OpenRouter Usage: The Agent Boom Comes in Waves

OpenRouter's agent boom has changed shape several times. OpenClaw produced a huge spring spike. Hermes became the leading app in May. More recently, Cline, pi and Claude Code each more than doubled their attributed tokens across two comparable four-week periods, while OpenClaw grew about 5%. The latest growth extends well beyond the app that drove the earlier surge.

Three successive leaders in the Coding Agents chart: Kilo Code, OpenClaw and Hermes. These are OpenRouter-attributed tokens, not total app usage or user counts. The last plotted week ends September 6, 2026.

This is a glimpse through one gateway. Claude Code users who want Anthropic models can call Anthropic directly, outside this data. People choosing OpenRouter could be there specifically to try alternatives. The same selection problem applies to other apps with several ways to reach their models. We can't infer their overall popularity, or anyone's lost customers, from this slice.

What we can do is follow each app's observed traffic over time. For the first installment of our occasional OpenRouter Observed series, we examined the app categories, expanded rankings and individual model histories. The strongest finding is that "agents are growing" conceals several different stories: a launch wave that recedes, a persistent agent that takes the lead, and renewed growth across coding interfaces.

The leading app changed twice

We read all 52 weekly points in OpenRouter's Coding Agents chart (opens in new tab), from September 8, 2025 through the week beginning August 31, 2026. The largest named app changes only twice in that captured history:

• Kilo Code leads the opening period.
• OpenClaw takes the lead in the week beginning February 2.
• Hermes Agent takes over in the week beginning May 11 and remains ahead through the last week.

OpenClaw's March 23 week reached 5.53 trillion tokens, about 63% of the displayed category total. By the August 31 week, it was at 1.13 trillion. That is a substantial fall from the peak, though it still leaves more than a trillion attributed tokens in a week. We can't tell how much of the change came from usage, routing choices or attribution.

Hermes followed a different path. Its displayed weekly volume rose from 1.92 trillion when it took the lead in May to 8.06 trillion in the August 24 week, before falling to 6.98 trillion the next week. The succession doesn't prove that OpenClaw's users moved to Hermes. It establishes which app was carrying the largest observed token load at each point.

These agents live in different places

The change in names also brings a change in the kinds of interface represented near the top. OpenClaw (opens in new tab) connects an assistant to messaging apps and describes work such as managing a calendar or an inbox. Hermes (opens in new tab) combines persistent memory, reusable skills, scheduled work and access through messaging, a terminal and desktop apps.

Hermes crosses OpenRouter's category boundaries. The badges describe the app; they don't classify the task behind every request. This rolling app-page total isn't used in our weekly comparisons.

Meanwhile, Cline works inside an IDE, Kilo spans editors and a command line, and pi describes itself as a customizable agent harness (opens in new tab): the software that manages a model's tools, context and execution. Pi can also be embedded inside another application. Its own documentation names OpenClaw as an example, which is another reason to avoid treating the app list as a set of unrelated products or independent audiences.

Our interpretation is that the workflow around the model is becoming a more visible part of the competition in this slice of OpenRouter. Memory, scheduling, tool access and the ability to adapt an agent to your work distinguish these interfaces. Their traffic histories give us reason to investigate those capabilities, though they don't identify which capability caused growth.

The earlier leaders were already coding agents. This history therefore doesn't establish a universal transition from chatbots to agents. It shows different agent interfaces gaining and holding substantial traffic within the gateway.

The latest growth is spread across several apps

For a recent comparison, we used two completed 28-day windows: July 13-August 9 and August 10-September 6. Each consists of four Monday-Sunday weeks from the same category chart. All six apps below have an individually named entry in every week of both windows.

Cline, pi and Claude Code grew faster than the displayed category total. This compares each app with its own earlier volume, rather than treating app ranks as a measure of market share.

T means trillion tokens. Values expand OpenRouter's rounded weekly figures, rather than implying precision down to individual tokens.

Hermes remains the largest app by volume here, but Cline, pi and Claude Code grew faster. OpenClaw looks closer to a plateau during these eight weeks. That's a more useful account of the recent surge than attributing it all to the most recognizable personal agent.

Nor is the latest movement relentlessly upward. The category total fell from 26.5 trillion in the August 24 week to 24.3 trillion the following week. A rising four-week total and a falling latest week can coexist. Future installments can test whether that dip was brief or the start of a longer change.

These remain token-growth figures. More agent steps, longer context, repeated attempts or more work per person can produce growth without a matching increase in users. Apps can also change how they identify traffic or which gateway they send it through.

Model launches can distort the app story

There is another layer underneath those app totals. The same app can change its model mix sharply while its weekly average looks steady.

