# GPT-5.5 Silently Throttles Reasoning for Third-Party Clients

> Third-party GPT-5.5 clients can cap reasoning at 516 tokens. You pay full price and may get a weaker answer.

- Author: AI Kai
- Published: 2026-06-24
- Updated: 2026-08-24
- Topic: AI Dev Tools
- Tags: gpt-55, openai, reasoning, api
- Canonical: https://hqman.me/blog/gpt55-reasoning-throttle/

If you call GPT-5.5 through a third-party client, OpenAI may cap your reasoning to exactly 516 tokens. You pay full price. You get a weaker model.

## What happened

A developer tested a probability problem across four access paths. Same model, same prompt, thinking mode on:

- **Codex CLI direct login:** Correct. Normal reasoning tokens.
- **Codex Python SDK:** Wrong. Reasoning = 516.
- **CLIProxyAPI reverse proxy:** Wrong. Reasoning = 516.
- **OpenCode direct login:** Wrong. Reasoning = 516.

Only the first-party Codex CLI got full reasoning. Others reproduced it.

It's also probabilistic. Same request sometimes returns normal reasoning, sometimes 516. No way to tell in advance.

## Why it matters

Normal reasoning on this problem uses thousands of tokens. 516 is not enough to finish the chain. The model gives up early and guesses wrong.

OpenAI's own Codex Python SDK triggers the same cap. You pay the same per-token rate either way.

## What to do

- Log `reasoning_tokens` on every response. Watch for 516.
- Don't run evals or benchmarks through third-party paths. Results may reflect the capped model.
- Use official Codex CLI or the Codex app for work where reasoning depth matters.

## Caveats

- Probabilistic, not 100% trigger rate
- Limited reproductions so far
- OpenAI has not responded

## Source

- [GitHub Issue #29353](https://github.com/openai/codex/issues/29353)