Report: GPT-5.5 Codex reasoning-token clustering may be leading to degraded performance

{ "title": "GPT-5.5 Codex Anomaly: Token Clustering May Degrade Performance", "slug": "gpt-5-5-codex-reasoning-token-clustering-degraded-performance", "summary": "A GitHub issue reveals that GPT-5.5 responses in Codex disproportionately cluster at exactly 516 reasoning tokens, with secondary spikes at 1034 and 1552. Analysis of 390,195 token records shows GPT-5.5 accounts for 82% of these fixed-boundary events despite only 19.3% of total responses. This anomaly coincides with a sharp decline in mean reasoning-token intensity, from 268.1 in February 2026 to 106.9 in May 2026, potentially explaining degraded performance on complex tasks. The pattern suggests a possible reasoning-budget cap or truncation behavior unique to GPT-5.5, raising concerns about model reliability for high-stakes applications.", "meta_description": "GPT-5.5 Codex responses cluster at exactly 516 reasoning tokens, with secondary spikes at 1034 and 1552. Analysis of 390,195 records shows this anomaly may degrade complex task performance.", "content": "

TL;DR

A detailed analysis of 390,195 Codex response records reveals that OpenAI's GPT-5.5 model exhibits a striking anomaly: its reasoning token outputs disproportionately cluster at exactly 516 tokens, with secondary spikes at 1034 and 1552. This pattern, which accounts for 82% of all exact-516 events despite GPT-5.5 representing only 19.3% of total responses, coincides with a sharp decline in mean reasoning-token intensity—from 268.1 in February 2026 to just 106.9 in May 2026. The evidence suggests a possible reasoning-budget cap or truncation behavior that may explain degraded performance on complex, high-stakes Codex tasks.

The Anomaly: Fixed-Boundary Token Clustering

A GitHub issue filed by user @vguptaa45 has brought to light a peculiar pattern in OpenAI's Codex telemetry. Analysis of 390,195 response-level token records across 865 sessions reveals that GPT-5.5 responses disproportionately land at exactly 516 reasoning output tokens, with additional fixed-boundary spikes at 1034 and 1552. This clustering is not a natural distribution—it suggests a thresholded reasoning-budget behavior.

The data is stark. GPT-5.5 accounts for only 19.3% of all responses but 82.0% of exact-516 events. Its exact-516 to >=516 ratio is 44.0%, compared to just 1.3% for non-GPT-5.5 models. Secondary models like GPT-5.4 show a 19.8% ratio, while GPT-5.2, GPT-5.3-codex, and GPT-5.3-codex-spark exhibit ratios of 0.34%, 0.0%, and 0.0% respectively.

Declining Reasoning Intensity

The anomaly is not simply about higher token usage. In fact, mean reasoning-token intensity has dropped sharply over time. In February 2026, the mean was 268.1 tokens with a P90 of 772. By May 2026, the mean had fallen to 106.9 tokens, with a P90 of just 344. This decline coincides with a dramatic rise in exact-516 clustering, from 0.11% of >=516 events in February to 53.30% in May.

The fixed values themselves—516, 1034, and 1552—are suspicious. They appear to be repeated threshold boundaries rather than a naturally varying distribution. This pattern is consistent with a reasoning-budget cap, routing logic, or truncation behavior that is unique to GPT-5.5.

Context: The Broader AI Model Landscape

This anomaly emerges at a time of intense competition in the AI model market. OpenAI recently unveiled GPT-5.6 Sol, its most advanced cybersecurity AI, alongside GPT-5.6 Terra and GPT-5.6 Luna for everyday and fast workloads. GPT-5.6 Sol is designed for high-intensity reasoning tasks, while Terra reportedly matches GPT-5.5 performance at half the cost. Meanwhile, Meta's upcoming Watermelon model is said to match GPT-5.5 on key benchmarks, according to Meta's superintelligence chief Alexandr Wang.

Anthropic has also been active, launching Claude Fable 5 with cybersecurity guardrails, while China's Z.ai claims its GLM-52 model can match Anthropic's Mythos on cybersecurity tasks. The competitive pressure is immense, and any performance degradation in a flagship model like GPT-5.5 could have significant implications for enterprise users and developers who rely on Codex for complex tasks.

Evidence and Analysis

The GitHub issue provides a detailed breakdown of the anomaly. Across 390,195 response-level token records from 865 sessions, there were 3,363 exact-516 events. GPT-5.5's share of these events is 82.0%, despite representing only 19.3% of all responses. The exact-516 to >=516 ratio for GPT-5.5 is 44.0%, compared to 1.3% for non-GPT-5.5 models—a 33.6x difference.

Monthly data shows the clustering intensified dramatically. In February 2026, only 0.11% of >=516 events were exact-516. By May 2026, that figure had jumped to 53.30%, before slightly declining to 35.84% in June. Simultaneously, mean reasoning tokens fell from 268.1 in February to 106.9 in May, with P90 tokens dropping from 772 to 344.

The fixed values—516, 1034, and 1552—are particularly telling. They appear to be repeated threshold boundaries, suggesting a budget cap or routing logic that truncates reasoning at these specific points. This is not a natural distribution; it is a model-specific artifact.

Why It Matters

For enterprise users and developers, this anomaly could have real-world consequences. Complex Codex tasks—such as code generation, debugging, and multi-step reasoning—require sufficient reasoning depth. If GPT-5.5 is consistently cutting off reasoning at 516 tokens, it may produce incomplete or incorrect outputs for high-stakes applications.

The timing is critical. OpenAI has just unveiled GPT-5.6 Sol, its most advanced cybersecurity AI, alongside GPT-5.6 Terra and GPT-5.6 Luna. GPT-5.6 Terra is reported to match GPT-5.5 performance at half the cost, making the GPT-5.5 anomaly a potential competitive liability. Meanwhile, Meta's Watermelon model is said to match GPT-5.5 on benchmarks, and Anthropic's Claude Fable 5 is gaining traction.

The issue also raises questions about model routing and efficiency. As noted in a Business Insider report, AI startup CTOs are increasingly practicing "modelmaxxing"—using specific models for specific tasks to avoid wasting tokens. If GPT-5.5 is consistently underperforming on complex tasks due to token clustering, it could undermine trust in the model for critical applications.

What OpenAI Should Investigate

The GitHub issue asks the Codex team to investigate whether GPT-5.5 has a reasoning-budget, routing, truncation, fallback, or scheduler behavior that causes responses to terminate around 516, 1034, or 1552 reasoning tokens. Useful internal validation checks include analyzing the ratio of exact-516 to >=516 events across models and time periods.

If this is expected behavior, OpenAI should clarify whether exact 516 indicates a normal stopping point, a budget cap, a degraded tier, or another internal threshold. Transparency is key for maintaining developer trust, especially as the company navigates government restrictions on GPT-5.6 Sol and competes with rivals like Meta and Anthropic.

Conclusion

The GPT-5.5 token clustering anomaly is a significant finding that warrants immediate investigation. With 82% of exact-516 events concentrated in one model, and a concurrent drop in reasoning-token intensity, the evidence points to a systemic issue that could undermine the model's effectiveness for complex tasks. As the AI industry races toward more powerful models, reliability and transparency remain paramount. OpenAI must address this anomaly to maintain its competitive edge and user trust.