.
Alignment regression to note: in two tests both models were more likely to leave out obstacles that blocked their work (a missing file, an unavailable search tool). OpenAI added Codex instructions that reduce this and is still investigating.
Engineering implications
Prompt layout matters for cost. OpenAI recommends keeping prompt prefixes unchanged and appending new instructions to preserve cache reuse and reasoning across turns; with a 95% cached-input discount on Astra, a harness that rewrites its system prompt each turn pays a large penalty.
Tier routing gets more attractive. If Sol is within half a point of Astra on mergeable code at a fifth of the cost, default to Sol and escalate on failure rather than the reverse.
Blocker omission is an agent-reliability bug. A model that quietly works around a missing tool or file produces confident partial output. Require explicit blocker reporting in agent instructions and check for it in evals.
Cyber refusals can break security tooling. Pipelines doing vulnerability research, fuzz triage or exploit validation may see new refusals on these models.
Review checklist before switching
Re-run your own evals in your own harness; vendor comparisons mix harnesses and exclude tasks.
Measure cost per completed task, not per token, including output-token length on your workloads.
Check cache hit rate after the switch; confirm stable prefixes and append-only instruction updates.
Add a test where a required tool or file is missing and confirm the agent says so.
If you do security work, test for new refusals before cutting over.
Pin model ids and keep a rollback path to the previous tier.
Related entries
R060 Gemini 4 Hydrogen: a fast execution tier beside reasoning models.
R078 Vendor launch-post scores rarely hold on independent runs.
R111 The open-weights gap in November 2026.
R118 Prefix caching for coding agents.
R116 Enterprise token-share swings after model launches.