AI model speed, measured in real agent sessions
Every published speed number is one prompt, timed after the first token. This page reports what Claude and Codex models actually deliver inside working sessions: output tokens per wall-clock minute, from 118,896 measured requests across our own Claude Code and Codex fleet, by model, by week and by hour.
Right now
Snapshot of 2 September 2026: medians over 118,896 requests. Live figures load when sessions are running.
Observed output speed by model, main sessions
Output tokens per wall-clock minute per request. Bar spans the 10th to 90th percentile, the dot is the median. 7 July 2026 to 2 September 2026.
Models with fewer than 500 measured requests in a scope are in the download, not the chart. Source: Jordan James Media session transcripts, provider-billed tokens.
Claude Fable 5.1
#claude-fable-5-1The middle 80 percent of 1,003 requests ran between 3,050 and 5,282 tokens a minute; the aggregate rate (total tokens over total minutes) was 4,697. First content block after a median 10.862 seconds. 39 percent of its output was thinking.
Claude Opus 5
#claude-opus-5The middle 80 percent of 64,675 requests ran between 1,952 and 4,754 tokens a minute; the aggregate rate (total tokens over total minutes) was 3,717. First content block after a median 8.085 seconds. 18 percent of its output was thinking.
Claude Fable 5
#claude-fable-5The middle 80 percent of 20,282 requests ran between 1,540 and 4,760 tokens a minute; the aggregate rate (total tokens over total minutes) was 3,556. First content block after a median 8.757 seconds. 21 percent of its output was thinking.
Claude Opus 4.8
#claude-opus-4-8The middle 80 percent of 4,024 requests ran between 1,588 and 4,298 tokens a minute; the aggregate rate (total tokens over total minutes) was 3,532. First content block after a median 11.347 seconds. 4 percent of its output was thinking.
Codex (GPT-5.6)
#codex-gpt-5-6The middle 80 percent of 18,368 requests ran between 943 and 3,639 tokens a minute; the aggregate rate (total tokens over total minutes) was 2,581. First content block after a median 3.516 seconds. 37 percent of its output was thinking.
Codex auto-review
#codex-auto-reviewThe middle 80 percent of 4,465 requests ran between 865 and 2,185 tokens a minute; the aggregate rate (total tokens over total minutes) was 1,493. First content block after a median 2.531 seconds. 41 percent of its output was thinking.
Subagent runs
The same models spawned as short-lived workers with a small context. Fewer than 300 requests are left out.
| Model | Requests | Median tokens/min | Tokens/s | First output |
|---|---|---|---|---|
| Claude Sonnet 5 | 2,254 | 4,114 | 69 | 5.83 s |
| Claude Opus 5 | 2,368 | 3,724 | 62 | 4.566 s |
| Claude Haiku 4.5 | 339 | 3,621 | 60 | 9.26 s |
| Claude Fable 5 | 835 | 3,017 | 50 | 4.958 s |
Weekly median speed, main sessions
One panel per model so the lines never cross. Shared scale, 2,000 to 4,200 tokens per minute. Weeks with fewer than 50 requests are omitted.
Median speed by hour of day (UTC), main sessions
Darker is faster within each model's own range of well-sampled hours. Outlined cells hold fewer than 500 requests and are left out of the shading range. Hover any hour for the value.
Hours are UTC. Sydney is UTC plus 10; New York is UTC minus 4 in September.
How one request is timed
A real Claude Fable 5.1 reply from the session that built this page: 1,988 output tokens in 26.7 seconds, 4,467 tokens a minute. The tool then ran for 63 seconds, which is not counted.
Methodology 1.0
What the number is, how it is produced, and what it is not.
The metric
Output tokens divided by wall-clock minutes for one model request. The request starts at the transcript record written immediately before it (the user message, tool result or attachment it went out with) and ends when its last streamed content block was written. Thinking and time to first token are inside the window. Tool execution is outside it, because tool results are written after the tool finishes.
