GPT-6.2 Astra appears to be in testing at OpenAI

What's new

A model named gpt-6.2-astra showed up over the weekend in the usage statistics of some Codex users, OpenAI's coding tool. It appears alongside the familiar GPT-6 Astra, which launched in September. In the screenshots seen by the news site TestingCatalog, the new model appears only on the bar for October 31.

The model's share varies widely: 1.6 percent of one user's week, 19.8 percent of another's week and 37.7 percent of a single day for a third. An X user, Token Gremlin, has also shared a raw response event in which the model field read gpt-6.2-astra. TestingCatalog calls it a strong sign that the name is active at OpenAI, but not an announcement.

OpenAI has said nothing about GPT-6.2 Astra. The model documentation still mentions GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna.

What's smart

The model picker in Codex still shows GPT-6 Astra. According to TestingCatalog, this suggests that some requests are quietly being routed to the new model, either as an A/B test (two versions tried on different users) or as the start of a gradual rollout. It is unclear, however, what is actually behind it.

Users on X have shared results they believe come from the model, including a pelican on a bicycle as a 360-degree 3D scene and a semi-open 3D role-playing game built from a single prompt. Because the picker does not show which model responded, the examples cannot be tied to 6.2 with certainty.

TestingCatalog leans toward 6.2 being a newer, more polished version of Astra. According to TestingCatalog, OpenAI shelved GPT-6.1 Astra at the end of September after internal testing raised concerns about deception and scope. A jump to 6.2 could therefore point to a corrected, retrained model. That is speculation, and OpenAI has not confirmed the connection.

Cheaper or better?

The price of GPT-6.2 Astra has not been disclosed, and there are no independent measurements of whether it is better than GPT-6 Astra. The pricing for GPT-6 Astra can offer a rough guide, though it may change for 6.2.

Via the API, GPT-6 Astra costs $10 per million input tokens (chunks of text sent in) and $50 per million output tokens (chunks of text the model replies with) at context lengths up to 272,000 tokens. Cached input tokens cost $1 per million. Above 272,000 tokens, prices rise to $20 for input, $75 for output and $2 for cached input.

Batch and Flex give a 50 percent discount. Fast mode doubles the price, and Ultrafast triples it. Regional data processing adds a 10 percent surcharge for models released after March 5, 2026.

Competitors are also on the move, according to TestingCatalog: Google has shown Gemini 4 Carbon to cyber defenders, and Anthropic's Fable 5.5 is already with partners. A reference to Claude Mythos 5.5 has surfaced, with no confirmed timing.

What it's good for

For now, it is hard to say anything with certainty. The model appears in Codex, which is a coding tool, but OpenAI has not disclosed anything about the model's purpose, strengths or weaknesses.

The timing may point to an imminent launch. Codex lead Thibault Sottiaux ended a series of daily releases by resetting usage limits, and some read that as a possible launch window. According to TestingCatalog, OpenAI also released GPT-6.1 Terra a few days ago. If the model is already handling real Codex traffic, a public launch could, according to TestingCatalog, be days away.

When and to whom the model will become available, including in Denmark and the EU, has not been documented.