A developer has analysed 390,195 Codex token-count records across 865 sessions and found that GPT-5.5 disproportionately terminates its reasoning at exactly 516 tokens — a pattern that grew from 0.11% of longer runs in February to 53.3% in May, with a corresponding collapse in mean reasoning-token intensity from 268 to 107 over the same period.
The finding, documented in detail on GitHub, landed on the Hacker News front page with 127 points and 40 comments in three hours. The reporter is careful about what they are claiming: not proof of hidden chain-of-thought truncation, but an aggregate clustering anomaly that looks consistent with a thresholded reasoning-budget behaviour — and one that is getting dramatically more common.
🔍 THE BOTTOM LINE
GPT-5.5 accounts for only 19.3% of Codex responses but 82% of all exact-516 reasoning-token events. Its ratio of exact-516 to ≥516 responses is 44%, versus 1.3% for every other model combined. That 33.6× gap is not a natural distribution. Something inside GPT-5.5 is making it stop thinking at a fixed boundary, and the pattern intensified precisely as mean reasoning-token counts fell by 60%.
What the Data Shows
The reporter, GitHub user vguptaa45, mined Codex telemetry metadata from February through June 2026. The aggregate picture is stark:
| Month | Exact-516 / ≥516 ratio | Mean reasoning tokens | P90 reasoning tokens |
|---|---|---|---|
| Feb 2026 | 0.11% | 268.1 | 772 |
| Mar 2026 | 2.45% | 256.8 | 723 |
| Apr 2026 | 4.25% | 228.7 | 669 |
| May 2026 | 53.30% | 106.9 | 344 |
| Jun 2026 | 35.84% | 168.5 | 515 |
Two things happened simultaneously: the model started thinking less on average, and a huge fraction of its longer runs started ending at exactly 516 tokens. Additional spikes appear at 1034 and 1552 — values the reporter notes “look like repeated threshold boundaries rather than a naturally varying reasoning-token distribution.”
The model-level breakdown confirms the pattern is GPT-5.5-specific. GPT-5.2’s exact-516 / ≥516 ratio is 0.34%. GPT-5.3-codex and GPT-5.3-codex-spark both sit at 0.0%. GPT-5.4 shows 19.8% — non-trivial, but still less than half of GPT-5.5’s 44%.
Why This Looks Suspicious
The anomaly is not simply higher reasoning-token usage. In fact, the opposite: mean and P90 reasoning-token intensity fell sharply from February to May, while exact-516 clustering rose. If the model were simply thinking more, you’d expect a broad rightward shift in the distribution. Instead, the distribution is collapsing onto a small set of fixed values.
The reporter links this to a prior issue (#29353) in which a user reproduced a case where GPT-5.5 runs ending at exactly 516 reasoning tokens returned the wrong answer on a complex task. That earlier report was anecdotal — a single task-level reproduction. The new issue adds aggregate evidence across 390,000 records and five months.
The framing is careful: “I am not claiming this proves hidden chain-of-thought truncation. The narrower claim is that Codex telemetry shows a GPT-5.5-specific fixed-token clustering anomaly that looks consistent with thresholded reasoning-budget behaviour.” That restraint is why the issue has attracted constructive engagement rather than dismissal.
The Quiet Part
If GPT-5.5 is hitting a reasoning-budget cap at 516 tokens on complex tasks, the user does not know that. Codex surfaces a final answer. The intermediate reasoning is hidden. A developer using Codex for a hard refactor would see a completed response and assume the model thought it through — when it may have hit a threshold and stopped early.
The timing is also uncomfortable. The clustering intensified in May, roughly when OpenAI was scaling GPT-5.5 across more Codex users. If a reasoning-budget throttle was introduced to manage compute costs, the side effect — degraded performance on the hardest tasks — would be invisible to the end user and would only surface in aggregate telemetry analysis like this one.
OpenAI has not yet responded on the issue. The ask from the reporter is straightforward: investigate whether GPT-5.5 has a reasoning-budget, routing, truncation, fallback, or scheduler behaviour that causes responses to terminate around 516/1034/1552 tokens, and if so, whether exact-516 indicates a normal stopping point, a budget cap, a degraded tier, or another internal threshold.
The NZ Angle
New Zealand developers using Codex on the gpt-5.5 tier are paying the same per-token rate as users everywhere else. If a portion of those tokens are being spent on a model that silently short-circuits its reasoning on complex tasks, the value equation shifts. For local engineering teams evaluating Codex against alternatives like GLM-5.2’s agentic coding harness or Kimi K2.7 Code on GitHub Copilot, the question is not just “which model is smarter” but “which model actually uses the reasoning budget it claims to.”
This also matters for the NZ sovereign AI infrastructure conversation. If frontier models are quietly degrading reasoning depth to manage cost, the case for local inference — where you control the full reasoning pipeline — gets stronger, not weaker.
❓ FAQ
What is reasoning-token clustering? When a model’s internal chain-of-thought terminates at the same token count with unusual frequency, rather than varying naturally with task complexity. A healthy distribution looks like a smooth curve. A clustering anomaly looks like spikes at specific values.
Does this mean GPT-5.5 is broken? Not necessarily. The reporter is careful not to claim that. The finding is that the pattern exists, is model-specific, is getting worse, and correlates with reduced reasoning depth. Whether it causes degraded performance on real tasks is the open question — but the prior issue (#29353) showed at least one case where an exact-516 run returned a wrong answer.
Why 516 specifically? Unknown. The values 516, 1034, and 1552 are roughly multiples of 518, which could suggest an internal buffer or batch boundary. OpenAI has not commented. The reporter suggests these “look like repeated threshold boundaries.”
Should I stop using GPT-5.5 on Codex? No. For the majority of tasks — which don’t max out the reasoning budget — GPT-5.5 works fine. The concern is specifically about complex, high-stakes tasks where the model might hit a threshold and stop early without you knowing. If your task is hard enough that reasoning depth matters, monitor the token_count metadata yourself.
What should OpenAI do? Respond on the issue. Either confirm the threshold and explain its purpose, or investigate whether it’s an unintended side effect of a cost-management change. The reporter provides clear validation queries. Silence is the worst option — it lets the pattern harden into a trust deficit.
🔍 THE BOTTOM LINE
390,000 records don’t lie about the shape of a distribution. GPT-5.5 is clustering its reasoning at fixed thresholds, the clustering intensified as overall reasoning depth fell, and the user has no way to know when it happened. OpenAI needs to explain what 516 means — because right now, the most charitable read is “a cost throttle,” and the least charitable is “the model is quietly thinking less and charging the same.”
