Skip to content

Which coding model is fastest right now?

Live response speed from Claude Code, Codex, Grok Build, etc. Privacy-safe open source clients. Free, no account needed.

Output tokens per second while the model is responding, excluding tool runs and your time.
Loading current observations…

Last 7 days

The current 7-day summary is unavailable until a published snapshot can be loaded.

Explore the 7-day board · Methodology and sample rules

Ask ChatGPT about the community data.

Connect Tokrate’s read-only MCP service to explore published comparisons. No desktop companion needed.

Use Tokrate in ChatGPT

One agent turn, two speeds

A coding turn alternates between the model responding and tools running. Response speed times only the responding.

  1. Your promptnot counted
  2. Responsecounted
  3. Tool runexcluded
  4. Responsecounted
  5. Tool runexcluded
  6. Final answercounted
Response speed
= output tokens ÷ time the model is responding. The first-token wait is included; tools and your time are excluded. Responses under 200 output tokens are not counted.
Turn speed
= output tokens ÷ the whole turn, from your prompt to the final answer. Tool runs and waiting are included.
Timeline: your prompt, a counted response, an excluded tool run, a counted response, an excluded tool run, then the counted final answer. Response speed uses the counted parts; turn speed uses the whole turn.
  • Response speedHeadline

    How fast the model answers. Typical coding responses: tens to a few hundred tok/s; bigger models are usually slower.

  • Turn speedSecondary

    How long your agent’s work takes. It reflects the workload: tool runs and waiting count too, so numbers are far lower than response speed.

  • First tokenCodex source

    The wait before the first output appears, in seconds. Reported by Codex, including supported T3 Code observations parsed from Codex. Claude Code, Grok Build, Antigravity, OpenCode and Kimi Code observations do not report it.

  • Efficiency indicatorIndicator

    Fewer output tokens per request scores higher. 100 = a typical request. Not a speed: it compares the output tokens a model spends per request, including subagent work. Needs Tokrate 0.1.16 or newer.

Your own pace, right on your desktop.

A quiet meter for your real coding sessions. See each model’s response speed, how it compares with the previous day, and your own local history, then compare the community’s observations. Available for Mac, with Windows and Linux alpha builds.

  • Claude CodeAnthropic
  • CodexOpenAI
  • Grok BuildxAI
  • AntigravityNewGoogle
  • OpenCodeNewAny provider
  • Kimi CodeNewMoonshot AI
  • T3 CodeNewT3

Product names are trademarks of their owners; Tokrate is independent and not affiliated with them.

Missing your coding tool? Suggest it:@mvilola on XCalltrin

Only numbers leave your computer.

Tokrate reads your local session logs to time each answer. The text stays on your machine; only the numbers are sent.

No ads or behavioral analytics. Sharing sends performance measurements under a random installation key.

Don’t take our word for it. Ask your AI agent to check the open-source code.

Each model, its own baseline.

Compare each model’s response speed, with every coding tool, provider and recorded effort setting kept separate. Turn speed and first-token waits sit alongside, and you can spot changes in your own sessions.

Real work. Honest numbers.

Different tasks take different time. Tools, context, effort and speed tier affect observations. Tokrate does not measure answer quality or streaming speed.

Questions before you install

Does Tokrate see my conversations?

Tokrate reads your local session logs on your computer to time each answer, but your conversations never leave it. No prompts, responses, code, file paths, session IDs, names or account details are ever uploaded. Sharing is optional: from version 0.1.11 it stays off until you choose “Yes, let's contribute”, and with “Only for local use” no measurements are sent at all (software update checks are a separate switch).

If you do share, each finished turn sends only performance numbers: coding tool, model, inference provider, recorded reasoning effort, client versions, token counts (including the output tokens of subagent work a turn started), durations, a five-minute time bucket and a random sample ID. The server stores your continent, never your country or IP address. Contribution data is not sold and is not used for advertising.

Does Tokrate use my tokens or cost me anything?

No. Tokrate never calls a model and never sends prompts on your behalf. It reads the session logs that Codex, Claude Code, Grok Build, Antigravity, OpenCode and Kimi Code already write on your computer; from Tokrate 0.1.25 it also recognizes supported T3 Code observations from Codex, Claude Code and Antigravity. It calculates speed from the token counts and timestamps in those records, so it adds nothing to your token usage, API bill or subscription limits.

It does not ask for API keys and needs no Tokrate account. The companions and the public board are free, with no ads.

Is Tokrate safe to install?

The Mac app is signed with an Apple Developer ID and notarized by Apple. Windows and Linux builds are early alpha releases: the Windows installer is not Authenticode-signed yet, so Windows may show an unknown-publisher warning. Every release publishes SHA-256 checksums, and in-app updates are signature-verified before you choose to install them.

The companions are open source under the MIT license and contain no advertising or behavioral analytics SDKs. You can read exactly what leaves your computer, or ask your AI agent to review the code for you. AI models have already reviewed it against our privacy promises, and the latest AI audit result is published unedited.

Is Tokrate free?

Yes. The public board at tokrate.dev, the companions for Mac, Windows and Linux, and the public API are free to use, with no account, no ads and no paid tier.

All questions