Introduction
Over the past few days, developers using OpenAI Codex have been discussing a strange pattern: some GPT-5.5 responses appear to stop around a very specific reasoning-token count — 516.
The original report does not prove that OpenAI is secretly cutting off reasoning. The narrower claim is more careful: telemetry shared in a public GitHub issue shows an unusual GPT-5.5-specific clustering pattern around reasoning_output_tokens = 516, with additional spikes near 1034 and 1552.
That detail matters because developers are not only asking whether GPT-5.5 is making mistakes. They are asking whether the model is sometimes taking a shorter reasoning path on complex tasks, then returning answers that feel less reliable than expected.
Source Note
- Original article: BAAI Hub article
- Original referenced GitHub issue: openai/codex issue #30364
- Related reproduction issue: openai/codex issue #29353
- The BAAI page states that the article was sourced from a WeChat article by Xinzhiyuan.
- The original article did not contain code blocks. It mainly contained screenshots, tables, and linked references.
- Image note: relevant screenshots, charts, and comparison images are preserved below. Decorative separators, brand-only images, QR codes, recruitment graphics, and unrelated promotional images were not included.
GPT-5.5 Gets Stuck at “516”: Developers Notice an Unusual Pattern
The discussion started with a surprisingly specific number: 516.
According to developer reports, GPT-5.5 has recently shown weaker performance on some complex programming and reasoning tasks inside Codex. What made the issue stand out was not just that the model sometimes produced wrong answers. It was that several failed or suspicious responses appeared to stop at the same reasoning-token boundary.
A number of Codex users then joined the discussion, saying they had seen similar behavior.
The central question is simple: why would a top-tier reasoning model repeatedly land on one exact token count?
The GitHub Issue: A Larger Data Window Behind the Claim
The most important public reference is GitHub issue #30364, opened in the openai/codex repository.
In that issue, the developer reported an aggregate pattern in Codex token_count metadata. The claim was that gpt-5.5 responses disproportionately landed at exactly reasoning_output_tokens = 516, with additional fixed-boundary spikes around 1034 and 1552.
The report covered a window from February 1 to June 27,
2026. It analyzed 390,195 response-level token records across 865 sessions.
Evidence Reported in Issue #30364
| Metric | Value |
|---|---|
| Response-level token records analyzed | 390,195 |
| Sessions represented | 865 |
Exact reasoning_output_tokens = 516 events | 3,363 |
| GPT-5.5 share of all responses | 19.3% |
| GPT-5.5 share of exact-516 events | 82.0% |
| GPT-5.5 exact-516 / >=516 ratio | 44.0% |
| Non-GPT-5.5 exact-516 / >=516 ratio | 1.3% |
The issue also compared GPT-5.5 with other GPT-family models. The gap was large enough to make developers suspect this was not just a normal distribution of reasoning lengths.
Model-Level Result
| Model | Response Records | Exact 516 / >=516 |
|---|---|---|
gpt-5.5 | 75,401 | 44.0% |
gpt-5.4 | 25,214 | 19.8% |
gpt-5.2 | 247,575 | 0.34% |
gpt-5.3-codex | 13,333 | 0.0% |
gpt-5.3-codex-spark | 26,179 | 0.0% |
The original article summarized the point sharply: GPT-5.5 accounted for only a minority of total responses, yet appeared to account for most exact-516 events in this dataset.
The More Suspicious Part: Overall Reasoning Intensity Fell
One possible defense would be that GPT-5.5 simply thinks more, so more responses naturally reach higher reasoning-token ranges.
But the reported data points in the opposite direction.
In May and June, when the exact-516 clustering reportedly became more obvious, GPT-5.5’s overall reasoning-token intensity dropped. Both the mean reasoning-token count and the P90 reasoning-token count were lower than earlier months.
Reasoning-Token Intensity by Month
| Month | Mean Reasoning Tokens | P90 Reasoning Tokens |
|---|---|---|
| Feb 2026 | 268.1 | 772 |
| Mar 2026 | 256.8 | 723 |
| Apr 2026 | 228.7 | 669 |
| May 2026 | 106.9 | 344 |
| Jun 2026 | 168.5 | 515 |
That is why developers found the pattern uncomfortable. On one side, exact-516 events became more frequent. On the other side, the model seemed to spend fewer reasoning tokens overall.
