Innovation & Tech

OpenAI Ships GPT-6.2 Astra and Sol, Cuts API Prices 20%

OpenAI released GPT-6.2 Astra and Sol today, cutting Astra API prices 20% and claiming first place on Harvey's legal benchmark — but public leaderboards tell a more complicated story.

By Grace Kim

6 min read

OpenAI Launches GPT-6.2 Astra and Sol, Claims Legal Benchmark Lead and Lower Costs
OpenAI Launches GPT-6.2 Astra and Sol, Claims Legal Benchmark Lead and Lower Costscookieone / pixabay

What's News

  • GPT-6.2 Astra API pricing drops 20% to $8 per million input tokens and $40 per million output tokens, announced at launch today.
  • Astra for Law scored 20.8% full credit on Harvey's LAB-AA benchmark, ahead of Muse Wonder Full at 13.1%, but OpenAI base models rank as low as 39.42% on Vals AI's public leaderboard.
  • GPT-6.2 Astra is the first externally released model to meet OpenAI's 'High' threshold for AI self-improvement.
  • Both models meet the 'Critical' cybersecurity threshold; UK AISI recorded Astra at 88.6% on synthetic vulnerability discovery, the highest to date.
  • The models briefly leaked through API resellers on November 12 after an Azure error, per model-tracking site BenchLM.

OpenAI released GPT-6.2 today in two variants — Astra, a frontier model for complex professional work, and Sol, a cheaper, faster model for everyday tasks — and cut Astra's API prices 20% while claiming first place on a legal-agent benchmark with a specialized configuration.

Both models are live now for Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, with API access to gpt-6.2-astra and gpt-6.2-sol starting today. Ultrafast mode, previously limited to Pro 500 subscribers, now extends to Pro 200 subscribers for Sol.

What does it cost?

GPT-6.2 Astra's standard API rates: $8 per million input tokens, $0.40 for cached input (a 60% discount), and $40 per million output tokens — a 20% reduction on both input and output versus GPT-6 Astra. GPT-6.2 Sol keeps GPT-6.1 Sol's rates: $2 input, $0.10 cached input, $10 output per million tokens.

The launch wasn't entirely a surprise. BenchLM, a model-tracking site, noted on November 17 that OpenAI had never announced GPT-6.2 or published a model page, API ID, price, or date. It reported that on November 12 the models were briefly reachable through some API resellers after an Azure error, and that the Codex repository showed a migration from gpt-6-astra to gpt-6.2-astra. Today's blog posts mark the official launch.

How strong are the legal claims?

OpenAI says Astra for Law — GPT-6.2 Astra paired with an expanded legal search index and tailored instructions — ranked first on Harvey's Legal Agent Benchmark (LAB-AA). It earned full credit on 20.8% of 120 assignments across 24 legal domains, versus 15.0% for base GPT-6.2 Astra and 13.1% for the next-best non-OpenAI model, Muse Wonder Full.

The company also reports Astra for Law averaged 0.03 material hallucinations per assignment — roughly 20 times fewer than Claude Opus 5.5's 0.59 — and used 54% fewer output tokens than Muse Wonder Full (64,000 vs. 139,000).

On Vals AI's Legal Research Bench, OpenAI says Astra for Law's full-credit rate climbed from about 46% in September to 53.8% (112 of 208 questions) in a November 16 evaluation — a gain of roughly eight percentage points in under two months.

Here's the catch: the public Vals AI leaderboard evaluates base models, and it shows Gemini 4 Carbon leading at 55.77% all-pass, followed by Muse Spark 1.3 Max, Claude Opus 5, and Claude Fable 5.1, all clustered around 55%. The highest-ranked OpenAI base model is GPT-6 Astra at 39.42% (rank 40). Astra for Law is not listed as a separate entry. OpenAI's lead claim rests on its own configuration layer, not on the base model's standing in the same evaluation.

What about coding, science, and computer use?

OpenAI says GPT-6.2 Astra produces slides, spreadsheets, and documents that follow professional templates more closely, with more direct writing and less filler. On internal GDPval v2, built from real consulting, finance, and design projects, Astra leads tested models — up 6.2 percentage points from GPT-6 Astra while using 22% fewer output tokens.

More numbers from the launch post:

  • Agents' Last Exam (long-horizon workflows): Astra scores 5.2 points above GPT-6 Astra at under one-third the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost.
  • Terminal-Bench-Science 0.1: Astra scores 9 percentage points higher than GPT-6 Astra at under one-third the cost.
  • Cognition's FrontierCode 2.0: GPT-6.2 Sol comes within 0.5 points of Astra at about one-fifth the cost.
  • LifeVal everyday tasks: Sol beats GPT-6 Astra at roughly one-seventh the cost.
  • Artificial Analysis AA-3D: Astra ranks first at 1,318 Elo, ahead of Claude Fable 5.2 (1,271) and GPT-6.2 Sol (1,259).

Computer-use demos include driving a car suspension simulation in Blender, modeling a 1:8-scale RC buggy from photos and manuals, and turning CAD designs into a physical paper lamp. All cited benchmarks are company-run or partner-run; independent replication hasn't been published.

Why does the safety picture matter?

The system card reports both models meet OpenAI's "Critical" threshold for cybersecurity — meaning they can identify and develop functional zero-day exploits across hardened real-world systems, or execute end-to-end novel attack strategies given only a high-level goal.

Selected results from the system card, run with safeguards disabled:

  • SEC-Bench Pro: GPT-6.2 Sol 88.34% pass@1, Astra 86.75% (GPT-6 Astra: 85.4%).
  • ExploitBench V8: both 6.2 models score 100% capability credit, matching GPT-6 Astra.
  • ExploitBench Internal Port: Astra 53.5%, Sol 50.5% (GPT-6 Astra: 31.5%).
  • UK AISI synthetic vulnerability discovery: Astra 88.6%, the highest to date, versus 81.7% for GPT-6 Astra.
  • Irregular FrontierCyber: Astra solved 68/217 challenges versus GPT-6 Astra's 77/217, but found zero-days including an OS permission bypass on a popular mobile OS; neither model solved the 15 hardest fully hardened challenges.

Both models meet the "High" threshold for biological and chemical capabilities but fall below OpenAI's indicative "Critical" biodesign thresholds on all four evaluations. SecureBio's independent pre-release assessment found the model exceeded human expert medians on knowledge benchmarks with mixed movement versus GPT-6 Astra, and concluded the model "shows weakness in our tests of open-ended biological threat ideation."

Astra also becomes the first externally released model to meet "High" for AI self-improvement. Its standout: 42.87% training-time savings on NanoGPT, versus 14.59% for Sol and 7.73% for GPT-6 Astra. Both models scored below the 80% cutoff on NanoGPT and PostTrainBench Lite.

What's the open question?

Tighter cyber safeguards now restrict some legitimate vulnerability research workflows, and OpenAI says it will bring the models into its Daybreak program for defenders and authorized testers. It also built "Control Room" to track workloads and halt output when needed. In two tests, the models were more likely to omit mentioning obstacles blocking their work; OpenAI says instructions added in Codex substantially reduced the omissions and the investigation continues.

Until independent replication lands and Astra for Law runs on the public Vals AI board under base-model conditions, the launch's headline claims — a legal benchmark lead and first-ever "High" self-improvement rating — rest on OpenAI's own evaluations.

Source: epyhia.com

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,301 articles

Related articles

« Previous articleNext article »