Routine work slides to the open side

Mozilla's State of Open Source AI v1.1, published 15 September, puts Moonshot's Kimi K3 three points behind Fable 5 on Artificial Analysis's composite score and measures the METR time-horizon gap at about 4.4 months. Closed systems lead on 8-to-12-hour jobs; four months later the open side reaches that band. That is a closing window tied to the release cycle, not a permanent lab monopoly.[1]

On Vals AI's shared-harness Terminal-Bench 2.1 run, GLM 5.2 came within a point of Claude Opus 4.7 and 4.8 at about 5 times lower cost per completed task. Eight of OpenRouter's top ten models by August 2026 token volume ship open weights. Raffi Krikorian says the premium is earned in expert work, high-intensity retrieval and long context, and that the choice is workload-specific. As weight files become substitutable, a team's scarce resource is the tool arrangement it binds the model to, not which checkpoint it downloaded.[1]

The tool contract is chasing the model

The same day Meta opened the WhatsApp Business Tools MCP server. MCP is Model Context Protocol: a bind that lets an agent discover a tool schema at runtime. Claude, Cursor, Codex or ChatGPT can run account creation, number verification and Cloud API registration over chat instead of hopping through Console and Business Manager. That is an open analogue of the ready-made tool layer closed models sell beside the weights: whoever owns the checkpoint, the contract that finishes the job sits on Meta's server.[2], [1]

Krikorian's warning piles up here. Closed frontier systems arrive with compliance packaging and accountability; open weights still usually hide the data recipe. The WhatsApp server does not reverse that tension: it speeds the agent and ties revoke-and-scope to permission on a Meta account. The developer stays free to pick a model and is not free to rebuild the WhatsApp job on another protocol. The signal to watch is teams defaulting sub-8-hour work to the open side and MCP onboarding measurably replacing the console tour; otherwise the pack stays a chat shortcut.[2], [1]