Why Codex’s GPT‑5.5 Suddenly Returns Short Answers: The Hidden 516 Token Bug

Recently a heated issue appeared in the Codex community: when using the gpt-5.5 model with the xhigh setting, some complex tasks finish in a few seconds and return an answer immediately, and many local logs contain the same token count 516.

What happened

The author collected Codex telemetry from 2026‑02‑01 to 2026‑06‑27, analysing 390,195 response‑level token records across 865 sessions. Key numbers include:

Total response records: 390,195

Sessions: 865

Events where the token count equals 516: 3,363 gpt-5.5 accounts for 19.3% of all responses

Among the 516 events, gpt-5.5 appears 82.0% of the time

44.0% of gpt-5.5 responses have a token count ≥ 516, compared with only 1.3% for other models

A second table shows that gpt-5.5 has 44.0% of its responses at ≥ 516 tokens, while gpt-5.4 is 19.8% and the other variants ( gpt-5.2, gpt-5.3‑codex, gpt-5.3‑codex‑spark) are near zero.

Why 516 is striking

The clustering of fixed token lengths (516, 1034, 1552) is unlikely to be natural variation; the author suggests it resembles a threshold‑based inference‑budget behavior specific to gpt-5.5.

In May 2026 the proportion of 516‑clustered responses rose to 53.30%, while the average inference token count dropped to 106.9 and the P90 to 344, giving the impression of a faster model but actually shortening reasoning on complex tasks.

The candy‑question case study

The issue also links to an earlier reproducibility experiment: a “candy” puzzle where the correct answer is 21. Some runs with gpt-5.5+xhigh finish in seconds and output 29, with the log snippet:

reasoning_output_tokens: 516
phases: ["final_answer"]

Successful runs show a short progress hint, a much larger token count (often thousands), and finally the correct answer 21. This contrast demonstrates that the fixed‑token shortcut leads to wrong answers on short paths while longer reasoning yields correct results.

Community replication

The issue now has over 100 comments. Users reproducing the bug with codex-cli 0.142.5 see only 3 correct answers out of 9 attempts, and all failures occur at 516 or 1034 tokens. Other logs report 516 proportions of 50.1% and 47.9% on different machines.

Further analysis reveals additional fixed lengths (1034, 1552, 2070, 2588, 3106) spaced by 518, a pattern the community calls 518n‑2.

Temporary work‑arounds

Prompt tricks such as “think slowly”, “spend at least 60 seconds”, or “fully verify complex tasks”. Some users report mitigation, others see limited effect.

Removing the “Intermediary updates” instruction (the command that forces periodic progress hints). After removal, a user got 20 correct answers on the candy test and no 516/1034/1552 occurrences, though the fix may not hold in real‑world coding tasks.

Local proxy or plugin that detects the 518n‑2 stop point, forces an extra reasoning round, and merges the result. Hook plugins can warn users when a high‑risk round ends.

These solutions can help but are not guaranteed to eliminate the issue for all users.

Practical recommendations

For heavy Codex users the author suggests a four‑step approach:

Never trust the first answer on complex tasks (architecture design, database migration, permission changes, payment flow, security fixes). If the response finishes in seconds without visible exploration, treat it with suspicion.

Require explicit verification: ask the model to run tests, read logs, reproduce errors, or compare outputs rather than just “explain”.

Cross‑check critical results with another model (e.g., gpt-5.3‑codex, gpt-5.3‑codex‑spark, or gpt-5.4), which do not exhibit the 516 clustering.

Treat the xhigh setting as a tendency, not a guarantee of deeper reasoning.

Why it matters

The issue reveals a hidden runtime strategy—routing, budgeting, caching, scheduling, front‑end streaming, tool protocols, and base instructions—that can silently affect output quality. Users cannot tell whether a higher‑tier model actually performed more reasoning, which undermines trust.

The author calls on OpenAI to (1) clarify what the fixed values (516, 1034, 1552, 2070, …) represent, and (2) provide a transparent quality signal for complex tasks, even if internal reasoning remains private.