Tokens rose 37% in October

Points forts

• Token volume kept rising: up 37% from September and almost 25x from January.
• Effective price per token fell: the blended average price per token kept marching down, and last month it was 57 to 70% below January averages.
• Decision models now see 21% of all requests: though the category is only weeks old, decision models now get more requests than any modality other than text models.

790 trillion tokens in October

October was the largest month in OpenRouter's history: 790T tokens, up 37% from September and a startling 25x from January.

The increase was broad based, but the refresh of DeepSeek V4.1 Flash, together with the introduction of Haiku 5.5 and the rest of the Claude 5.5 family, were major contributors.

Agents consume 76% of all tokens

Agent dominance has only increased over the course of 2026. We classify every API key as agentic, human or mixed with a weighted classifier of seven signals, including tool-call rate, number of turns and request timing.

A year ago agentic keys took 8% of tokens. In October they took 76%, about 600T tokens.

A distinguishing feature of agentic usage is resending the same, often extensive, context on every turn, so prompt caching matters more for agents than for human usage. Most providers served cached prefixes for 85 to 97% of agentic prompt tokens in October. The median cache hit rate has risen steeply, from 76.1% in June to over 90% in October.

Open weights bounced back

The open-weight share of paid tokens fell to 52% in the week of September 28, when Claude Opus 5.5, Sonnet 5.5 and GPT-6.1 Sol all launched. Two weeks later it was back above 70%, primarily due to DeepSeek V4.1 Flash.

Open weights take about a quarter of paid spend, up from 11% in January, and that share has barely moved over the last quarter or so.

Effective price per token fell 57 to 70% since January

The blended price per million tokens fell from $1.21 to $0.52 for closed-weight models and from $0.28 to $0.084 for open weights. The gap between them widened from 4.3× to 6.2×.

This effective price can move for many reasons: the list prices of the models, the model mix itself, or the cache hit rate per session. Most of the August drop in closed-weight prices was a mix shift: GPT-5.6 Luna, at about $0.07 per million tokens, became the most used paid closed model in the days after its launch.

Individual lab model share

Typically, each lab's traffic is concentrated in a few models, and the newest release usually takes over within a month. Pick a lab to see how its token and spend mix moved through the year.

Open-weight labs lead in growth

Since May, Z.ai's monthly tokens grew about twentyfold and DeepSeek's about twelvefold. On the spend side, Z.ai's grew about sixfold, Anthropic's rose 76%, and MiniMax's fell 20% even as its tokens grew substantially.

Growth for individual labs is rarely linear. In an ever more competitive release calendar, the timing of a new model can move a lab's growth sharply in a single month.

The task leaderboard features two models prominently

We classify a sample of requests into the task categories on our rankings page. By tokens, DeepSeek V4.1 Flash leads most coding and agent tasks. Switch to spend and Claude Opus 5.5 leads most of the board.

Other models making an appearance include Jev 1.13, the decision model from TypeSafe, and GLM 5.3 Flash, with strong showings in both Data Extraction and Data Transformation.

The closed-weight labs (and DeepSeek) lead in user retention

There are many ways to measure retention for a model or a lab. Here we took the accounts that sent at least 100 paid requests in a month to models from a given lab and checked how many used the same lab again the next month. The threshold filters out one-off trials, and the order of labs was very similar at lower and higher cut-offs.

From September to October, Google kept 78% of those accounts, Anthropic 74% and OpenAI 74%. The leading open-weight labs tend to have lower retention month over month, and October was no different.

Decision models reached 21% of requests

Decision models (or System One models, in TypeSafe’s nomenclature) judge a narrow question, such as routing, moderation or a yes-or-no check. They were only born as a category in mid-September and rose to 21% of all requests by the week of October 26.

Text models here means every request to the chat completions API, including requests with image inputs. On tokens or spend instead of requests, decision models would take a much smaller slice of weekly traffic.

Open-weight inference providers: Together climbs to first

Among providers serving open-weight models, Together moved from 18th in June to first in October by total tokens served. DeepSeek fell to ninth in September and returned to second with V4.1 Flash, and InferenceNet entered the top ten at fourth.

Notes on methodology

• Data covers requests through the OpenRouter API from January 1 to October 31, 2026; the agentic chart starts in September 2025. Weekly charts use weeks that start on Monday.
• Tokens are prompt plus completion tokens. Spend is billed usage. Labs are model makers, and a model is open-weight when its weights are publicly released.
• Each API key is classified every month from seven weighted signals: tool-call rate, input-to-output ratio, turns per request, cache hit rate, tools offered per request, time between requests and cancel rate. A score of 0.50 or more is agentic, 0.25 to 0.49 is mixed and below 0.25 is human.
• Paid traffic excludes free model variants.
• Effective price, cache hit rate and provider rank exclude bring-your-own-key (BYOK) traffic and stealth models. Cache hit rate is cached prompt tokens over prompt tokens on traffic from agentic keys, for providers with at least 1T agentic tokens that month.
• Task leaders come from a sample of requests classified into the rankings task taxonomy and weighted by sampling rate.
• Retention counts accounts with at least 100 paid requests to a lab's models in a month, and calls an account retained if it sent any paid request to that lab the next month. Banned accounts are excluded.
• Decision models are requests to the decisions API. Text models are requests to the chat completions API, including requests with image inputs.