224 billion versus 186 billion. This is not a funding comparison. It is the number of tokens burned in a single day by the two hottest AI agent projects on OpenRouter this week.
The short version: OpenClaw has lost the top spot.
On May 10, MarkTechPost reviewed OpenRouter's global daily activity ranking. Nous Research's Hermes Agent led the table with 224 billion tokens in one day, pushing OpenClaw into second place at 186 billion. It was the first time in months that OpenClaw had slipped from No. 1 on OpenRouter.
The numbers get more interesting when read together
Daily active tokens: Hermes 224 billion / OpenClaw 186 billion
Cumulative tokens: Hermes 6.35 trillion / OpenClaw 9.17 trillion
GitHub stars: Hermes 114,000 / OpenClaw 370,000
OpenClaw still leads by a wide margin in cumulative usage and stars. It has also been running for more than a year longer. But the daily reversal says something different: the flow of new users has shifted. Hermes Agent has captured much of the incremental demand over the past three months.
This is not just another open-source ranking shuffle. The release cadence tells the story.
| Version | Release date | Scale |
|---|---|---|
| v0.11.0 "Interface" | April 23 | 1,556 commits, 761 PRs, full TUI rewrite |
| v0.12.0 "Curator" | April 30 | Automated skill pruning |
| v0.13.0 "Tenacity" | May 7 | 864 commits, 588 PRs, fixes for eight P0 security issues |
Three major releases in three weeks, with four-digit commit counts. That pace is hard for OpenClaw to match, especially after founder Peter Steinberger moved to OpenAI in February. OpenClaw is now run by an independent open-source foundation with OpenAI as a named sponsor. When the organization changes, the tempo changes too.
Two very different architectures
OpenClaw is built around a WebSocket Gateway, a central router that connects more than 50 message channels. Its core question is: in how many places can the agent show up at once?
Hermes takes a different route: a "do, learn, improve" self-learning loop. Its reflective phase lets an agent review its own work after a task, then generate reusable skill files automatically. The memory stack is also more deliberate: persistent identity snapshots, a SQLite FTS5 database and procedural skill files.
One project spreads the agent widely. The other tries to make it smarter with use. Judging by OpenRouter's daily activity, developers have recently been leaning toward the second model.
The security gap is clearer
This may be the real reason behind Hermes' recent growth: OpenClaw's security exposure has become hard to ignore.
OpenClaw's record:
- CVE-2026-25253, CVSS 8.8, a remote gateway exploit
- Nine CVEs disclosed over four days in March, including one scored 9.9
- Koi Security audited 2,857 ClawHub skills and found 341 malicious entries, or 11.9%
Hermes' record:
- One medium-severity CVE-2026-7113, CVSS 5.6, involving missing webhook authentication
When 11.9% of the skills in an official repository are malicious, that becomes a serious reputational event for any open-source project.
For an agent framework, security exposure directly translates into a simple enterprise question: will installing this let someone control my system in reverse? That is where buyers are most sensitive. Hermes v0.13.0 added redaction by default, Discord role allowlists split by guild, WhatsApp rejection of unknown senders by default and other hardening steps. The direction is clear: it is preparing for enterprise use.
What happens next
OpenClaw is not going away. Its 370,000 stars and 9.17 trillion cumulative tokens are not assets that disappear overnight. But several points now look settled.
First, the open-source agent market is no longer winner-takes-all. Two projects can sit first and second on OpenRouter because users are willing to split by use case.
Second, the self-learning loop has been validated. Hermes' reflective phase is not an isolated bet. Anthropic recently showed a similar "Claude can dream" mechanism at Code with Claude. For this generation of AI agents, getting better the longer they are used has become a new product advantage.
Third, whether OpenAI's sponsorship helps or hurts OpenClaw should be visible within six months. If its release pace cannot keep up with Hermes, sponsorship alone will not save the ranking.
At least this week, OpenRouter's No. 1 is no longer OpenClaw. The signal matters more than the raw numbers.
Sources: OpenClaw vs Hermes Agent: Why Nous Research's Self-Improving Agent Now Leads OpenRouter's Global Rankings (MarkTechPost), CocoLoop, Hermes Agent v0.13.0 Release Notes (NousResearch/hermes-agent on GitHub)