Today's take
The strongest signal today is a shift toward making agents improve over time and operate reliably. Skill learning, auditing and local inference are attracting open-source attention, while coding agents occupy prominent positions in application usage: the execution layer has concrete workflows to serve.
Model usage tells a more uneven story, with growth at the top alongside sharp gains further down the rankings. Model and app figures cover 2026-10-02 and represent only traffic routed through OpenRouter with public attribution; tokens measure usage, not users or market share.
Open source
Three themes stand out across the Top 50: turning experience into skills, adding execution controls and simplifying local deployment. Rank gains indicate attention, while star growth alone does not establish deployment scale.
- Skills become maintained assets: tigerless-labs/autoharness extracts, updates and prunes skills from real sessions. walkinglabs/learn-harness-engineering rose from #30 to #12 with +640 stars in 1 day, suggesting that practical tooling and educational entry points are developing together.
- Execution gains a control layer: auditing project ifixai-ai/iFixAi added 1,049 stars, while parallel-agent workspace stablyai/orca added 967. These address checking outcomes and managing execution, extending the focus beyond task completion alone.
- Local inference offers another engineering path: Niko1221/Strata reached #1 with consumer-hardware installation and compatible local APIs. This signals interest in easier deployment, without establishing verified performance or cost advantages.
Four projects best capture the movement across learning, deployment, tooling and quality control.
- #7 tigerless-labs/autoharness: +917 stars in 1 day, reaching 7,096 and climbing from #26. Its distinctive idea is to maintain an evolving skill layer from working sessions.
- #1 Niko1221/Strata: +1,375 stars in 1 day, reaching 5,341 and moving from #2. Installation tooling and local API compatibility make local models easier to connect to existing workflows.
- #15 earendil-works/pi: +601 stars in 1 day, up from #29. It packages a unified model API, agent loop and CLI; the app ranking also records pi at 456B tokens, providing a directionally consistent usage signal.
- #4 ifixai-ai/iFixAi: +1,049 stars in 1 day, reaching 19,000. Independent auditing addresses whether an agent actually performed the requested task, highlighting interest in quality controls.
Models
Usage leadership and growth leadership are distinct. The movers list covers models outside the top ten but within the Top 50, so rapid growth does not imply top-tier volume.
- Leaders continue expanding: Space Bunny Alpha processed 5.89T tokens, +14% day over day; DeepSeek V4.1 Flash (batch) recorded 3.48T, +3%; GLM 5.3 Flash (batch) reached 1.61T, +12%.
- Fastest growth comes from lower ranks: #36 Hy4 preview recorded 61.7B, +219%; #15 GPT-6 Luna Pro (batch) reached 214B, +117%; #24 Claude Opus 5 (batch) recorded 121B, +94%.
- Related models diverge: GPT-6 Luna (batch) fell 27% to 729B, contrasting with Luna Pro's growth. These figures do not establish a transfer of traffic between them.
Apps
Coding agents appear in both high-volume and growth rankings, adding usage evidence alongside open-source attention. Growth ranks compare usage with the average of three preceding equal-length periods; percentage gains were not supplied.
- Memory and skills have a usage counterpart: Hermes Agent leads at 2.01T tokens. Its persistent memory and reusable-skills positioning aligns with the open-source emphasis on accumulated experience.
- Coding workflows occupy leading positions: Claude Code recorded 1.23T, Kilo Code 1.14T, Cline 1.06T and Codex 1.01T. The top five include three CLI agents and one IDE extension.
- Developer tools lead growth: Freebuff ranks first in growth with 656B tokens, followed by Codex, omp, Claude Code and OpenHands, at 1.01T, 621B, 1.23T and 130B respectively.
What to watch
Watch whether skill learning, auditing and local deployment tools continue to advance alongside agent usage; today's evidence supports thematic alignment, but cannot establish causal links between individual repositories, apps and models.
