Stream number 13 opens with a technical complaint that becomes the whole day's theme: since the Opus 4.7 launch, token consumption is up around 35%, and Sonnet's weekly limit was already at 100% usage hours before the stream. On top of that, the old recurring problem — VS Code with disconnected workspaces, conversation sessions lost when closing a terminal — had become unsustainable for someone running six Claude Code panels at once trying to orchestrate with Opus and execute with Sonnet.
The turn happened in the early morning: an alert from his own AI-launch monitoring system on WhatsApp flagged DeepSeek V4 Pro, then Kimi K2, until it landed on Mimo V2.5 Pro, from Xiaomi — a model almost as smart as Opus 4.7, faster, and costing a fraction of the price. Discovering a repository called "Free Cloud Code" showed it was possible to take the shell (the framework) of the Claude Code CLI and plug another LLM in behind it via Xiaomi's official protocol — without relying on OpenRouter or losing the official API's rate limit.
That's where Overclock was born, live: a terminal with multiple panels organized by workspace, persistent sessions (the problem of losing context on reopen got solved on screen, with Claude Code restoring the session after a restart), and a provider system letting you run Opus 4.7 as orchestrator/planner in one panel and Mimo V2.5 Pro as executor in another, in parallel. The install got tested live with the Jarvis assistant doing QA — mapping onboarding bugs, model defaults that didn't persist, and mislabeled panels — all logged in a markdown checklist to pick back up after restarting the session.
The commercial launch came in the middle of the stream itself: an Overclock subscription at overclock.sh for R$47/month (with a lifetime 50% founder discount for anyone joining that day), Mac-only for now, with Windows promised for the coming days because one person on the team only uses that platform. Lachuk revealed he'd already installed Overclock on his agency's own employees' computers to scale Mimo usage without paying a full subscription per person — the real reason the tool existed before it became a product.
The stream closes with an attempt to benchmark four Mimo variants (V2.5 Pro, V2.5, V2, and V2 Flash) running the same task in parallel — which gave inconsistent, visibly "lying" results, forcing the decision to build a real benchmark script, outside the models' own control, for the next day.