DEV Community

UULab
UULab

Posted on

Introducing GPT-6.2 Astra & Sol: Benchmarks, API Pricing, and Practical Takeaways

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. The headline is not "smarter" — it is token efficiency. Scores went up while output tokens and per-token prices came down, and the everyday model (Sol) closed most of the gap with the frontier model (Astra). Here is what actually matters when you are deciding what to route where.

GPT-6.2 Astra & Sol overview

Two models, two jobs

Astra handles long-horizon professional work: documents, spreadsheets and slides, CAD/3D design, and computer use. Sol is the default engine for everyday and agentic tasks. Both are available in ChatGPT Work and Codex on Plus, Pro, Enterprise and Edu, and through the API as gpt-6.2-astra and gpt-6.2-sol.

Model ID Position ChatGPT Work / Codex Ultrafast Price change
gpt-6.2-astra Frontier model for complex professional work Plus, Pro, Enterprise, Edu Pro 500 Input/output 20% lower
gpt-6.2-sol High-efficiency default engine Plus, Pro, Enterprise, Edu Pro 500, Pro 200 Same as GPT-6.1 Sol
  • Enterprise and Edu workspaces need an admin to enable access.
  • Both models are available from launch day on the plans above.

Benchmarks: higher scores, fewer output tokens

The most useful row is GDPval v2, an internal evaluation built on long-running, project-style work: Astra scores 6.2 points higher than GPT-6 Astra while using 22% fewer output tokens. That combination — better answers with less generation — is what moves per-task cost.

Evaluation Result Notes
GDPval v2 Astra: +6.2%p vs GPT-6 Astra, 22% fewer output tokens. Sol: +5.3 points vs GPT-6.1 Sol OpenAI internal, long-horizon work
Agents' Last Exam Astra: +5.2 points vs GPT-6 Astra at under 1/3 the cost; +5.5 points ahead of Claude Fable 5.2 at ~5% of the per-task cost Fable 5.2 leaderboard figure excludes 9 tasks that used a fallback model
AA-3D (blind preference) Astra 1,318 Elo (1st) / Claude Fable 5.2 1,271 / GPT-6.2 Sol 1,259 Artificial Analysis, human pairwise voting
Harvey LAB-AA v1.1 (legal) More than 2× the pass rate at ~1/6 the cost vs Claude Fable 5.2
Terminal-Bench-Science 0.1 +9%p at under 1/3 the cost vs GPT-6 Astra
FrontierCode 2.0 Sol within 0.5 points of Astra at ~1/5 the cost Cognition, both at max settings
LifeVal (everyday tasks) Sol ahead of GPT-6 Astra at ~1/7 the cost Moving, budgeting and similar tasks

What changed in document and 3D work

Astra manipulates CAD and simulation tools down to individual parts: it can check how components mate, where load moves when the design changes, and test revisions before anything is manufactured. It also builds interactive 3D experiences that run in the browser.

Document output is where most teams will notice the difference. Boilerplate phrasing is reduced, and output follows company templates and industry conventions. Given only screenshots of an existing slide deck, it matches layout, color and font sizing — useful when you want a design template to act as a fixture in an automation pipeline.

GPT-6.2 Astra & Sol benchmarks and pricing

API pricing and prompt caching

Model Input Cached input Output
gpt-6.2-astra $8.00 $0.40 $40.00
gpt-6.2-sol $2.00 $0.10 $10.00

Per million tokens. Astra is 20% cheaper on input and output than GPT-6 Astra, with cached input 60% lower. Sol keeps GPT-6.1 Sol's pricing while scoring higher.

Cached input is 1/100 of output pricing ($0.40 vs $40.00), so cache hits dominate real workloads. The recommended pattern is simple:

  • Keep the prompt prefix stable: system prompt, tool definitions and fixed context.
  • Append new instructions at the end of the conversation instead of editing the prefix.
  • Preserve reasoning across turns where the API supports it.

Editing the prefix on every request invalidates the cache and silently gives up the discount.

Practical takeaways

  • Batch, classification and summarization pipelines should default to Sol. FrontierCode 2.0 puts it within 0.5 points of Astra at roughly one fifth of the cost.
  • Document deliverables, CAD/3D and front-end work justify Astra: template-conforming output and fewer output tokens per task.
  • Check your plan before designing around Ultrafast. Pro 500 gets both models; Pro 200 starts with Sol.

Safety notes

Both models fall at the Critical cybersecurity and High biological/chemical thresholds under OpenAI's Preparedness Framework. Higher-risk cyber support gets stronger safeguards, which can restrict some vulnerability research and security testing workflows. Astra is the first publicly released OpenAI model to meet the High threshold for AI self-improvement.

Bottom line

This release raises scores and lowers prices at the same time. Route by cost per task, not by model tier: Sol for high-volume repeated work, Astra for documents and design artifacts. Keep prompt prefixes stable so the cache actually pays you back.

Source: OpenAI — Introducing GPT-6.2 Astra & Sol

Figures, evaluation names and evaluators are taken from OpenAI's official announcement (dated November 17, 2026) and public evaluations by Artificial Analysis, Cognition and Harvey. Benchmark scores vary by evaluator.

Top comments (0)