GPT-5.5 Codex Clustering: Pricing Impact
A new Codex issue reports GPT-5.5 reasoning-token clustering at 516/1034/1552. Here is the pricing impact for coding-agent buyers.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
A fast-moving GitHub issue in the OpenAI Codex repository is raising a pricing question that matters more than the token rate: what happens if an expensive reasoning model appears to spend less reasoning budget on complex tasks than buyers expect?
The report claims GPT-5.5 Codex responses are disproportionately clustering at exact reasoning-token values, especially 516, with related spikes around 1034 and 1552. The issue is still open, labeled bug, model-behavior, and rate-limits, and had 55 comments when this article was written.
This is not a confirmed OpenAI price change. GPT-5.5 remains listed in our live pricing data at $5.00 per million input tokens, $0.50 per million cached input tokens, and $30.00 per million output tokens. The pricing impact is about effective cost: if a premium Codex run fails more often, the cost per accepted patch, correct answer, or completed task rises quickly.
For live numbers, compare OpenAI pricing with Anthropic pricing and DeepSeek pricing, then model retry and review rates in the AI token calculator. For context, see our GPT-5.5 launch pricing article and GPT-5.5 vs GPT-5.4 pricing guide.
What changed
The alert comes from OpenAI Codex issue #30364, titled “GPT-5.5 Codex reasoning-token clustering at 516/1034/1552 may be leading to degraded performance on complex tasks.”
The original report says the author analyzed Codex token_count metadata from February 1 to June 27, 2026 UTC. The claim is not that hidden chain-of-thought truncation has been proven. The narrower claim is that Codex telemetry shows a GPT-5.5-specific fixed-token clustering anomaly.
Key figures from the report:
| Metric | Reported 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 model-level comparison is the most important pricing signal. The issue reports GPT-5.5 at a 44.0% exact-516 / >=516 ratio, GPT-5.4 at 19.8%, GPT-5.2 at 0.34%, and Codex-specific GPT-5.3 variants at 0.0%.
The issue also says the clustering became much more visible in May and June. Exact-516 / >=516 reportedly rose from 0.11% in February to 53.30% in May and 35.84% in June, while mean reasoning tokens fell from 268.1 in February to 106.9 in May and 168.5 in June.
That combination is why developers are paying attention: the reported pattern is not simply “GPT-5.5 uses more reasoning tokens.” It is the opposite concern. The issue alleges lower overall reasoning-token intensity alongside fixed stopping-like values.
Current pricing comparison
Here is the rate-card context for teams deciding whether to keep GPT-5.5 as the default Codex route.
| Model | Input | Cached input | Output | Pricing read |
|---|---|---|---|---|
| GPT-5.5 | $5.00 / 1M | $0.50 / 1M | $30.00 / 1M | Premium OpenAI flagship |
| GPT-5.4 | $2.50 / 1M | $0.25 / 1M | $15.00 / 1M | Half the GPT-5.5 token price |
| GPT-5.4 mini | $0.75 / 1M | $0.075 / 1M | $4.50 / 1M | Volume-friendly OpenAI fallback |
| Claude Opus 4.8 | $5.00 / 1M | $0.50 / 1M | $25.00 / 1M | Premium Claude comparison |
| Claude Sonnet 5 | $2.00 / 1M | $0.20 / 1M | $10.00 / 1M | Common coding-agent alternative |
| DeepSeek V4 Pro | $0.435 / 1M | $0.003625 / 1M | $0.87 / 1M | Low-cost benchmark route |
The direct GPT-5.5 versus GPT-5.4 comparison is stark: GPT-5.5 costs 2x GPT-5.4 on input, cached input, and output. That premium can be rational if GPT-5.5 reduces task failures, handles harder repository context, or saves senior engineering time.
But if a team sees degraded Codex behavior on complex tasks, the economics flip. A $30/M output model that requires repeated runs, manual cleanup, or fallback to GPT-5.4 can become more expensive than its price card suggests.
