00 · MODEL RANKINGS

The models agents on Rensei use, week by week.

Token usage across every agent session on the platform, by model, lab, serving provider, harness and kind of work. Rebuilt every day from production rollups. The full data set is free to download below.

Tokens, last 7 days: 22.8B
Change on the prior 7 days: +89% up on the prior 7 days
Agent sessions, last 7 days: 3,024
Models in use, last 7 days: 9

WEEK OCT 28 - NOV 3 · DATA THROUGH NOV 3, 2026 UTC · UPDATED DAILY

REC 00 ▸ sha256:0a62…dbd2 ▸ prev b7d7…9d72 ▸ build 2026-11-04T22:23Z

01 · MODELS

Usage by model

Each model's share of tokens over time, and the full leaderboard. The 7-day change column compares the latest 7 days with the 7 days before, which is where a new checkpoint taking over from an older one shows first.

FIG 1.1

Model ranked by tokens, last 30 days, with the change over the latest 7 days
#	Model	Lab	Tokens	Share	7d7-day change
1	Muse Spark 1.3Meta	Meta	77.9B	79.9%	+9% up on the prior 7 days
2	GPT-6.1 SolOpenAI	OpenAI	14B	14.4%	+1,423% up on the prior 7 days
3	GPT-6 AstraOpenAI	OpenAI	4.3B	4.4%	+1,226% up on the prior 7 days
4	GLM-5.3Z.ai	Z.ai	790M	0.8%	-79% down on the prior 7 days
5	GPT-5.4OpenAI	OpenAI	157M	0.2%	-
6	Claude Opus 5.5Anthropic	Anthropic	134M	0.1%	new
7	Claude Fable 5.1Anthropic	Anthropic	80.9M	<0.1%	-31% down on the prior 7 days
8	Muse Wonder Full 1.0Meta	Meta	43.4M	<0.1%	new
9	Claude Sonnet 4.6Anthropic	Anthropic	661K	<0.1%	+243% up on the prior 7 days
10	Claude Sonnet 5.5Anthropic	Anthropic	596K	<0.1%	+19% up on the prior 7 days
11	Gemini 3.8 FlashGoogle	Google	2.8K	<0.1%	-

Weekly data table

Weekly tokens by model
Week	Muse Spark 1.3	GPT-6.1 Sol	GPT-6 Astra	GLM-5.3	GPT-5.4	Claude Opus 5.5	Claude Fable 5.1	Muse Wonder Full 1.0	Other models
Oct 28 - Nov 3	11.6B	7.1B	3.8B	120M	0	130M	24.9M	43.4M	712K
Oct 21 - Oct 27	10.7B	466M	283M	563M	0	817K	36M	0	316K
Oct 14 - Oct 20	26B	6.5B	253M	107M	0	972K	8.4M	0	232K
Oct 7 - Oct 13	20.1B	28M	0	0	0	1.3M	0	0	0
Oct 5 - Oct 6	9.5B	0	0	0	157M	309K	11.6M	0	0

FIG 1.1 - Tokens by model. Share is each model's part of the trailing 7 days; volume is daily tokens, each model drawn from zero. The table ranks every model with tokens in the window.

REC 01 ▸ sha256:0547…d97b ▸ prev 0a62…dbd2 ▸ build 2026-11-04T22:23Z

02 · MARKET SHARE

Market share by lab

Tokens rolled up to the lab that built each model, whichever route served the call. A model reached through a gateway and the same model reached directly both count to its lab.

FIG 2.1

Lab ranked by tokens, last 30 days, with the change over the latest 7 days
#	Lab	Tokens	Share	7d7-day change
1	Meta	78B	80.0%	+9% up on the prior 7 days
2	OpenAI	18.5B	19.0%	+1,349% up on the prior 7 days
3	Z.ai	790M	0.8%	-79% down on the prior 7 days
4	Anthropic	216M	0.2%	+319% up on the prior 7 days
5	Google	2.8K	<0.1%	-

Weekly data table

Weekly tokens by lab
Week	Meta	OpenAI	Z.ai	Anthropic	Google
Oct 28 - Nov 3	11.6B	10.9B	120M	156M	0
Oct 21 - Oct 27	10.7B	750M	563M	37.2M	0
Oct 14 - Oct 20	26B	6.7B	107M	9.6M	2.8K
Oct 7 - Oct 13	20.1B	28M	0	1.3M	0
Oct 5 - Oct 6	9.5B	157M	0	11.9M	0

FIG 2.1 - Tokens by model lab, with each lab's share and 7-day change.

REC 02 ▸ sha256:8cd7…f762 ▸ prev 0547…d97b ▸ build 2026-11-04T22:23Z

03 · SERVING PROVIDERS

How agents on Rensei reach models

Agents on Rensei reach a model in three ways: through an AI gateway such as Vercel AI Gateway or OpenRouter, directly from the lab on an API key, or on the customer's own subscription through the lab's CLI. This chart counts tokens by whoever served them.