Sources
Claude Code session transcripts under ~/.claude/projects/ and Codex rollouts under ~/.codex/sessions/ on our own workstation, every record timestamped, every reply carrying the provider's usage counters. 3,595 files at the last snapshot. Main interactive sessions and subagent runs are reported separately. One measuring point in Sydney, so every first-output figure includes that network path.
Exclusions, counted
Replies under 20 output tokens (a short tool call over a few seconds says nothing about generation), placeholder and fallback turns, replies with no counters, and requests longer than an hour. Every exclusion is counted by reason in the JSON download, so a zero is a measured zero.
Units and what it is not
Tokens are provider-billed tokens, the ones invoices and rate limits count; on the same Claude text they run about 1.6 times the OpenAI o200k count that benchmark sites use. This is observed throughput, never a rate-limit quota, so it is written out in full and never abbreviated. Medians and 90th percentiles are reported, never means. Models with fewer than 500 requests in a scope stay in the download only.
Data and API
Rows carry model names, ISO weeks, counts and durations. No project, prompt, session identifier or dollar figure is in the data.
Free to download and reuse with a link back to this page.
Questions people ask
How fast is Claude Opus 5 in a real coding session?
Across 64,675 of our main-session requests between 7 July and 2 September 2026, Claude Opus 5 delivered a median 3,628 output tokens per wall-clock minute (60 tokens a second) with a 90th percentile of 4,754, and its first content block arrived a median 8.1 seconds after the request was sent. Those are provider-billed tokens measured end to end, including thinking, not a benchmark decode rate.
Is tokens per minute the same as the TPM rate limit?
No. Anthropic, OpenAI, Azure and Google all quote rate-limit quotas in tokens per minute, which is how many tokens you are allowed to send or receive. The figures on this page are observed throughput, how many output tokens a model actually delivered per minute inside real sessions. We write the metric out in full and never abbreviate it so the two are not confused.
Why is my agent session slower than the benchmark number?
Benchmarks time one prompt after the first token arrives and count tokens in OpenAI's o200k tokenizer. A session request also waits for the first token and for thinking, and the billed tokens you see run about 1.6 times the o200k count for the same Claude output. Put Opus 5's 3,628 billed tokens a minute in benchmark units and it is about 37 tokens a second end to end against a published 52 decode-only.
Is fast mode worth it for AI agents?
Fast mode raises decode speed by up to 2.5 times on both Anthropic and OpenAI, at 2 to 2.5 times the price or credit burn, and it does not shorten time to first token or tool execution. A request that waits eight seconds and streams for twenty goes from 28 seconds to about 16, a 1.75 times gain, and the tool runs between requests do not change. It pays for long generation with few tool calls and not for sessions dominated by waiting.
How do I measure model speed in my own Claude Code or Codex sessions?
Every Claude Code transcript under ~/.claude/projects/ and every Codex rollout under ~/.codex/sessions/ carries a timestamp on each record and the provider's usage counters on each reply. Anchor each request to the record written just before it, end it at its last streamed block, divide output tokens by minutes, drop replies under 20 tokens, and report the median and 90th percentile per model. Our measurement module and dataset are linked on this page.
Which AI coding model is fastest right now?
In our main sessions Claude Fable 5.1 has the highest median at 4,311 tokens a minute, but 39 percent of that is thinking, so Claude Opus 5 writes more visible tokens per minute. Claude Sonnet 5 is the fastest subagent at 4,114. Codex on GPT-5.6 has the lowest rate at 2,490 but the shortest wait for a first token at 3.5 seconds. Speed drifted by 11 to 20 percent within weeks, so check the live figure on this page before deciding.
Why benchmarks time a sprint while your agent runs a marathon, how the same model runs at three speeds, and how to choose a model on speed that is real.
Read itCustom platforms, integrations and agent workflows, built and run by an AI-native studio in Sydney. This page is what that instrumentation looks like from the outside.
Talk to us about AI development