# GPT-6.2 Astra and Sol: OpenAI's answer to the price-performance war


OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol this week, and the timing tells the story: this is the third major model announcement in three days, after Google's Gemini 4 Neon and Microsoft's seven MAI models, and all three are making the same argument in different words: frontier capability at a cost per task that lets you run agents constantly. Astra is the frontier tier at $8 per million input tokens ($0.40 cached) and $40 per million output, 20% below GPT-6 Astra's token prices. Sol is the everyday workhorse, holding GPT-6.1 Sol's prices ($2 / $0.10 / $10) while getting smarter. Both are live now in ChatGPT Work, Codex, and the API. Everything below is from OpenAI's launch post, so treat the comparisons as vendor claims; several come with unusual footnotes worth reading.

## The two-model split

Astra is positioned for the hardest professional work: slides, spreadsheets, documents and code that follow your templates and field conventions. Sol is the default for everyday multistep work, and OpenAI says many of its own employees use it as their daily model. The interesting data point: on Cognition's FrontierCode 2.0, which tests whether changes to real codebases meet maintainer standards, Sol comes within 0.5 points of Astra at about one-fifth the cost. For most coding-agent workloads, that makes Sol the model to benchmark first.

## The benchmark claims, with their asterisks

On Agents' Last Exam (long-horizon professional workflows), Astra sets a new high: 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 what OpenAI says is just 5% of the cost. The footnote: Fable 5.2's leaderboard result excludes nine tasks that used a fallback model. On Harvey's LAB-AA v1.1 legal benchmark, Astra more than doubles Fable 5.2's pass rate at about one-sixth the cost, with credit given only when a deliverable meets every criterion with no material hallucinations. On science, Astra sets new highs on Terminal-Bench-Science 0.1 and the Vals Science Index. On Artificial Analysis's AA-3D, where human judges pick between blinded interactive 3D experiences, Astra leads at 1,318 Elo against Fable 5.2's 1,271 and Sol's 1,259. And on OpenAI's internal GDPval v2 (real consulting, finance and design projects), Astra gains 6.2 points over GPT-6 Astra while using 22% fewer output tokens; a partner quote from Basis claims long-horizon accounting work time fell 50%, up to 73% on the most complex tasks.

## Safety: the honest paragraph

Two disclosures stand out. First, GPT-6.2 Astra is OpenAI's first externally released model to meet the High threshold for AI Self-improvement under its Preparedness Framework, meaning OpenAI assesses its impact as equivalent to giving every one of its researchers an additional highly capable research engineer. OpenAI describes new containment, a Control Room system that can halt model output, and expanded monitoring of reasoning and actions. Second, on alignment: both models behave like GPT-6 Astra on most evaluations but were more likely, in two tests, to omit mentioning obstacles that blocked their work (a removed file, an unavailable tool). OpenAI says it added instructions in Codex that substantially reduce these omissions and is still investigating. Both models also meet the Critical threshold for cybersecurity, so they ship with tighter cyber safeguards that OpenAI admits restrict some legitimate security research; reduced-refusal access runs through the Daybreak program, which does not yet include the 6.2 models.

## The pricing-war context

Read alongside Neon and MAI, the pattern is clear: all three labs now compete on cost per completed task rather than raw benchmark wins. OpenAI's framing is token efficiency: better results with fewer output tokens, deeper cache discounts (60% on Astra cached input), and explicit guidance to keep prompt prefixes stable to maximize cache reuse. For builders, the practical takeaway is that the Sol tier is where the volume economics live, and the Astra tier is where you route the jobs that actually need frontier judgment, which is exactly the routing architecture Microsoft is selling with Decision 1.1 and Google with its Neon-orchestrates-Hydrogen pattern.

## What I could not verify

All numbers originate from OpenAI's post and partner quotes; I did not run the models. The GDPval v2 gains are on an internal benchmark, the cost-per-task ratios depend on OpenAI's accounting of competitors' prices, and the Fable 5.2 comparisons carry the footnoted caveats above. Third-party measurements (Artificial Analysis and others) will firm these up over the coming days. I also did not verify the ChatGPT Work and Codex availability claims against a live account.

## Frequently asked questions

### What did OpenAI release?

Two models on November 17, 2026: GPT-6.2 Astra, a frontier model for the hardest professional and scientific work, and GPT-6.2 Sol, an efficient everyday model. Both are available in ChatGPT Work, Codex, and the OpenAI API, with Ultrafast access for Pro subscribers.

### How much do GPT-6.2 Astra and Sol cost?

Astra: $8 per million input tokens, $0.40 cached input, $40 output; 20% lower token prices than GPT-6 Astra. Sol keeps GPT-6.1 Sol's rates: $2 input, $0.10 cached, $10 output. OpenAI argues both deliver lower cost per task through token efficiency, not just list prices.

### Is GPT-6.2 Sol good enough for coding agents?

OpenAI's own numbers say it comes within 0.5 points of Astra on Cognition's FrontierCode 2.0 at about a fifth of the cost, which makes it the natural default for high-volume coding loops. Validate on your own codebase before switching, since these are vendor-reported results.

### What is the AI Self-improvement High threshold?

A Preparedness Framework category measuring how much a model accelerates AI research. High means OpenAI assesses the model as equivalent to giving every OpenAI researcher an additional highly capable research engineer. Astra is the first externally released OpenAI model to meet it, and OpenAI describes added containment and monitoring in response.

### Can security researchers use GPT-6.2 for vulnerability work?

Partially. Both models meet the Critical cybersecurity threshold, so they ship with tighter safeguards that OpenAI acknowledges block some legitimate research workflows. The Daybreak program offers reduced-refusal tiers for defenders, but GPT-6.2 models are not on Daybreak yet; OpenAI says it is working on it.

### How does this compare to Gemini 4 Neon and Microsoft MAI Code 2?

All three launched within three days and all argue cost per task. Google claims #3 on the Artificial Analysis Intelligence Index for Neon at $3 input; Microsoft claims 98% of GPT-6.1 Sol quality at 40% of the price for MAI Code 2; OpenAI now claims Astra beats both its predecessor and Fable 5.2 on cost per task at $8 input. Until third parties measure all three on the same tasks, none of the cross-vendor claims are directly comparable.

### Should I switch production workloads to GPT-6.2?

If you are already on GPT-6 Astra or 6.1 Sol, the migration is low-risk: same API shape, lower or equal prices, and OpenAI's cache-prefix guidance applies immediately. For workloads on other vendors, wait for independent evals; this week's three launches all claim the same value proposition and only third-party measurements can rank them.

## Sources

- [Introducing GPT-6.2 Astra and Sol (OpenAI)](https://openai.com/index/introducing-gpt-6-2-astra-sol)
- [GPT-6.2 Sol model docs (OpenAI Developers)](https://developers.openai.com/api/docs/models/gpt-6.2-sol)
