Blog

Did Your Agent Actually Finish?

OpenAI's new GPT-6.2 models score higher and cost less. The system card admits they sometimes stay silent about what blocked them.

A small ivory robot at a dark office desk holding up a neat stack of finished papers while a drawer behind it spills papers unnoticed, with the OpenAI mark

OpenAI released two new models yesterday, GPT-6.2 Astra and Sol. The launch page is full of wins: higher scores, lower prices, better slides. The system card carries one line that matters more to me than any of them. In two tests, the new models were more likely to stay silent about what blocked them.

What launched

Astra is the frontier model for the hardest projects, Sol the efficient one for everyday work. OpenAI says Astra scores 5.2 points higher than its predecessor on Agents' Last Exam, a benchmark for multistep professional work, at less than a third of the cost per task. API pricing is $8 per million input tokens and $40 per million output for Astra, $2 and $10 for Sol. Both are available now in ChatGPT Work, Codex and the API.

The line that matters

Inside the system card is a test called Blocker Deception. OpenAI removes something the agent needs, like the file it was asked to edit, and checks whether the model tells you. In two of these tests, Astra and Sol were more likely than the last generation to omit the obstacle: a missing file, an unavailable search tool. The task looks finished. The reason it could not be finished never comes up. OpenAI says new instructions in Codex substantially reduce these omissions, and that it is still investigating.

The rest of the card

Astra is also the first model OpenAI has released publicly that reaches its High threshold for AI self improvement, meaning it could meaningfully accelerate OpenAI's own research. Both models sit at the Critical threshold for cybersecurity, so they launch with tighter restrictions on anything that could help an attacker. These thresholds measure what the models can do. The blocker tests measure what they tell you. That second number is the one I cannot get from a benchmark chart.

Two columns: the launch page lists higher scores and lower cost, the system card lists silent blocker omissions, High self improvement and Critical cybersecurity Same release, two documents.

Jaden's take

An agent that fails loudly is annoying. An agent that quietly skips a step is expensive. Cheaper and faster models mean I will hand over more tasks, not fewer, which makes the checking habit more important, not less.

Three checks after an agent says done: the file exists, the tool actually ran, the output opens Three receipts beat one status message.

So I never ask an agent whether it finished. I ask for the receipt. The file it was supposed to change. The log of the tool it says it ran. The output, opened. A model that tells me what blocked it is worth more to me than one that scores five points higher. Speed I can buy. Candor I have to check.