NewsAI and automationGPT-6.2 Astra and Sol

GPT-6.2 Astra and Sol: pricing and the real cost of your tasks

GPT-6.2 Astra and Sol: API pricing, conditions and a method to compare quality, turnaround time and real cost on your company’s tasks.

Mehdi B.

Co-founder · SEO & digital strategy

GPT-6.2 Astra and Sol: pricing and the real cost of your tasks
In brief

OpenAI launches GPT-6.2 Astra and Sol with separate API pricing for complex projects and everyday use. Compare their quality, speed and cost on your own tasks, counting the human corrections.

On 17 November 2026, OpenAI launches GPT-6.2 Astra and Sol for professional work, in ChatGPT Work, Codex and its API. This interface lets you connect the models to your company’s tools. To choose, we focus on the cost of a task completed correctly, correction time included, rather than the headline price alone.

GPT-6.2 Astra and Sol: what does this release change?

OpenAI positions Astra for complex projects and Sol for everyday use, with a focus on efficiency. Both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscriptions. Some features require activation by the workspace administrator.

For integrations, the credentials are gpt-6.2-astra and gpt-6.2-sol. A token is a fragment of text used to measure consumption. Input is what is sent to the model; output is what it produces.

At the standard rate, Astra costs 8 USD for input and 40 USD for output per million tokens. Reading cached input, already stored for reuse, costs 0.40 USD. Sol charges 2 USD, 10 USD and 0.10 USD respectively.

According to OpenAI, Astra's input and output prices fall by 20% compared with GPT-6 Astra. The drop reaches 60% for cached inputs. The company also announces better performance in its professional task evaluations.

Why is the price list not enough for your budget?

GPT-6.2 Astra and Sol: pricing and the real cost of your tasks

Take an online shop that uses AI to prepare its product pages and review its site code. For a product page, the team checks the features, the tone and the absence of any invented promise. For a change to the basket, it also checks how it works after the fix.

In this scenario, compare the models on a result accepted by the person in charge. Add up the calls needed, the retries and the review time. A fast answer that requires several redos is not the same amount of work as a proposal that is usable straight away.

Watch the pricing conditions too. Beyond 272,000 input tokens, input and cache prices double for the whole request; output prices are multiplied by 1.5. Writing to the cache also has its own rate, separate from reading it.

So avoid systematically sending the whole catalogue and all the code with every request. Prepare the information useful for the task and ask for an output of suitable length. For a product sheet, specify the expected fields and provide the validated characteristics.

How do you compare models on your own tasks?

Before changing an automation, build a reproducible test of 20 tasks. Split them across Next.js development, application security, SEO writing and business analysis. This number is meant to organise a first trial; it does not guarantee a measurement representative of your whole activity.

For our shop, this means fixing a site component, reviewing access rights, writing a product page and analysing sales. Give each model the same documents and the same instructions. Keep the settings, the answers and the results of the checks so you can repeat the comparison.

Define success before you send the requests. A product sheet respects the product data; a sales analysis finds the figures in the file provided. A code fix passes the planned tests without breaking an existing function.

For security checks, our news item on GitHub AI Scan complements this topic. Include the confirmed alerts and the diagnostic errors in your evaluation grid. Do not score only the clarity of the report.

Which checks should you ask your team for?

  1. Quality and corrections: who validates each result, against which criteria, and how many rework rounds are needed? Also keep the failures in the review, to avoid a comparison limited to the successes.
  2. Actual turnaround: how much time passes between the request and the usable result? Measure the model’s wait time and the human intervention separately, especially for tasks that block an order.
  3. Total cost: how much does each accepted task cost, including calls, tools used and review? Present the results by task type before choosing a model for each use.

Sources

OpenAIArtificial intelligenceAutomation
Mehdi B.
About the authorMehdi B.

Co-founder of Smart Impact. Hooked on the web from the start, he launched his first project in 2006: an online music magazine still running today. With close to 20 years of SEO experience, a Swiss federal diploma in marketing and a solid geek streak, he turns clients' ideas, sometimes hazy ones, into concrete digital projects with his team.

LinkedIn
Further reading