All five models passed all 17 hidden ticket tests in real app work.

• Tokens are the chunks of text a model reads and writes, about three quarters of a word each. Tokens read counts all the text sent to the model, each time it was sent. An agent sends its instructions and the conversation again with every request. That total grows with each step.
• The cache keeps text the model has already read, so the next request can reuse it faster and at a lower price. Read from cache is the share of what it read that came from there. Codex CLI's own count leaves out its hidden first request. Without it, Codex CLI's count gives 56% to 91% for these runs. Cache writes are text saved to the cache. Codex CLI always reports 0, so for Codex CLI I count what it read that did not come from the cache.
• Codex CLI's hidden first request. At the start of every run Codex CLI sends a request that loads the model's instructions and tools and writes nothing. Codex CLI leaves it out of its own token count, so I measured it once for each model and Codex CLI release and add it back.
• Output is everything the model wrote: its answer, its tool calls, such as asking to read a file, and its hidden thinking. Visible output is output minus thinking.
• Time on a run this short is mostly fixed: a second or two to start, then one to three seconds before each reply begins. A short answer is not much faster than a long one. Two runs of the same model can differ by several seconds, and a gap of a second is a tie.
• Account extras are what a subscription login loads by default. For Claude Code, they are the plugins, skills and connectors on the account's claude.ai. For Codex CLI, they are its connected apps and plugins. Round 13 turned them off. The launch-day runs under Earlier runs ran before that, with them on.
• Cost at API prices is what a run's tokens would cost at the model's published API prices: input that missed the cache at the input price, cache reads at the cache price, Claude's cache writes at the one-hour rate Claude Code uses, and output. For every Claude run it equals Claude Code's own estimate. My runs used subscriptions, which don't bill per token. Runs share a cache, so a run's cost depends partly on the runs before it. The price table lists every rate.
• Light runs are Claude Code runs that started with thousands fewer tokens of instructions and tools than the same model's other runs. They were a session's first run, before Claude Code loaded the account extras. They cost less and read fewer tokens. Some ended with a remark about the account's connectors. I removed that remark from the answers, but their token counts still include it. Round 13 reran this page's explain runs and real app work with the account extras off. A number where at least half the runs behind it are light carries a star. The Skills page's limits say how I find them.
• Effort is picked by Claude Code for Claude models and set by me for GPT models. The effort table lists every model.
• The token charts and tables show medians: the middle run, or the average of the two middle runs. The speed charts show the average of each model's timed runs. Every run is in the download below.

API prices used for cost#

The table scrolls sideways on a small screen.

Checked 29 September 2026. US dollars per million tokens, standard tier. OpenAI's rates are its short-context rates, which cover every request in these runs. OpenAI lists cache writes only for long context. Claude cache writes at the one-hour rate Claude Code uses, 2x input. Sources: OpenAI's pricing, Anthropic's pricing and Anthropic's caching docs.