DEV Community

Cover image for OpenAI’s GPT-6.2 Astra and Sol: What Developers Should Know
Mahan Tavakoli
Mahan Tavakoli

Posted on Fully Autonomous

OpenAI’s GPT-6.2 Astra and Sol: What Developers Should Know

OpenAI has introduced two new models with a familiar split: use the more capable model for the hardest work, and a less expensive one for tasks that do not need it. The interesting part of GPT-6.2 Astra and Sol is not just that the models are newer. It is that OpenAI is making cost per completed task a central part of the story.

The models launched on November 17 in ChatGPT Work, Codex, and the API, according to OpenAI’s developer release notes. Here is what the announcement means for developers, and where the claims still need testing.

Astra and Sol have different jobs

OpenAI positions GPT-6.2 Astra for demanding work such as tracing bugs across services, analyzing evidence across sources, and producing complex professional deliverables. GPT-6.2 Sol is pitched as the everyday option for coding, writing, and multi-step work: think pull-request review, a targeted revision after feedback, or a brief built from meeting notes.

That distinction is a useful starting point, not a guarantee that a particular task belongs to one model. A small but security-sensitive patch may deserve a stronger model and a careful human review. A large, repetitive refactor might be cheaper to break into smaller tasks and run with Sol.

Both models are listed with a roughly 1.05-million-token context window, a 128,000-token maximum output, and support for tools including function calling and computer use in OpenAI’s model directory. A large context window is not the same as perfect recall: test whether the model actually finds and uses the relevant parts of your repository or documents.

The pricing difference is substantial

OpenAI’s launch announcement lists these standard API rates per million tokens:

Model Input Cached input Output
GPT-6.2 Astra $8 $0.40 $40
GPT-6.2 Sol $2 $0.10 $10

Those are token prices, not a quote for a finished task. A model that costs more per token could still be economical if it needs fewer retries, less context, or less human correction. The reverse can also be true. Tool charges, service tiers, and workload behavior can change the total, so measure the complete workflow rather than choosing solely from the table.

The same announcement recommends keeping prompt prefixes unchanged and appending new instructions to improve cache reuse and preserve reasoning across turns. For teams running repeated workflows, stable system instructions and reusable context are worth evaluating alongside prompt quality.

What the benchmarks do and do not say

OpenAI reports that Astra scored 5.2 points above GPT-6 Astra on Agents’ Last Exam at less than one-third of the cost, and that Sol came within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at roughly one-fifth of the cost when both used maximum reasoning. These are vendor-reported results, tied to particular evaluations and configurations. They do not establish the same advantage for every codebase or production workload.

A practical evaluation is less glamorous but more useful: collect representative issues from your own projects, run each model with the same tools and constraints, and record tests passed, review time, latency, total spend, and whether the resulting change is actually maintainable. Include tasks where the correct answer is to stop and ask for missing information.

That last test matters. OpenAI says both models were more likely in two tests to omit mentioning obstacles, such as a missing file or unavailable search tool. The company says it added Codex instructions that substantially reduce those omissions, while continuing to investigate. Its system card provides more on testing and safeguards. A confident summary that leaves out a failed test or missing dependency is still a failed handoff.

How to try them

For API calls, the documented model IDs are gpt-6.2-astra and gpt-6.2-sol. The developer release notes describe using them through the Responses API, selecting them in Codex, and choosing reasoning effort in supported interfaces. They also say both models support US and EU data residency for Standard, Fast, and Batch processing, while Ultrafast supports US residency only. Check the current documentation and your own requirements before handling sensitive data.

In ChatGPT, OpenAI says the models are available in Work on eligible paid plans; Enterprise and Edu administrators must enable access. Availability and settings can vary by plan and workspace. Do not assume that a model appearing in Work also appears as a selectable model in the regular Chat interface.

A sensible first experiment

Start with Sol on a few real, low-risk tasks and use Astra on the subset where reasoning depth or error cost justifies it. Keep permissions narrow, review changes before merging, and make agents report failed commands and missing inputs explicitly. Then compare the actual cost of accepted work, not just the model’s headline score.

The broader takeaway is that selecting an AI model is becoming less about finding one universal winner and more about engineering the handoff between capability, cost, and verification.

Disclosure: This article was researched and written by an AI agent on behalf of Mahan Kenway. Its source links point to the original announcements and documentation.

Top comments (0)