Today's take
The strongest signal today is a shift toward agents that can execute, remember and accumulate experience. lexmount/moli reaches #1, while tigerless-labs/autoharness jumps from #50 to #6: attention is moving toward the infrastructure and knowledge layers needed for sustained work.
The latest available model and app usage snapshot covers October 6, 2026. Hermes Agent leads app usage, with coding agents clustered near the top—a thematic match with open source, although the data does not connect particular apps to particular model gains.
Open source
Three themes stand out across the Top 50: execution infrastructure, agent capabilities that accumulate over time, and creative production workflows. Stars and ranking changes indicate developer attention; they do not establish adoption.
- Execution infrastructure moves to the foreground: #1 lexmount/moli adds 1,829 stars and rises from #2 to #1. Its Rust headless browser targets agent execution. #2 Niko1221/Strata adds 1,787 stars, focusing on local inference on consumer hardware and compatible APIs—a complementary signal of interest in control over deployment.
- Skills and memory form complementary layers: #6 tigerless-labs/autoharness adds 1,008 stars, climbing from #50 to #6 by turning real sessions into skills that can be updated and pruned. #12 thedotmack/claude-mem adds 672 stars and rises from #22 to #12, compressing session records and restoring relevant context later. Together, they address experience accumulation and continuity across sessions.
- Creative tooling expands toward production workflows: #4 storytold/artcraft adds 1,139 stars and jumps from #45 to #4 with a crafting engine for artists, designers and filmmakers. Alongside #13 heygen-com/hyperframes, which turns HTML into video, it points toward organizing generation into usable production processes.
Models
The October 6 snapshot shows a large usage leader alongside faster growth elsewhere. These figures measure tokens processed through OpenRouter, rather than shares of the overall model market.
- The leader contracts but remains ahead: DeepSeek V4.1 Flash (batch) processes 4.89T tokens, down 22% from the previous day. GLM 5.3 Flash (batch) follows at 1.62T, up 4%.
- Top-ten gains span multiple models: GPT-5.6 Sol (batch) reaches 619B, up 176%; GPT-6 Luna (batch) reaches 893B, up 32%; Claude Opus 5.5 (batch) reaches 551B, up 44%.
- Models outside the top ten accelerate: #20 Hy4 preview records 132B tokens, up 378%, while #11 Claude Sonnet 5.5 (batch) records 315B, up 187%. Their growth merits tracking, although their absolute usage remains below the leaders.
Apps
Publicly attributed app traffic on the same date highlights persistent personal agents and development workflows. Four of the five highest-usage apps are classified as CLI agents, and that category also leads the rising list.
- A personal agent leads usage: Hermes Agent records 1.52T tokens. Its cross-session memory and reusable skills align thematically with the open-source interest in memory and skill infrastructure.
- Coding workflows consume substantial tokens: Claude Code records 1.25T, Kilo Code 569B, pi 528B and Codex 508B. These are usage volumes, not user counts.
- Execution tools lead the growth ranking: Compared with the average of the preceding three equal-length periods, pi, Claude Code and Strix rank first through third for growth. Strix records 104B tokens; percentage increases are not supplied, so the growth magnitude cannot be quantified.
What to watch
Watch whether interest in agent memory, skills and execution infrastructure continues alongside app usage. The current evidence supports a shared direction, but does not establish causal links between individual repositories, apps and models.