This led to a more serious concern: on complex or high-risk tasks, GPT-5.5 may sometimes be hitting a hidden reasoning budget, truncation point, fallback route, or scheduler behavior before it has completed a deeper reasoning path.
To be clear, this remains a developer-reported anomaly, not an official explanation from OpenAI.
GitHub Developers Push Back
The GitHub discussion quickly attracted other users who said they had experienced similar problems.
The issue also connects to an earlier report, #29353, where a developer described a reproducible pattern in Codex Desktop using gpt-5.5 with xhigh reasoning.
In that reproduction, some fresh runs went directly to a final answer, used exactly 516 reasoning output tokens, and returned the wrong answer. Other runs spent thousands of reasoning tokens, produced a visible intermediate phase, and returned the expected answer.
This earlier issue did not settle the question, but it made the later aggregate report feel less isolated.
Reddit and Hacker News Discussions Add More Pressure
The conversation moved beyond GitHub as well. Screenshots in the original article show developers discussing whether a portion of high-risk Codex requests might be silently degraded due to reasoning truncation.
One comment shown in the article argued that some reasoning problems need 6,000 to 8,000 thinking tokens before the correct answer appears. If a model stops around 516 tokens on those cases, it may produce an answer too early.
Another screenshot showed users comparing Codex and Claude, with some saying they switch tools depending on which one feels less broken on a given week.
What Developers Want OpenAI to Clarify
The community’s core request is not complicated. Developers want OpenAI or the Codex team to clarify what is happening around 516, 1034, and 1552.
The open questions include:
幾分鐘搭建展示站並增長獲客
輸入一句想法，We0 AI 即可生成展示站、頁面與 CMS。發佈上線後並幫你獲取客戶和流量。
- Is this caused by a reasoning budget?
- Is it a routing issue?
- Is it a truncation or streaming behavior?
- Is it a fallback path?
- Is a scheduler or backend system creating fixed stopping boundaries?
- Is exact
516a normal stop point, a degraded tier, or an internal threshold?
The original GitHub issue is careful not to claim that it proves hidden chain-of-thought truncation. The stronger claim is simply that the clustering looks model-specific and threshold-like enough to deserve investigation.
Not Just Less Reliable: Users Also Complain About Personality
The second half of the original article shifts from performance to personality.
A developer named Angel compared ChatGPT using GPT-5.5 Instant with Claude Fable 5 using side-by-side screenshots. The complaint was not about whether the model could answer. It was about how the assistant behaved.
The article highlights three recurring frustrations.
Problem 1: Everything Becomes a Bullet List
The first complaint is that ChatGPT tends to over-format even simple conversational responses.
When asked to be more natural and less AI-like, ChatGPT reportedly responded with a structured explanation of how it would be natural. Claude, by contrast, gave a much shorter and more casual response.
The issue is not that bullet points are always bad. They are useful in technical explanations. The problem is that a chat assistant can feel stiff when it turns every small request into headings, bold text, lists, and follow-up suggestions.
Problem 2: It Always Wants to Correct Something
The second complaint is about over-editing.
When asked to check a sentence or a tweet, ChatGPT often tries to improve it, rewrite it, or offer alternatives, even when the user may only want a simple “this is fine.”
The original article contrasts this with Claude-style behavior, where the assistant more often says the text is acceptable as-is when no real fix is needed.
For users who write casually, this can create friction. A model that always “fixes” the user may feel less like a helper and more like a strict editor.
Problem 3: You Ask for One, It Gives Three
The third complaint is that ChatGPT often gives more than requested.
In the article’s example, a user asks for a joke. ChatGPT gives one joke, adds another, adds a third, and then asks for the user’s preferred style of humor. Claude gives a shorter response.
This is a subtle product problem. More output can look helpful in a benchmark, but in a conversation it can feel like the assistant is not listening.
Why This Matters for AI Assistants
For a coding agent, reliability matters. If a model stops early, takes a shorter reasoning path, or falls into a fixed-token boundary on difficult tasks, developers need to know. They rely on these tools for debugging, code review, architecture decisions, and production changes.
For a chat assistant, personality matters too. If every answer is over-formatted, over-corrective, or over-supplied with options, the user experience gets heavier over time.
