AI news for business · checked November 18, 2026

Evaluate GPT-6.2 Astra and Sol for real work

OpenAI introduced GPT-6.2 Astra and Sol on November 17, 2026. Sol is aimed at recurring everyday tasks, while Astra is designed for complex assignments where output quality matters more. The announcement highlights improvements across documents, spreadsheets, presentations and software development. For your team, the decision comes down to how quickly a reviewer can approve the result and what the complete workflow costs.

LOL.MARKETING / CONNECTING INTELLIGENCE

The essential answer

Start by measuring Sol on recurring assignments. Test Astra on the same input when a difficult deliverable consistently needs extensive revision despite a clear template. Record the output, review effort and access conditions separately.

01 / FIT

A good fit when

  • For teams producing recurring drafts, analyses or software changes with reviewed reference outputs.

02 / LIMITS

Not the first choice when

  • For automatic approval of figures, contracts or production code changes without an established review.

Which tasks make a fair comparison?

Choose two assignments from an ordinary week. A routine case might be a status update from approved project notes. A demanding case might be a decision memo that reconciles conflicting sources and keeps open assumptions visible. Specify the output format and the review criteria for both tasks before running either model. This helps you distinguish an improvement from switching models from an improvement you could get by clarifying the assignment.

What does an accepted API result cost?

OpenAI lists standard API pricing per million tokens at USD 8 for input, USD 0.40 for cached input and USD 40 for output for Astra. Sol is priced at USD 2, USD 0.10 and USD 10 for the same categories. Use those rates as the basis for the estimate. Record repeat calls, any separately billed tools and the time a reviewer spends on the result. A more expensive run can pay off if the time needed for corrections drops substantially. Use actual usage from your test.

As an illustration, a single call with 20,000 uncached input tokens and 2,000 output tokens costs USD 0.24 with Astra and USD 0.06 with Sol at those rates. If API usage reports some input as cached, price that portion at the cached rate. Compare the number of calls required to produce an acceptable result after it clears review.

What availability has OpenAI confirmed?

The launch announcement lists ChatGPT Work and Codex access for Plus, Pro, Enterprise and Edu, and gives the API model IDs gpt-6.2-astra and gpt-6.2-sol. Before a pilot, confirm the model picker in the intended account and access from the correct API project. A model family name is too broad for a useful comparison. Record the exact version, the surface you used and the reasoning setting where applicable. This gives a second reviewer enough information to reproduce the conditions.

What makes a deliverable better?

For a decision memo, review the evidence and figures first. Each material claim should point to a passage the reviewer can verify. Record omissions, unclear assumptions and any error that could change the decision. Only then assess structure, tone and formatting. Polished writing can make a substantive error easy to overlook. Give both models the same reference material, the same assignment and comparable tool permissions. Repeat across several tasks before turning one strong result into a recommendation.

How do you keep the pilot contained?

Start with synthetic or explicitly approved material. A folder with versioned documents and precomputed reference figures is enough for the first comparison. Provide only the files needed for the assignment and keep drafts in a separate location. Before anything is shared, have a reviewer check for omissions, unsupported claims and accidentally copied confidential passages. If the team wants to connect new sources later, the people responsible for those datasets can decide what information to include and who should have access.

Choose a default for each kind of work

For each assignment, record which run cleared review, how long approval took and what it cost to get there. If both models pass reliably, decide based on total expense and staff time. If Astra consistently passes the difficult assignment on its own, reserve it for that class of work and keep Sol as the routine default. Rerun the reference cases after the next model change. This gives you evidence for the decision and makes it easier to return to the previous process if quality falls below the agreed standard.

Keep it verifiable

Primary sources

The next sensible step

Choose a model for your workflow

Eight steps from a general interest in AI to a clearer decision for your business.

Start AI System Check
FreeProvider-neutralNo credentials