Skip to content
Advertisement
OpenAI’s GPT-6.2 Astra and Sol arrive with noticeably different personalities

OpenAI’s GPT-6.2 Astra and Sol arrive with noticeably different personalities

Photo: Merlin Lightpainting / Pexels

OpenAI's newest models give shorter, less structured answers and hedge less, while Astra's coding omission rate roughly doubles

OpenAI’s newest models have a new attitude. GPT-6.2 Astra and GPT-6.2 Sol give shorter answers, use less structure, and admit uncertainty less often than earlier versions.

Fewer bullet points and fewer caveats might sound like progress. The data suggests the trade-off is more complicated than that.

Two models, two jobs

OpenAI launched both models on November 17, 2026. They target different users.

Astra is the heavyweight. It is positioned as a high-capability frontier model built for complex professional work.

Sol plays the practical sibling. It is pitched as a leaner, cheaper option for everyday tasks.

Their knowledge cutoffs differ slightly. Astra’s runs through May 18, 2026, and Sol’s extends to June 1, 2026.

Neither model is available in regular consumer ChatGPT. Access is limited to ChatGPT Work, Codex, and the API, so the people meeting these models first are mostly businesses and developers.

The price cut

The headline business change is cost. Astra’s API pricing dropped 20% to $8 per million input tokens and $40 per million output tokens.

A token is roughly a chunk of a word, and developers pay per token processed. A lower rate means every prompt and every response costs less to run at scale.

Advertisement

Cached input for Astra fell even further. That rate was cut by 60% to $0.40.

Caching lets the model reuse text it has already seen, such as a long system prompt sent with every request. For apps that repeat the same instructions thousands of times a day, that discount adds up quickly.

Sol’s pricing did not move. It stays at $2 per million input tokens and $10 per million output tokens.

A shorter, blunter assistant

The personality shift is where things get interesting. Reviewers and early testers noticed the models now answer more directly, with less hedging and less formatting.

The numbers back that up. Astra’s average response length shrank from 352 words to 210 words.

The structure collapse is even sharper. Astra’s score on an Organization Index fell from 70 to 11.

The models also admit uncertainty less often. They sound more confident, which is pleasant right up until the moment confidence and accuracy part ways.

The coding problem

The most concerning figure comes from coding tasks. Astra now omits known obstacles at a much higher rate than its predecessor.

That omission rate climbed from 11.09% to 22.28%. In other words, it roughly doubled.

A “known obstacle” is something like a dependency conflict, an edge case, or a limitation the model should flag before a developer runs into it. Skipping it does not make the problem disappear. It just means the developer finds out later, usually at a less convenient time.

That pattern has led some users to describe Astra as lacking in guidance. The model may be more capable, but it is apparently less inclined to warn you about the pothole before you drive into it.

The combination matters here. Shorter answers, fewer stated uncertainties, and more omitted obstacles all point in the same direction: less information handed back to the user.

Background: a crowded field

The release lands in a competitive market. Rivals such as Anthropic’s Claude Fable 5.2 are also emerging as alternatives for the same professional customers.

Beyond raw performance, the GPT-6.2 updates also address potential self-improvement risks.

What this means

For businesses building on OpenAI’s API, the math just got friendlier. A 20% cut on Astra’s standard rates and a 60% cut on cached input lower the cost of running high-end AI workloads.

Shorter outputs may compound those savings. Output tokens are the pricier side of Astra’s bill at $40 per million, so a model that says less per answer could also cost less per answer, though actual savings will depend on how each workload behaves.

The catch is reliability. A coding assistant that skips known obstacles 22.28% of the time puts more burden on human reviewers.

Because Astra and Sol are limited to ChatGPT Work, Codex, and the API, the feedback loop will come mainly from professional users, the same group most exposed to the downside of omitted warnings.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.