A developer combing through Codex telemetry found something odd: GPT-5.5 responses keep terminating at exactly 516 reasoning tokens, far more often than chance would predict. According to Hacker News, where the finding was posted and quickly climbed to 186 points, this fixed-token clustering lines up with a drop in overall reasoning effort and may help explain why the model has felt weaker on complex, high-stakes coding tasks.
What stands out here is the precision. Reasoning-token counts are supposed to vary with how hard a task is. Instead, GPT-5.5 keeps landing on the same round numbers, with extra spikes near 1034 and 1552. Those look less like natural variation and more like repeated threshold boundaries.
What the researcher did
The analysis pulled from Codex token_count metadata across a February-to-June 2026 window. It’s aggregate, not anecdotal:
- 390,195 response-level token records analyzed
- 865 sessions represented
- 3,363 events landing at exactly 516 reasoning tokens
The method is simple and reproducible: count how often each model stops at exactly 516 tokens versus how often it uses 516 or more, then compare across model families and over time.
The numbers that raise eyebrows
GPT-5.5 makes up just 19.3% of all responses but 82.0% of the exact-516 events. Here’s how the models compare on the share of runs that stop at exactly 516 out of all runs reaching 516 or more:
- gpt-5.5: 44.0%
- gpt-5.4: 19.8%
- gpt-5.2: 0.34%
- gpt-5.3-codex: 0.0%
- gpt-5.3-codex-spark: 0.0%
That puts GPT-5.5 roughly 33x above the non-5.5 baseline. The clustering also grew fast month over month, jumping from 0.11% in February to 53.30% in May before settling at 35.84% in June.
At the same time, the model started thinking less on average:
- February: mean 268 reasoning tokens, P90 of 772
- April: mean 229, P90 of 669
- May: mean 107, P90 of 344
- June: mean 169, P90 of 515
So it’s not that GPT-5.5 is burning more reasoning and hitting a ceiling. It’s reasoning less overall and piling up at fixed stopping points. That combination is what makes the pattern look like a budget cap, a routing decision, or a degraded tier rather than random noise.
Why it matters for practitioners
The finding connects to an earlier bug report (#29353) where a GPT-5.5 run that stopped at exactly 516 tokens returned the wrong answer. If a hidden reasoning budget is cutting the model off mid-thought on hard problems, that’s a direct hit to reliability on exactly the tasks where you need the model to grind: complex refactors, multi-file debugging, tricky logic.
What you can do with this today:
- If GPT-5.5 gives shallow or wrong answers on a hard task, retry or route the task to GPT-5.2 or another variant that doesn’t show the clustering.
- Treat suspiciously short reasoning traces as a signal to double-check the output, not trust it.
- If you have access to your own telemetry, run the same check: count responses at exactly 516, 1034, and 1552 tokens by model and by day. It’s a five-minute query.
The limits of the claim
The author is careful, and that carefulness is worth repeating. This is not proof of hidden chain-of-thought truncation. The narrow claim is that Codex telemetry shows a GPT-5.5-specific fixed-token clustering anomaly consistent with thresholded reasoning-budget behavior. Correlation with degraded performance is suggestive, not confirmed. Only the Codex team can see whether 516 is a normal stopping point, a budget cap, a fallback tier, or something else.
The ask to the Codex team is reasonable: replay matched complex tasks across GPT-5.2 and GPT-5.5 with quality evals, and separate the exact-516 responses from the longer-reasoning ones to see if the short ones score worse.
This is significant because it shows how much you can learn from metadata alone. A single developer, no internal access, spotted a pattern across nearly 400,000 records that points at a real product question. Whether OpenAI confirms a budget mechanism or explains it away, the pressure is now public. Full breakdown and the raw tables are at the original source.