OpenAI Releases GPT-6.2 Astra and Sol with Lower API Rates and Reduced Token Usage

529 words
3 minutes
OpenAI Releases GPT-6.2 Astra and Sol with Lower API Rates and Reduced Token Usage

What changed#

OpenAI has introduced two new models: GPT-6.2 Astra, designed for complex professional tasks and agentic workflows, and GPT-6.2 Sol, aimed at efficient everyday work.

GPT-6.2 Astra Capabilities and Benchmarks#

GPT-6.2 Astra generates cleaner code, improved CAD/3D visuals, and formatted spreadsheets, slides, and legal briefs using fewer output tokens per task.

  • API Pricing: Input and output token prices are 20% lower than GPT-6 Astra (8permillioninputtokens,8 per million input tokens, 0.40 per million cached input tokens, and $40 per million output tokens). Cached input receives a 60% discount.
  • Agent Benchmark: On Agents’ Last Exam, Astra scored 5.2 points higher than GPT-6 Astra at less than one-third the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of its cost.
  • Scientific Workflows: Astra set a new high score on Terminal-Bench-Science 0.1, scoring 9 percentage points higher than GPT-6 Astra at less than one-third the cost. It also reached a top score on the Vals Science Index.
  • Legal and Knowledge Work: On Harvey LAB-AA v1.1, Astra more than doubled Claude Fable 5.2’s pass rate at roughly one-sixth the cost, with a tailored Astra for Law variant scoring higher. On GDPval v2, Astra improved by 6.2 percentage points over GPT-6 Astra while generating 22% fewer output tokens.
  • Data Analysis: On Artificial Analysis’s AA-AnalystAgent evaluation, Astra answered 62.5% of questions correctly across five independent attempts on a specific task calculating medical transport spending from California Medicaid expenditure reports.
  • Design and Interactive Web: Astra achieved 1,318 Elo on Artificial Analysis’s AA-3D human preference benchmark, placing first ahead of Claude Fable 5.2 (1,271 Elo) and GPT-6.2 Sol (1,259 Elo).

GPT-6.2 Sol Capabilities#

GPT-6.2 Sol maintains the API token pricing of GPT-6.1 Sol while upgrading reasoning, writing, and coding capabilities.

  • Code Evaluation: On Cognition’s FrontierCode 2.0, Sol finished within 0.5 points of Astra at approximately one-fifth the cost (both evaluated at maximum settings).
  • Everyday Tasks: On the LifeVal evaluation, Sol outscored GPT-6 Astra at approximately one-seventh the cost.
  • GDPval v2: Sol improved by 5.3 points over GPT-6.1 Sol.

Why it matters#

Reducing the number of output tokens needed to resolve complex tasks directly decreases total workflow cost and execution time for multi-step agents. Astra’s performance gains on scientific, legal, and agent benchmarks lower the barrier to running long-horizon evaluations at scale.

Meanwhile, GPT-6.2 Sol’s performance on FrontierCode 2.0 makes high-volume code maintenance and daily interactive tasks viable at lower cost tiers. Practitioners evaluating modern API options across providers (such as those comparing architectures in our guide to architecting with Claude) now have updated cost-per-task metrics to consider.

Limitations and caveats#

  • On the Agents’ Last Exam leaderboard, Claude Fable 5.2’s published results excluded nine tasks that required a fallback model.
  • On the AA-AnalystAgent benchmark, Astra’s 62.5% accuracy rate specifically reflects performance across five attempts on a single California Medicaid expenditure task, rather than an overall score across all financial spreadsheet benchmarks.

What to do next#

  1. Implement Prompt Caching: Ensure system prompts and repetitive contexts use Astra’s cached input endpoint to take advantage of the $0.40/1M cached token rate.
  2. Benchmark Sol for High-Volume Tasks: Test GPT-6.2 Sol on automated PR reviews, background evaluations, and everyday user requests where Astra’s frontier reasoning isn’t strictly required.
  3. Audit Token Limits: Adjust downstream parser timeouts and token budgets to account for Astra’s reduced output token footprint on complex tasks.

Sources#

Share Article

If this article helped you, please share it with others!

Comments

Table of Contents