Seriously, It’s AI Seriously, It’s AIAI News • Tools • Trends Subscribe
Latest
AI News

GPT-6.2 Astra and Sol Are Here: OpenAI's New Models Beat Claude on Cost — and One Honest Catch

November 18, 2026 By Seriously, It’s AI 0 Comments


OpenAI released two new models yesterday, November 17: GPT-6.2 Astra, its new frontier model for the hardest professional work, and GPT-6.2 Sol, a cheaper everyday model. I read OpenAI's full announcement and the accompanying system card so you don't have to take the launch numbers at face value. Here is what actually changed, what it costs, and the one part of this launch that deserves a second look.

What OpenAI actually launched

Astra is the flagship: built for presentations, spreadsheets, documents, coding, websites, games and 3D design. OpenAI says its slides and spreadsheets now follow the templates and conventions of each field more closely, and read less like generic AI output. Sol is the efficient sibling for everyday work — OpenAI says many of its own employees now use Sol as their default model.

Both models are live now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol. Neither was added to regular Chat — this rollout is aimed squarely at work. Both carry a 1,050,000-token context window. It continues a busy month for OpenAI — last week it put computer-use agents inside ChatGPT for Healthcare, and days before that it launched ChatGPT Wallet.

The price cut is the real story

In the API, Astra costs $8 per million input tokens and $40 per million output, with cached input at $0.40 — 20% below GPT-6 Astra, and 60% lower on cached input. Sol keeps GPT-6.1 Sol's prices: $2 in and $10 out per million tokens.

This is a price war in plain sight. Alibaba's Qwen4-Max arrived two days earlier with its own token savings, and Google made Gemini 4 Carbon generally available last week at $2 in and $10 out until November 30.

The benchmarks — impressive, but read the fine print

OpenAI's headline claim: on Agents' Last Exam, which tests long professional workflows, Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost — and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost. On AA-3D, where human judges compare blinded 3D experiences, Astra ranks first at 1,318 Elo, ahead of Fable 5.2 (1,271) and Sol itself (1,259). In legal work, Astra more than doubles Fable 5.2's pass rate on the Harvey LAB-AA benchmark at about one-sixth the cost.

Sol is the quiet surprise of the pair. It comes within 0.5 points of Astra on FrontierCode 2.0 — a test of whether code changes meet real maintainers' standards — at about one-fifth the cost. And on LifeVal, OpenAI's test of everyday tasks, Sol outscores GPT-6 Astra at about one-seventh the cost.

Here is the fine print: every one of those numbers is OpenAI's own measurement. The one independent score I could check tells a slightly different story — Artificial Analysis puts Astra second on its Intelligence Index at 62, behind Claude Fable 5.2 at 65. That does not make OpenAI's claims wrong, but the "beats Claude" headline depends on which test you look at.

Why this matters

For anyone paying per token, this launch changes the maths. For most teams, the default should now be the cheaper tier, not the flagship. Sol doing everyday work at one-seventh of GPT-6 Astra's cost is the kind of number that quietly rewrites budgets.

The honest catch

Astra is the first OpenAI model released externally at the "High" level for AI self-improvement — OpenAI's own rating for a model's ability to accelerate AI research. Both models also meet the "Critical" threshold for cybersecurity; Astra is OpenAI's most capable cyber model yet, 22 points above GPT-6 Astra on ExploitBench Internal Port.

Because of that, OpenAI launched both models with tighter safeguards for higher-risk cybersecurity help — protections it admits will also restrict some legitimate vulnerability research. And there is a wrinkle OpenAI itself discloses in the system card: the UK AI Security Institute found a regression in jailbreak robustness, with its red-teamers building a universal cyber jailbreak using half the queries GPT-6 Astra had required. OpenAI says it believes the issue is fixed; the institute had not confirmed that at launch.

My take: the capability jump is real and the prices are genuinely lower, but use this launch with your eyes open. If you do security work, check whether the tighter safeguards block your legitimate workflows before you switch — OpenAI points defenders to its Daybreak program. Treat the benchmark claims the way you would treat any vendor's own numbers: as a starting point, not a verdict.

What to watch next

First, whether independent benchmarks confirm OpenAI's cost-per-task claims once the models are in wider hands. Second, whether the tighter cyber safeguards frustrate legitimate security researchers. And third, the wider race: Anthropic begins trading as a public company on November 19.

Would you switch your daily work to Sol at a fifth of the cost — or does the flagship still earn its premium for you? Tell me in the comments.

Sources

Comments

Browse by Category

Stay Ahead with Seriously, It’s AI

Get the latest AI news, trends and expert insights — explained simply, every day.

Follow Us →

Stay updated with AI

One short email a week. No spam, unsubscribe anytime.