The article’s broader argument is that both problems point to the same product risk: an assistant can become good at “delivering an answer” while becoming worse at actually helping the user in the moment.
FAQ
What is the GPT-5.5 516 reasoning-token issue?
It refers to developer reports that some GPT-5.5 Codex responses appear to cluster at exactly reasoning_output_tokens = 516. The main public reference is GitHub issue #30364 in the openai/codex repository. The issue claims this pattern is much stronger for GPT-5.5 than for several other models.
Does this prove OpenAI is secretly truncating GPT-5.5 reasoning?
No. The GitHub issue itself says it does not prove hidden chain-of-thought truncation. The safer conclusion is that the reported data shows an unusual fixed-token clustering pattern that may be consistent with thresholded reasoning-budget behavior.
Why is the number 516 important?
The number matters because repeated exact stopping points can look less natural than a normal spread of reasoning lengths. In the reported dataset, 516, 1034, and 1552 appeared as fixed-boundary spikes. Developers are asking whether these are caused by a budget, routing behavior, fallback path, or another backend mechanism.
What is OpenAI Codex used for?
OpenAI Codex is a coding agent for software development. According to OpenAI’s developer documentation, Codex can help read codebases, edit files, fix bugs, review code, and work on software tasks across local or cloud environments.
Is the 516 issue specific to Codex?
The public discussion focuses mainly on Codex and Codex Desktop metadata. The strongest claims in the article are tied to Codex token_count data and GitHub issues in the openai/codex repository. It should not be generalized to every ChatGPT or OpenAI API use case without separate evidence.
Why does the article compare ChatGPT with Claude?
The article uses the comparison to discuss assistant “personality,” not only raw reasoning performance. Screenshots show complaints that ChatGPT can be overly structured, overly corrective, and too eager to provide multiple options, while Claude-style responses are sometimes shorter and more conversational.
What should developers do if they see similar Codex behavior?
Developers should save reproducible examples, metadata, timestamps, model settings, and task prompts where possible. A clear report with token counts, expected behavior, actual behavior, and reproduction steps is more useful than a vague complaint.
Related Tools
- OpenAI Codex: OpenAI’s coding agent for reading, editing, debugging, and reviewing code.
- Codex CLI: A local terminal version of Codex for working with code directly on your machine.
- Codex Web: OpenAI’s cloud Codex environment for delegating coding tasks in the background.
- GitHub Issues: GitHub’s built-in system for tracking bugs, feedback, tasks, and technical discussions.
- ChatGPT Custom Instructions: OpenAI’s official feature for guiding ChatGPT’s style and behavior.
- Claude: Anthropic’s AI assistant, referenced in the article’s personality comparison.
- Claude Code: Anthropic’s coding agent for working with codebases, files, commands, and development tools.
Related Links
- Original BAAI Hub Article: The republished Chinese article that this English version is based on.
- GitHub Issue #30364: The main public issue reporting GPT-5.5 reasoning-token clustering around 516, 1034, and 1552.
- GitHub Issue #29353: A related reproduction report involving
gpt-5.5,xhighreasoning, and exact 516 reasoning tokens. - OpenAI Codex Developer Documentation: Official OpenAI documentation for Codex.
- OpenAI Codex CLI Documentation: Official documentation for running Codex locally from the terminal.
- OpenAI Codex Web Documentation: Official documentation for using Codex in the cloud.
- GitHub Docs: About Issues: Official GitHub documentation explaining how issues are used to track bugs and discussions.
- OpenAI ChatGPT Custom Instructions Help: Official guidance for customizing ChatGPT responses.
Summary
This article explains the developer-reported GPT-5.5 Codex 516 reasoning-token anomaly, including the main GitHub issue, the reported aggregate data, and the concern that some complex tasks may be ending too early.
It also covers a second user-experience complaint: ChatGPT’s tendency to over-format, over-correct, and over-answer simple requests. That part of the discussion is more subjective, but it matters because assistant personality directly affects everyday product experience.
The important point is not to treat the 516 pattern as proven hidden truncation. The public evidence is better understood as a model-behavior anomaly that deserves investigation.
For developers, the practical takeaway is simple: when an AI coding agent feels suddenly worse, collect metadata, compare runs, and report reproducible patterns instead of relying only on impressions.
