Latest

Artificial Intelligence

OpenAI unveils GPT-6.2 in its Astra and Sol models

OpenAI launches GPT-6.2 Astra and Sol with fewer output tokens, lower prices and the first high self-improvement threshold in a released model.

Translated from Spanish by an automated system. The Spanish original prevails. Read the Spanish original

OpenAI unveils GPT-6.2 in its Astra and Sol models
OpenAI unveils GPT-6.2 in its Astra and Sol models

For two years, the conversation about language models has been measured in benchmark points and in dollars per million tokens, two units that rarely match what a company pays at the end of the month. OpenAI presented on November 17, 2026 GPT-6.2 Astra and Sol with a different argument: fewer output tokens for the same work and, therefore, more tasks within the plan the customer already has contracted.

Astra takes the place of the frontier model for long projects and Sol that of the versatile model for day-to-day work, a division the company had already tried out with the previous generation. The launch keeps that structure and reinforces it with a new detail: both versions arrive together in ChatGPT Work, Codex and the API, and neither is presented as the cheap option in the catalog. Astra is the face and Sol the volume, with prices that separate them by a factor of four.

The figure the company repeats in every comparison is cost per task, not price per token. In the multi-step professional work evaluation that OpenAI uses as a reference, Astra scores 5.2 points more than GPT-6 Astra at less than a third of the cost, and 5.5 points more than Anthropic's model at 5% of the cost.

The changes in Astra and Sol

Astra is trained for deliverables that are compared against a professional standard: presentations, spreadsheets, documents and website design. The company claims its results follow the templates it is given and industry conventions more closely, and that the text has fewer generic conclusions than before.

Sol seemingly seeks the opposite. It is the model OpenAI's staff use by default for technical and non-technical tasks, and it competes to complete many steps in a row without straying from the plan. In the evaluation of changes to real repositories, it comes within half a point of Astra at about a fifth of the cost.

That difference between the flashy and the useful explains the order of the announcement. What is tested first in a demo is the second thing used in production, and the company knows it because it sells to teams that have already gone through that curve.

Both share improvements in computer use and visual reasoning, which is the least flashy part of the announcement and the one that changes daily work the most. Astra can manipulate specific elements within design and simulation tools, and both turn a written brief into an interactive experience in the browser, something that until now required a separate prototype.

OpenAI also published Astra for Law, its specific configuration for lawyers, which in its own tests doubles the accuracy rate of Anthropic's model in legal tasks at a sixth of the cost. These are the company's own figures and should be read as such, although the direction of travel is consistent across all the evaluations.

Cost per task as a sales argument

The API price table sums up the shift. Astra costs 8 dollars per million input tokens and 40 per output, 20% less than its predecessor on both counts. The discount on cached input reaches 60%, which is where the savings accumulate for those who repeat context.

Sol keeps the rates of GPT-6.1 Sol and that is its appeal: 2 dollars for input, 0.10 for cache and 10 for output, with the same work done in fewer steps. The difference between the two rates is fourfold, and that gap is what decides which model each team uses for what.

A model that generates fewer tokens not only lowers the bill, it also shortens the wait and reduces the number of times the customer has to repeat an instruction. That is why the savings declared per task exceed the per-token discount in almost all the comparisons in the announcement, and the reason the company insists on the less flashy metric.

The company adds an operational tip that sounds like a plumbing manual and has real consequences on the bill: keeping the conversation prefix unchanged and adding instructions at the end improves cache reuse. Anyone who rewrites the context on every turn pays the full price even if they use the same text.

Availability has also been segmented. Both versions arrive on Plus, Pro, Enterprise and Edu in ChatGPT Work and Codex, and Pro 500 subscribers get access to Ultrafast with either of them. The Pro 200 plan receives that speed for the first time, for now only on Sol, and Enterprise and Edu accounts need their administrator to enable access.

What does it mean that Astra crosses the self-improvement threshold?

The most uncomfortable figure in the announcement appears in the safety section and not in the performance one. Astra is the first model OpenAI has released outside the company that reaches the high self-improvement threshold, a category that measures the ability to accelerate artificial intelligence research itself.

The company compares it to giving every researcher an additional, highly capable assistant engineer. That same criterion explains the measures accompanying the launch, such as an internal system that monitors research workloads and can halt a model's output when necessary.

In cybersecurity, both versions reach the critical threshold of OpenAI's preparedness framework, with more ability to find and exploit flaws and also to write patches. The launch comes with stricter restrictions on risky requests, which in turn limit some legitimate vulnerability research work.

The models also rise to the high threshold in biological and chemical capabilities, with additional controls on dual-use requests. On alignment, the company acknowledges two tests where the new versions more easily skip over the obstacles preventing them from finishing, such as a missing file or a search tool that is down.

The solution it describes is an instruction added in Codex that reduces those omissions, not a change in training. Anyone using the models in workflows with external tools should know that the flaw exists and that the mitigation depends on the configuration they apply themselves.

Price is no longer measured in tokens

The race of the last two years consisted of having the most capable model and charging for usage. The proposal of this generation is that the customer pays per completed task and the provider makes sure that task consumes as little as possible, because the margin is no longer in the token but in efficiency.

That shift in messaging has an obvious commercial reading: if the competitor charges a fifth for an almost identical result, the difference stops being in the ranking and moves to the bottom line. The pattern appears in every launch this week.

Alibaba's model arrived with 2.4 trillion parameters and the same promise of capability at lower cost, and OpenAI's comparisons are now made against a specific rival by name. Competition has moved from the leaderboard to the monthly bill.

For a small team, the decision is less heroic than it seems. Astra is for the deliverable you defend before a client, and Sol for the volume work that nobody audits. It remains to be seen how much of the improvement survives if the customer changes the template, the language or the sector.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
US imposes licensing on chip rentals and tightens its security

Keep reading