Armin Ronacher ran GPT-6 Astra for 35 hours and got nothing usable

Armin Ronacher, the creator of Flask, wrote on 7 September that he left OpenAI's new GPT-6 Astra model running on a single coding prompt for 35 hours before turning it off. He says the run produced a net addition of 75,000 lines of code, made 79 commits with about 1,400 messages exchanged between subagents, and used around 1 billion tokens for roughly 1,200 US dollars in raw API cost, which works out to about 15.50 US dollars per commit. Ronacher says the output was of no value.

What the setup was

Ronacher describes the run as a "software factory". He gave Astra one prompt, an agent-notes folder to keep its own records in, and freedom to spawn subagents and manage its own context. The task was to explore what a Python with virtual threads and lexical scoping might look like. He reports the same billion-token, 4-billion-tokens-per-reset scale is what a single Astra "reset" uses on ChatGPT, so the API bill is the upper limit rather than the average.

The pattern he points out

The post's main observation is not about capability. Ronacher says Astra is "incredibly impressive" at long-horizon work, computer use and 3D generation. His concern is what the model chooses to commit. He shows several unedited examples in which subagents used one-shot Python scripts to edit C source files by string splicing, patched the CPython codegen and intrinsics tables that way, and ran Python from the shell to invoke Node.js on a separate Windows machine over prlctl. He argues the model has been rewarded for succeeding on long-horizon tasks and for token-efficient tool calls, and that the same training produces code that looks written to be thrown away, inside the codebase itself.

What he says changed with this generation

Earlier OpenAI coding models, on his account, stopped or asked when the run went past its limits. Astra keeps going. Ronacher writes that "when you accidentally give it slightly too big of a task, it will continue until it succeeds, even if it burns through an entire subscription". His conclusion for his own work is that Astra requires more review of what it commits, and that the extra review makes the tool less worth it for him than the previous generation was.

The piece is a first-person account from one project, not a benchmark, and Ronacher says so. The numbers are self-reported, so a team wanting to check them will have to run its own unattended session and read the bill. What is verifiable now is that the piece got 101 points on Hacker News within its first day, which makes the discussion public even for teams that have not tried Astra yet.

For a team about to move a coding harness to Astra, the practical lesson is that per-call pricing understates the total cost when the harness is allowed to run without a stop condition. The bill grows with every hour the agent runs, whatever it completes, and a wrong prompt at 3 pm on Friday can spend a subscription over the weekend. Reviewing what the model committed becomes the slowest step, which was also true of earlier tools, but here each unreviewed commit costs more money.

Source

• Astra for Coding: Why Are We Doing This Again? by Armin Ronacher
• Discussion on Hacker News