GPT-6.2 Astra and Sol: what commerce teams should test

OpenAI has launched GPT-6.2 Astra and Sol, positioning Astra for demanding projects and Sol for everyday work. Its launch announcement highlights document and spreadsheet quality, template following…

Will Slatton3 min read

OpenAI has launched GPT-6.2 Astra and Sol, positioning Astra for demanding projects and Sol for everyday work. Its launch announcement highlights document and spreadsheet quality, template following and frontend development. Those are relevant tasks for commerce teams, but the company’s evaluations do not establish results on a particular store’s workload.

Prices and access

Standard Astra API prices are $8 per million input tokens, $0.40 for cached input and $40 for output. Sol is $2, $0.10 and $10 respectively. OpenAI says both are available in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu users, as well as through the API. Workspace enablement and Ultrafast access vary by plan.

OpenAI also describes tighter safeguards for higher-risk cybersecurity assistance. A team evaluating the models for authorised app security work should test its actual workflow rather than assume identical behaviour to an earlier model.

Compare finished work, not just token prices

Our recommendation is to run the same small set of tasks through the existing model and the replacement. Include a product-data cleanup, an analysis with a known answer and a code change with meaningful acceptance checks. Keep the input files and requirements identical.

For a reporting workbook, check formulas against source data and inspect the exported file. For a storefront change, test the resulting interaction on mobile and desktop. For customer-facing copy, check product claims and tone before release.

Record total cost, correction time and failures as well as speed. A cheaper response that needs extensive repair can be the more expensive workflow. Keep production changes behind the team’s existing review process until the comparison supports a switch.

ℹ️
Launch benchmarks are vendor evidence, not a guarantee of accuracy or savings on your own store’s tasks.

Context length and speed change the bill

The official Astra and Sol model pages specify a 1,050,000-token context window. That capacity should not be confused with the short-context price: requests above 272,000 input tokens double input and cache rates and charge 1.5 times the output rate for the full request.

At Standard short-context pricing, a 30-minute cache write costs $10 per million tokens for Astra and $2.50 for Sol. A 12-hour write costs $20 and $5 respectively. These write charges are separate from the lower cached-input prices above. Teams repeatedly supplying a catalogue or instruction set should measure cache reuse before estimating savings.

Fast mode costs twice Standard for the same context length. Ultrafast costs three times Standard through 28 February 2027, then six times Standard from 1 March 2027, Pacific time. Batch costs half the Standard rates. Record the processing mode and context size alongside token totals when comparing costs.

An independent coding comparison adds a caution

In its 17 November Coding Agent Index report, Artificial Analysis says Astra is flat against GPT-6 Astra, while Sol scores three points below GPT-6.1 Sol, using the Codex harness at maximum effort. That is one evaluator’s specific coding setup, not a finding about every commerce task. It reinforces the need to test finished work against your own acceptance checks rather than treating the version number as proof of improvement.

Watch the official introduction

OpenAI’s official introduction to the models, 17 November 2026. Watch the original video on YouTube.

View the original OpenAI announcement on X, 17 November 2026.

19 November: shorter answers and delayed-work pricing

Artificial Analysis reports changes in its personality evaluation: average responses fell from 352 to 210 words for Astra and from 337 to 223 for Sol compared with the prior releases. It also reports less structure and a lower tendency to admit uncertainty. These describe that evaluation, not a guarantee about merchant copy or factual accuracy.

For product descriptions and support drafts, test brevity alongside completeness: a shorter answer still needs the correct conditions, limitations and source evidence. Separately, OpenAI has announced discounted background completion windows for eligible work that can wait. The timing target applies to the initial response, not an entire agent workflow.

Filed underAI

More from AI

See all →