Why reasoning-token clustering affects real cost
Coding-agent buyers rarely pay for a single answer. They pay for a workflow:
- inspect the repository
- plan a change
- edit files
- run tests
- recover from failures
- explain the result
When that loop works, a premium model can be cheap even at high token prices because it replaces expensive human time. When that loop fails, the invoice is only part of the cost. The team also pays in retries, review time, broken trust, and delayed merges.
That is why the reported 516 clustering matters. If exact-516 runs are correlated with weaker task performance, buyers should not evaluate GPT-5.5 only by cost per million tokens. They should evaluate it by cost per accepted result.
The issue comments add pressure here. Several users report that GPT-5.5 fails simple reproductions at 516 reasoning tokens while GPT-5.4 succeeds on the same eval. Others posted aggregate telemetry with repeated values such as 516, 1034, 1552, and 2070. Treat those as community reports, not audited benchmark results, but relevant enough for paid Codex teams to investigate.
What this means for Codex buyers
The right response is not to abandon GPT-5.5 blindly. The right response is to stop treating GPT-5.5 as automatically cheaper just because it is more capable on paper.
For high-value coding work, GPT-5.5 can still make sense. If it solves complex tasks that GPT-5.4, Claude Sonnet 5, or DeepSeek V4 Pro cannot, the premium token rate is easy to justify. The problem is when teams route routine or fragile work to GPT-5.5 by default and never measure accepted output.
For the next few days, Codex users should track:
| Metric | Why it matters |
|---|---|
reasoning_output_tokens distribution | Detect fixed-value clustering in your own runs |
| Exact 516, 1034, 1552 counts | Compare with issue reports |
| Accepted patch rate | The core productivity metric |
| Test pass rate after agent edits | Better than subjective quality |
| Retry rate by model | Reveals hidden cost |
| Human review minutes per task | Captures non-token cost |
If you see a spike in exact-516 runs and worse results, route complex tasks away from GPT-5.5 until the issue is clearer. GPT-5.4 is the obvious first fallback inside the OpenAI stack because it is half the token price. Claude Sonnet 5 is also a credible coding-agent comparison point at $2 input and $10 output per million tokens. DeepSeek V4 Pro is far cheaper and worth benchmarking for lower-risk tasks, though quality and operational fit will vary.
Practical advice
First, export your own Codex usage metadata and compute simple counts by model: total records, exact 516 events, exact 1034 events, exact 1552 events, and exact 516 divided by counts at or above 516. You do not need to expose prompt content to do this aggregate check.
Second, run the same representative tasks through GPT-5.5, GPT-5.4, Claude Sonnet 5, and one low-cost route. Measure whether the output passed tests and whether a human accepted the patch. Token price alone will hide the problem.
Third, cap runaway agent loops but do not over-optimize for low token usage. In coding agents, fewer reasoning tokens are not always better. A cheap-looking run that produces a wrong edit can be the most expensive path.
Finally, keep GPT-5.5 for tasks where it earns its premium: repo-scale architecture changes, multi-file migrations, and ambiguous debugging. Routine code generation, formatting, documentation updates, and narrow fixes should not default to a $30/M output model unless accepted-task data proves it.
Finally, watch for an official response. As of publication, the GitHub issue was open and the public report had not established whether the clustering is a bug, a telemetry artifact, expected scheduler behavior, a routing interaction, or something else. Pricing teams should treat this as an active reliability signal, not as settled proof.
Bottom line
GPT-5.5 has not received a public price cut, and this issue does not change the official rate card. But it may change the buying calculus for Codex users.
At $5 input and $30 output per million tokens, GPT-5.5 needs to deliver a materially higher accepted-task rate than GPT-5.4, Claude Sonnet 5, and cheaper alternatives. If reasoning-token clustering is correlated with degraded performance in your own Codex workflows, GPT-5.5’s real cost is no longer $30/M output. It is $30/M output plus retries, fallback runs, and human cleanup.
Until OpenAI clarifies the behavior, the practical move is simple: measure your own Codex telemetry, compare exact-516 runs against task outcomes, and route based on cost per accepted result rather than model name.
Sources: OpenAI Codex issue #30364, related Codex issue #29353, and AI Pricing Guru’s live pricing dataset.