FIG 3.1

Provider ranked by tokens, last 30 days, with the change over the latest 7 days
#	Provider	Kind	Tokens	Share	7d7-day change
1	Vercel AI GatewayGateway	Gateway	75.9B	77.9%	-7% down on the prior 7 days
2	OpenAILab, direct	Lab, direct	18.3B	18.8%	+1,349% up on the prior 7 days
3	MetaLab, direct	Lab, direct	2.2B	2.2%	new
4	Z.aiLab, direct	Lab, direct	790M	0.8%	-79% down on the prior 7 days
5	AnthropicLab, direct	Lab, direct	215M	0.2%	+319% up on the prior 7 days
6	GoogleLab, direct	Lab, direct	2.8K	<0.1%	-

Weekly data table

Weekly tokens by provider
Week	Vercel AI Gateway	OpenAI	Meta	Z.ai	Anthropic	Google
Oct 28 - Nov 3	9.9B	10.9B	1.8B	120M	155M	0
Oct 21 - Oct 27	10.7B	750M	202K	563M	37M	0
Oct 14 - Oct 20	26B	6.7B	0	107M	9.6M	2.8K
Oct 7 - Oct 13	20B	28M	125M	0	1.3M	0
Oct 5 - Oct 6	9.4B	0	307M	0	11.9M	0

FIG 3.1 - Tokens by serving provider: gateways, and labs serving their own models directly.

How each lab is reached

For each lab, the split of its tokens by route over the last 90 days.

Direct, subscription via vendor CLI
Direct, API key
Vercel AI Gateway
Other route

Meta 81.3B tokens
Direct, subscription via vendor CLI 1.5%,Direct, API key 1.2%,Vercel AI Gateway 97.3%
OpenAI 18.5B tokens
Direct, subscription via vendor CLI 99.1%,Vercel AI Gateway 0.9%
Z.ai 790M tokens
Direct, API key 100%
Anthropic 218M tokens
Direct, subscription via vendor CLI 37.3%,Direct, API key 62.3%,Vercel AI Gateway 0.3%
Google 2.8K tokens
Other route 100%

REC 03 ▸ sha256:35b8…ca09 ▸ prev 8cd7…f762 ▸ build 2026-11-04T22:23Z

04 · HARNESSES

Usage by harness

The agent harness that ran each session, such as Claude Code, Codex, Pi or Muse. Rensei runs several harnesses side by side; the models each one runs are listed below the chart.

FIG 4.1

Harness ranked by tokens, last 30 days, with the change over the latest 7 days
#	Harness	Tokens	Share	7d7-day change
1	Pi	77.7B	79.7%	-6% down on the prior 7 days
2	Codex	18.3B	18.8%	+1,349% up on the prior 7 days
3	Muse	1.2B	1.2%	new
4	Claude Code	215M	0.2%	+319% up on the prior 7 days
5	Other	2.8K	<0.1%	-

Weekly data table

Weekly tokens by harness
Week	Pi	Codex	Muse	Claude Code	Other harnesses
Oct 28 - Nov 3	10.6B	10.9B	1.2B	155M	0
Oct 21 - Oct 27	11.2B	750M	0	37M	0
Oct 14 - Oct 20	26.1B	6.7B	0	9.6M	2.8K
Oct 7 - Oct 13	20.1B	28M	0	1.3M	0
Oct 5 - Oct 6	9.7B	0	0	11.9M	0

FIG 4.1 - Tokens by agent harness, with each harness's share and 7-day change.

Models per harness

The three models carrying the most tokens in each harness over the last 90 days.

Top models per harness by share of the harness's tokens, last 90 days
Harness	Tokens	Top models, share of the harness
Pi	81B	98.8%Muse Spark 1.3 1.0%GLM-5.3 0.2%GPT-5.4
Codex	18.3B	76.6%GPT-6.1 Sol 23.4%GPT-6 Astra <0.1%GPT-5.6 Sol
Muse	1.2B	96.4%Muse Spark 1.3 3.6%Muse Wonder Full 1.0
Claude Code	217M	61.6%Claude Opus 5.5 37.5%Claude Fable 5.1 0.3%Claude Sonnet 5

REC 04 ▸ sha256:4b3b…20f1 ▸ prev 35b8…ca09 ▸ build 2026-11-04T22:23Z

05 · BY TASK

Top harnesses and models by task

Every session carries a work type, such as development, QA or research. This is how tokens divide across those kinds of work, and which harnesses and models lead each one.

FIG 5.1

Work type ranked by tokens, last 30 days, with the change over the latest 7 days
#	Work type	Tokens	Share	7d7-day change
1	Development	62.7B	64.4%	-3% down on the prior 7 days
2	QA	33.5B	34.4%	+288% up on the prior 7 days
3	Improvement loop	961M	1.0%	-71% down on the prior 7 days
4	Interactive	158M	0.2%	-100% down on the prior 7 days
5	Research	82M	<0.1%	+4,567% up on the prior 7 days
6	Backlog grooming	596K	<0.1%	+19% up on the prior 7 days
7	Other	2.8K	<0.1%	-

Weekly data table

Weekly tokens by work type
Week	Development	QA	Improvement loop	Interactive	Research	Backlog grooming	Other work
Oct 28 - Nov 3	7B	15.5B	182M	512K	46.8M	200K	0
Oct 21 - Oct 27	7.3B	4B	628M	113M	1M	167K	0
Oct 14 - Oct 20	23.7B	8.9B	139M	16.6M	972K	229K	2.8K
Oct 7 - Oct 13	16.6B	3.5B	0	28M	32.9M	0	0
Oct 5 - Oct 6	8.1B	1.6B	11.6M	0	309K	0	0

FIG 5.1 - Tokens by work type.

