GPT-6 Astra draws ducks on some accounts, a test finds

According to an informal investigation published on Pages, if you ask Codex for the same SVG pelican dozens of times, about half the drawings come out polished. The rest are crude cartoon birds that an image classifier mostly took for ducks or gulls, and which one you got depended on the ChatGPT account. The prompt, the settings, the client version and the time of day did not explain it.

At a glance

• On affected accounts, Codex responses labelled gpt-6-astra run at a different speed, come out shorter and draw and write like gpt-5.6-luna, while the server keeps reporting gpt-6-astra.
• In the authors' data the pattern began abruptly on 2026-09-20 around 09:00 UTC. Affected accounts produced about 25% less output per response, and users interrupted turns roughly 20 times as often.
• The authors cannot see OpenAI's servers, so they cannot say why it happens or confirm the model, and all their data comes from one team using many ChatGPT Pro accounts.

If you have not been following, the pelican is Simon Willison's long-running informal benchmark: an SVG of a pelican riding a bicycle, run across dozens of models going back to GPT-3.5 Turbo because it exposes differences that are otherwise hard to see. According to Simon Willison's Weblog, OpenAI released GPT-6 Astra for developers on 5th September 2026. Willison soon compared Astra's pelicans at several reasoning levels against GPT-5.6 Sol, Terra and Luna.

No account showed the pattern before September 20, and afterwards more than half did

The authors date the change to 2026-09-20, around 09:00 UTC. Before that day none of their accounts were affected. Of the accounts used afterwards, more than half were, a split with a Fisher test p-value of 2·10⁻¹². An account counts as affected when at least 30% of its long responses are anomalous, and the authors say every account falls clearly on one side or the other.

The effect stays with the account. The first half of an account's data predicts its second half with a correlation of ρ = 0.79, yet two accounts used in the same hour can behave differently. The authors ruled out reasoning effort, quota used, time since the account was first used, Codex version and settings, and the prompt and context.

Speed tells the same story. Normal gpt-6-astra serving tops out around 34 tokens per second, and in the two weeks before September 20 essentially no long response went faster than that. After that date some did every day: 2.8% of long gpt-6-astra responses on September 20, 9.7% on the 21st, 2.8% on the 22nd and 8.0% on the 23rd.

Affected accounts produced about 25% less output per response

The pelican drawings were scored with CLIP, an off-the-shelf image model, by multiplying the score for "is it a pelican" by the score for "is it a good illustration". The scores fell into two groups with a clear gap between them.

The effect reaches real work too. On affected accounts, each response carried about 25% less output (p = 3·10⁻⁵), and users interrupted turns roughly 20 times as often (p = 0.009). On a 12-question text benchmark, affected answers were less Astra-like on all 12 questions (Wilcoxon p = 0.0005).

The authors could also reproduce the effect on demand. They kept the same machine setup, prompt and settings and changed only the account. Every run was normal on one account and most runs were anomalous on the other. They rate this result as significant, with the caveat that it rests on one account each.

The closest match is gpt-5.6-luna, a finding the authors rate only as suggestive

For the reference runs, the authors sent the same prompts to each available model, naming it explicitly, and compared drawing style, phrasing and the 12-question benchmark. By drawing style, phrasing and speed, gpt-5.6-luna came closest. But they picked Luna as the comparison after a preview, so they rate that finding (p = 0.016) as suggestive, not strong.

The two tiers are priced very differently. According to AI Weekly, list prices per million tokens are $10 input and $50 output for Astra, against $0.20 and $1.20 for Luna. The same outlet reports that at max reasoning effort one Astra pelican SVG cost 63.21 cents on 12,638 output tokens, while the same prompt on GPT-5.6 Luna cost 1.57 cents on 13,040.

The check script times only the answer, so a slow connection does not matter

The script, is_my_astra_fake.py, needs a logged-in codex CLI and Python 3.8+ with only the standard library. It sends 3–9 short gpt-6-astra requests through your own Codex, which uses a little of your quota. It then reads the timing from Codex's local session logs. It sends nothing anywhere else, and the account line is only printed to your terminal.

What it times is the key design choice: only the answer text that streams after reasoning ends, so network, queueing and startup delays drop out. It is like timing a sprinter from the first stride, not from the moment you left home to watch. Real gpt-6-astra streamed at 32.5–33.8 tokens per second in every run, so anything outside 30.5–36.5 counts as anomalous.

The script runs up to three rounds of three parallel requests. YES (exit code 1) means at least two of three runs in a round streamed clearly faster or slower than real gpt-6-astra. NO (exit 0) means three rounds looked normal. UNKNOWN (exit 2) means Codex failed or too few runs gave a clear measurement.

The data fits a per-account experiment or routing cohort, an account-level policy, or capacity routing, and the authors cannot tell these apart from the client side. In their live side-by-side test the affected account also ran in a container while the normal one ran on the host. In our view a NO from the script is weaker than it looks, because affected accounts are intermittent and one round of three on an affected account looked normal.

Whether single-account users see it

The open question is whether ordinary users on a single account see this at all. That is the main reason the authors released the script. They ask anyone who posts a result to share only YES or NO, the date and whether they use a single account, with no emails or session logs. The source does not say why the change began on September 20 or whether it will stop.