GPT-5.5 Codex Reasoning-Token Clustering and Performance Degradation

Reasoning-Token Clustering Causes Performance Drops in GPT-5.5

GPT-5.5 Codex is exhibiting a phenomenon where reasoning_output_tokens cluster at fixed values, specifically starting at 516 and increasing in increments of 518. This clustering is strongly correlated with "short-circuiting," where the model stops reasoning prematurely and returns incorrect results for complex logic puzzles and coding tasks.

This behavior is specific to GPT-5.5; it is significantly less prevalent in GPT-5.4 and almost entirely absent in versions 5.2 and 5.3. The issue suggests a throughput optimization or batching mechanism on the server side that is inadvertently capping the model's reasoning effort.

Evidence of Token Clustering

Technical analysis of Codex session logs reveals a distinct statistical spike in reasoning tokens at specific thresholds. Users have identified a pattern where tokens cluster at $516 + 518n$ (where $n$ is an integer).

Observed Token Distribution

Data shared by users indicates a high frequency of responses hitting these exact limits:

516 tokens: The most common spike, often associated with incorrect answers.
1034 tokens: ($516 + 518$)
1552 tokens: ($516 + 2 \times 518$)
2070 tokens: ($516 + 3 \times 518$)

One user hypothesized that the 516-token mark represents an initial 512-byte buffer plus a 4-byte header, with subsequent increments representing additional buffers and linked list references.

Impact on Model Accuracy

When the model "short-circuits" at these token thresholds, the quality of the output degrades significantly. In contrast, when the model is allowed to use a larger number of reasoning tokens (e.g., 6,000 to 8,000), it consistently returns the correct result for the same prompts.

Case Study: The Candy Puzzle

Users tested the model with a complex logic prompt involving counting candies of different shapes and flavors. The results demonstrated the instability of the current version:

Failed Runs: Multiple attempts resulted in exactly 516 reasoning tokens and incorrect answers (e.g., results of 24, 27, 12, and 21).
Successful Runs: When the model exceeded the clustering threshold, it provided the correct reasoning and final number.

User Experience and Community Response

Developers have reported a noticeable decline in the reliability of GPT-5.5 compared to earlier versions. Some users have noted that the "outstandingly thorough coding" seen earlier in the year has been replaced by "incredibly stupid implementations" appearing intermittently.

Key Community Insights

"Out of 10 runs of the exact same prompt, 4/10 had this 516 thinking token issue, and every one of these had the wrong solution."

"I’ve definitely experienced step jumps down in quality on an almost daily basis... I’m seeing incredibly stupid implementations intermittently."

Technical Concerns and Transparency

There is significant frustration regarding the "black box" nature of the reasoning process in GPT-5.5. Because reasoning contents are encrypted or obfuscated, users cannot see the internal chain of thought, making it difficult to diagnose why a model failed until they analyze the token counts in the metadata.

Furthermore, some users expressed concern that these regressions are the result of silent server-side optimizations intended to reduce compute costs, noting that similar issues were previously reported on GitHub but closed without public acknowledgement or resolution.