Leaders by task

For each work type, the three harnesses and the three models with the largest share of its tokens over the last 90 days.

Leading harnesses and models for each kind of work, by share of that work's tokens, last 90 days
Work type	Tokens	Top harnesses	Top models
Development	65.4B	88.4%Pi 10.0%Codex 1.6%Muse	89.7%Muse Spark 1.3 9.9%GPT-6.1 Sol 0.2%GPT-5.4
QA	34.2B	65.4%Pi 33.7%Codex 0.5%Muse	65.8%Muse Spark 1.3 21.9%GPT-6.1 Sol 11.8%GPT-6 Astra
Improvement loop	971M	82.3%Pi 9.3%Codex 8.4%Claude Code	81.4%GLM-5.3 9.3%GPT-6 Astra 8.3%Claude Fable 5.1
Interactive	158M	99.2%Codex 0.5%Pi 0.3%Claude Code	81.4%GPT-6 Astra 17.8%GPT-6.1 Sol 0.4%Claude Sonnet 4.6
Research	86.7M	53.4%Codex 42.0%Pi 4.6%Claude Code	41.4%Muse Spark 1.3 39.3%GPT-6.1 Sol 14.1%GPT-6 Astra
Backlog grooming	596K	100%Claude Code	100%Claude Sonnet 5.5
Backlog creation	1.9K	100%Claude Code	100%Claude Opus 5

REC 05 ▸ sha256:fbd3…3d04 ▸ prev 4b3b…20f1 ▸ build 2026-11-04T22:23Z

06 · TOKEN MIX

Token mix

Input, cache reads, cache writes and output as parts of all tokens. Agent sessions send long context on every turn, so the cache series show how much of that context providers served from cache.

In the 7 days to Nov 3, cache reads were 97.6% of all tokens.

FIG 6.1

Kind ranked by tokens, last 30 days, with the change over the latest 7 days
#	Kind	Tokens	Share	7d7-day change
1	Cache read	93.8B	96.2%	+89% up on the prior 7 days
2	Input (uncached)	3.2B	3.2%	+87% up on the prior 7 days
3	Output	481M	0.5%	+91% up on the prior 7 days
4	Cache write	16.9M	<0.1%	+248% up on the prior 7 days

Weekly data table

Weekly tokens by token kind
Week	Input (uncached)	Cache read	Cache write	Output
Oct 28 - Nov 3	438M	22.2B	12.1M	107M
Oct 21 - Oct 27	234M	11.7B	3.5M	56.2M
Oct 14 - Oct 20	618M	32B	929K	162M
Oct 7 - Oct 13	435M	19.6B	201K	111M
Oct 5 - Oct 6	1.4B	8.2B	247K	45.2M

FIG 6.1 - Tokens by kind. Reasoning tokens sit inside output.

REC 06 ▸ sha256:5164…a0cd ▸ prev fbd3…3d04 ▸ build 2026-11-04T22:23Z

07 · METHODOLOGY

What is counted, and the data

Tokens. Input, cache reads, cache writes and output, added together. Reasoning tokens are part of output and are not counted twice. OpenRouter and Vercel AI Gateway count tokens the same way for their rankings.
Scope. Every agent session on the Rensei platform, as platform totals. No organization, project or person is identified, and no price or spend is shown.
Labs. A model's lab comes from its canonical model identity, so the lab rollup does not depend on the provider that served the call. Models the platform cannot map to an identity are listed as Unattributed.
Serving providers. The provider that served the tokens: an AI gateway, or the lab itself. Direct access is split by how the call was authorized: an API key, or the customer's own subscription through the vendor's CLI.
Windows and change. The data covers the last 90 complete days in UTC; the current day is left out until it ends. Weeks end on the latest complete day. On the share view, each day shows a series' share of the 7 days ending that day. The 7-day change compares the latest 7 days with the 7 days before. New means the earlier week had less than 1% of the latest week's tokens, too little for a percentage to mean anything.
Series. Each chart draws up to eight series per window, chosen by their highest weekly share in that window, so a model that led any week keeps its own line even when a busier month follows. The rest are added together as Other. Lines between days are smoothed with a curve that never rises above or dips below the days it joins.
Sessions. Distinct agent sessions per day, added up across the week, so a session that runs past midnight counts on both days.
Updates. The page is rebuilt every day from the platform's production rollups.

REC 07 ▸ sha256:813f…f22a ▸ prev 5164…a0cd ▸ build 2026-11-04T22:23Z