Skip to content
Ruvicode logo
Ruvicode
USD
ReportNovember 3, 2026·6 min read

What AI Agents Ran On in October 2026: Ruvicode Usage Report

Real usage data from 185,421 chat requests on Ruvicode in October 2026: the most used models, how much of each prompt came from cache, and how big agent prompts get.

In October 2026, Ruvicode served 185,421 chat requests across 56 models. The numbers have the shape of coding agents and other tool-using apps: huge prompts, short answers, and a prompt that is mostly cache.

The headline numbers:

  • 69.7% of chat requests went to DeepSeek models.
  • 83.1% of all prompt tokens were served from cache.
  • The median request carried 37,633 prompt tokens. One in ten carried more than 267,809.
  • Requests used about 85 prompt tokens for every output token.
  • 42.3% of all output tokens were reasoning tokens.
  • Compared with September 2026, requests grew 2.1x and tokens 2.3x.

How we counted

This report uses Ruvicode's usage records for completed chat requests between 1 and 31 October 2026 (UTC). Requests from Ruvicode's own internal and test accounts are left out; requests made with the free community key are included. Token counts are the ones each response reported in its usage object. Image requests (209 in October) are not part of the chat figures.

Ruvicode does not log or store prompt or response content, so this report is built from token counts and model names only. It describes Ruvicode's customers, mostly developers running agents, not the AI market as a whole.

The most used models

RankModelShare of chat requestsPrompt tokens from cacheMedian prompt (tokens)
1DeepSeek V4 Flash26.8%82.7%7,342
2DeepSeek V4.1 Flash26.4%84.0%52,840
3DeepSeek V4.1 Flash Uncensored10.8%95.9%122,746
4GPT-5.6 Luna7.2%77.6%43,820
5GPT-6 Luna6.9%47.9%6,865
6DeepSeek V4 Pro5.0%87.6%95,172
7GLM 5.3 Flash4.2%74.1%98,574
8GLM 5.33.8%79.5%60,960
9Claude Haiku 5.51.2%82.7%187,848
10Kimi K31.0%79.5%64,213

DeepSeek models took the top three places, and with DeepSeek V4 Pro and the turbo variant they handled 69.7% of all chat requests. The common thread is a large context window (around a million tokens) at a low per-token rate, which is what a long-running agent needs.

The median prompt size says a lot about how each model is used. DeepSeek V4 Flash and GPT-6 Luna sit around 7,000 tokens, which points to shorter tasks and chat. DeepSeek V4.1 Flash Uncensored, DeepSeek V4 Pro, GLM 5.3 Flash and Claude Haiku 5.5 sit between 95,000 and 190,000 tokens, which is what long agent sessions with big codebases look like.

Current rates for every model are on the models catalog. We do not print them here because they move with the market every 2 minutes.

Prompts are huge, answers are short

An agent sends its whole conversation on every tool call: the system prompt, the tool definitions, every file it has read, and every earlier turn. That is why the median request carried 37,633 prompt tokens and the top 10% carried more than 267,809.

The answers are small by comparison. Across October, requests used about 85 prompt tokens for every output token. For anyone running agents, the input side of the bill is the one that matters.

Most of the prompt is cache

83.1% of all prompt tokens in October were cache reads. When a request repeats a prompt prefix the provider has already processed, that part is billed at the cached input rate, which on most models is several times lower than the full input rate. On Ruvicode this happens automatically.

The share varies by model and by how it is used. DeepSeek V4.1 Flash Uncensored reached 95.9%, which fits long sessions that keep extending one conversation. GPT-6 Luna, with short prompts, reached 47.9%. Prompt caching for AI agents explains how the billing works.

Reasoning takes a large share of the output

42.3% of all output tokens in October were reasoning tokens: text the model wrote while thinking, before its answer. The share depends heavily on the model:

ModelReasoning share of output tokens
DeepSeek V4 Flash62.6%
DeepSeek V4.1 Flash Uncensored57.5%
GLM 5.330.0%
GLM 5.3 Flash24.2%
DeepSeek V4.1 Flash22.5%
DeepSeek V4 Pro20.6%
GPT-6 Luna10.2%
GPT-5.6 Luna7.0%

Reasoning tokens bill as output tokens, and they count against max_tokens. A budget that is fine for a model that thinks little can be used up entirely on a model that thinks a lot, which ends in an empty answer. Reasoning model returns empty content shows what that looks like and how to avoid it.

Growth from September

Chat requests grew from 90,224 in September to 185,421 in October, 2.1x. Tokens grew from about 7.5 billion to about 17.3 billion, 2.3x. Tokens grew faster than requests, so the average request got bigger, which fits more and longer agent sessions.

What this means if you run agents

  • Look at the cached input rate, not only the input rate. With 83% of prompt tokens coming from cache, the cached rate decides most of an agent's input bill. Each model page lists both.
  • Pick the context window for your longest session. One in ten requests carried more than 267,809 prompt tokens. A model with a small window will run out on long tasks.
  • Budget max_tokens for reasoning. On models that reason a lot, more than half of the output can be thinking.
  • Cap the spend per key. Agents run unattended. A daily or monthly spend limit on the API key turns a runaway loop into a small, bounded cost.

The realtime page shows the same traffic live, request by request.

Frequently asked questions

Which AI models were used most on Ruvicode in October 2026?
DeepSeek V4 Flash (26.8% of chat requests), DeepSeek V4.1 Flash (26.4%), DeepSeek V4.1 Flash Uncensored (10.8%), GPT-5.6 Luna (7.2%) and GPT-6 Luna (6.9%). DeepSeek models together handled 69.7% of all chat requests.
How much of an AI agent's prompt is served from cache?
On Ruvicode in October 2026, 83.1% of all prompt tokens were cache reads. Agents resend their whole conversation on every tool call, so most of each prompt repeats what the model has already processed.
How large are AI agent prompts?
The median chat request on Ruvicode in October 2026 carried 37,633 prompt tokens, and one in ten carried more than 267,809. Requests used about 85 prompt tokens for every output token.
How many output tokens do reasoning models spend on thinking?
Across all chat requests on Ruvicode in October 2026, 42.3% of output tokens were reasoning tokens. The share depends on the model: 62.6% on DeepSeek V4 Flash, 7.0% on GPT-5.6 Luna.
Where does this data come from?
From Ruvicode's own usage records for completed chat requests between 1 and 31 October 2026 (UTC), excluding Ruvicode's internal and test accounts. Token counts are the ones each response reported. Ruvicode does not store prompt or response content, so no content was read.

Ready to start building?

Get your API key in under 2 minutes.

Get Started →
R

Ruvicode Team

The people who build and run Ruvicode