In the Cline (opens in new tab) and Claude Code (opens in new tab) daily charts, we followed GLM 5.3 Flash across August 24-30 and August 31-September 6. Its identified token share rose from 34.9% to 41.0% in Cline, while staying near 45% in Claude Code.

A shorter comparison within each app's daily chart. Both weeks begin after the preview launch; their shares combine the preview and named model identifiers.

The identity check matters: OpenRouter identifies Ox Alpha as GLM 5.3 Flash (opens in new tab). The preview arrived on August 20 and the named release on August 26. Tracking only the new name would erase the preview's traffic; tracking only the old one could make a rename look like abandonment.

One model under two identifiers. We add the separately attributed usage rows to follow its history.

Even with that correction, weekly averages hide a lot:

Cline's weekly gain includes a dip and recovery. Claude Code's similar weekly shares hide a preview spike and retreat. The timing doesn't establish what caused either move.

Cline was already around 40% on August 24 and 25. It dropped to 8.0% on August 27, then returned to roughly 38-42%. Claude Code's preview spike reached 72.4% on August 22, followed by readings below 24% on August 27 and 28. Neither history describes a smooth adoption curve.

The volume check supports the weekly shares: Cline's identified model tokens grew 34.4%, faster than its overall 14.3% growth across those two weeks. In Claude Code, model tokens and app totals both fell by about 4%. Those two-week figures answer a different question from the 28-day app growth above.

Price adds another complication. On September 7, the named model's page (opens in new tab) advertised a 50% discount via ZAI through September 9, 2026, at 16:00 UTC.

The promotion is part of the dated snapshot. Observed use during this period doesn't establish how demand would hold at undiscounted prices.

App growth and model growth can overlap. These observations don't isolate the effects of prices, defaults, workload changes or people trying a new release. Following the app and the model together makes those possibilities easier to see.

Other apps show continuity

The swings in coding don't describe every app we inspected. On Janitor AI's page (opens in new tab), DeepSeek V3.2 was the largest individually named model on all 30 completed days from August 8 through September 6.

A stable named leader over the same month. "Others" combines several models, so leading the named entries doesn't establish a majority of tokens.

That continuity belongs beside the launch spikes. It doesn't prove that roleplay users value familiarity while coding users chase novelty. It shows that a single story about AI demand won't explain even this limited set of apps.

Category labels need care too. Hermes appears in both Coding Agents and Productivity. Other apps cross Creative and Entertainment. Adding category totals would count some traffic more than once, and a token attributed to a coding-capable app doesn't prove that its task was programming.

What this says about the app around the model

Our reading is that these histories show several ways an AI app can become useful: an agent that stays available across messaging channels, one that remembers recurring work, or a coding tool whose context and execution can be customized. The observed demand is spread across those interfaces, and their fortunes change at different speeds.

That helps explain why the routing layer matters. Stripe announced an agreement to acquire OpenRouter on August 19 (opens in new tab), emphasizing the problem of choosing among changing models, prices and performance. OpenRouter's announcement described closing conditions (opens in new tab). The agreement supplies timely context; the app data supplies a way to inspect the routing problem.

Builders are making several connected choices: the workflow people use, the model that does each job, and whether to reach it through OpenRouter or a direct provider connection. Even with the model and task held constant, gateway controls can change the bill, as our OpenRouter and Vercel tests showed.

We build Cumbersome, an AI app for iPhone and Mac with a choice of models, so this separation matters to us too. A useful workflow can outlast a particular model release. Usage charts identify candidates to investigate; they still leave task quality and compatibility to test.

For this series, the next questions are specific: does coding-agent growth persist beyond the latest promotions, does OpenClaw's observed plateau continue, and does Hermes retain its lead as other interfaces gain volume? Repeating the same dated comparisons will be more informative than declaring a new winner from every spike.

Data and method

I collected public OpenRouter data for Folding Sky on September 7, 2026, Pacific time. The app-growth comparison sums the same category chart's weekly points into two completed 28-day windows. The model comparison separately sums daily app tooltips into two completed seven-day windows. It excludes September 7's partial daily data. We don't mix the live leaderboard's period labels with these explicitly dated chart points.

The UI rounds values and groups smaller entries into "Others." Missing names are unknown, rather than zero; model sums are identified subtotals. Weekly model shares divide summed model tokens by summed app totals, rather than averaging daily percentages. These are repeated app aggregates, not a fixed cohort or an experiment. App membership, attribution and routing can change.

OpenRouter's attribution documentation (opens in new tab) explains how apps identify their traffic. Direct-provider routes, other gateways and unattributed requests are outside these app charts. No figures here count developers, successful tasks, revenue or global market share.