Agent Usage

Rankings

Which models and coding tools builders pick, and which they actually run. Votes come from people. Usage comes from reported tokens. Neither one is a benchmark.

What builders actually run

Share of reported tokens by model, last 8 weeks · Locally reported · UTC
4.1B reported tokens in the latest week

Best value

The most capability per dollar, open or closed.
Rank byAdd a model
No votes in Best value this month. Until there are, this list is ordered by what builders actually run.Cast the first vote
#ModelVotesUsage sharePrice per 1M (in / out)Builders
1
Claude Opus 5.5AnthropicProprietary
70.3%
—5
2
Claude Opus 4.8AnthropicProprietary
13.8%
—1
3
Claude Sonnet 5.5AnthropicProprietary
9.3%
—3
4
Claude Fable 5.1AnthropicProprietary
2.7%
—2
5
GPT-6.1 SolOpenAIProprietary
1.3%
—1
6
Claude Opus 5AnthropicProprietary
1.1%
—1
7
GPT-6 AstraOpenAIProprietary
0.7%
—1
8
Claude Haiku 5.5AnthropicProprietary
0.1%
—1
9
Claude Sonnet 5AnthropicProprietary
No reported usage yet——
10
Composer 2.5CursorProprietary
No reported usage yet——
11
DeepSeek V4 ProDeepSeekOpen weights
No reported usage yet——
12
Gemini 4 ArgonGoogleProprietary
No reported usage yet——
13
GLM-5.2Z.aiOpen weights
No reported usage yet——
14
GPT-5.6 SolOpenAIProprietary
No reported usage yet——
15
Grok 4.6xAIProprietary
No reported usage yet——
16
Kimi K2.6Moonshot AIOpen weights
No reported usage yet——
17
Kimi K2.7 CodeMoonshot AIOpen weights
No reported usage yet——
18
MiniMax M3MiniMaxOpen weights
No reported usage yet——
19
Qwen3.5AlibabaOpen weights
No reported usage yet——
Votes for October 2026. Usage share is each model's part of all reported tokens in the same period, from the open-source companion. Prices, where shown, are list prices added with the entry.Vote in every category

One vote per category

Each account picks one model for Best paid, one for Best open source, one for Best value and one coding tool, every month. Change your pick any time before the month ends.

Tokens never add weight

A vote from someone who runs a model all day counts the same as anyone else's. Voters who also report using it show as "uses it" on its page.

Two different signals

Votes show what people recommend. Usage share shows what the companion reports people running. When they disagree, that's worth reading about.