Insights
Harness Restriction and the model TPS ledger
Slowtest is a clock and a compliance check. It is also a simple question: will this model run inside an instrumented harness at all? That answer is not IQ.
Harness Restriction
Some models will paste the Slowtest script, call the start beacon, write the audit, and post finish. Others will not. They refuse the outbound hook — the telemetry / beacon connection that lets this page time Generation without you copying the essay back.
That refusal is useful. It tells you the model (or the chat product around it) will not cooperate with an automated, instrumented run. We call that Harness Restriction. It is about willingness to run in a wired-up harness. It is not a quality score, not an IQ number, and not a claim that a blocked model is worse at writing.
How we observe it, in plain English: on 1-paste, Slowtest waits for a start beacon, then a finish beacon, for the same serial. If the model sets that up, the handshake completed. If it will not POST — common on locked-down consumer chat — the listener stays empty. We do not invent a percentage from that. We use three labels, plus Unrated when we have not seen enough Slowtest runs.
Source: observed on Slowtest runs. Dates are when we last wrote the note — not a published lab paper. Leave room to update.
| Model | Harness | Observed | Note |
|---|---|---|---|
| Grok 4.6 | Open | 2026-09-10 UTC | Typically completes the 1-paste telemetry handshake on Slowtest. |
| Claude (Opus 5 / Fable 5 and typical Claude chat) | Blocked | 2026-09-10 UTC | Often refuses to set up automated telemetry. Use own-key or the keyboard worksheet if you still want a clock. |
| GPT-5.6 / ChatGPT | Unrated | — | Consumer chat often blocks outbound scripts; coding-agent or own-key paths vary. No Slowtest band yet. |
| Gemini 3.7 Flash | Unrated | — | Own-key stream works in-browser. 1-paste depends on whether that Gemini surface will POST. No band claimed. |
Model TPS ledger
Generation tok/s is the decode-ish clock when we have a real generation window (own-key first-to-last token, or 1-paste start/finish beacons). Full-turn tok/s is the whole job: think, tools, queue, copy slack, and the POST home. Empty cells are empty on purpose.
This table is a curated ledger Tim/CoS can update in data/model-tps.json and on this page. It is not a crowdsourced average. Rows marked placeholder have no Slowtest number yet. Do not paste a vendor chart into the tok/s columns.
Ledger updated 10 Sep 2026 UTC. All tok/s cells are placeholders until a real Slowtest receipt is entered.
| Model | Gen tok/s | Full-turn tok/s | n runs | Last updated | Notes |
|---|---|---|---|---|---|
| Grok 4.6 | — | — | 0 | — | Typically completes 1-paste. Replace with a live Slowtest run. placeholder |
| Claude Opus 5 | — | — | 0 | — | Often refuses 1-paste. Prefer own-key if you need a number. placeholder |
| Claude Fable 5 | — | — | 0 | — | Same harness pattern as other Claude chat surfaces. placeholder |
| Gemini 3.7 Flash | — | — | 0 | — | Own-key stream is the clean Generation clock. placeholder |
| GPT-5.6 | — | — | 0 | — | Leave null until a Slowtest receipt is pasted in. placeholder |
| DeepSeek V4 | — | — | 0 | — | Hosted vs local will differ. Do not copy a vendor chart here. placeholder |
How to run, and how to update
1-paste: copy the serial script on the home page, paste it into a model that will POST start, write the ~1000-word audit, then POST finish. Own-key: stream from your browser; the key stays in localStorage. A finished run can mint a share page at /r/SLOW-… with metrics only.
To update the ledger: edit data/model-tps.json, put the measured tok/s in, set placeholder to false, bump last_updated_utc, and change the matching row on this page. Same for harness notes in data/harness-restriction.json. No public index of every /r/ id — those pages live on share links.
Home · Claude Opus 5 vs Grok 4.6 · Fable 5 vs Gemini 3.7 Flash · GPT-5.6 vs DeepSeek V4