Moonshot shipped the largest open weight model yet while Alphabet’s stock took a hit on word that Gemini 3.5 Pro is behind schedule on coding, and a new manual for coding agent harnesses landed the same week Anthropic published its own migration playbook.
Frontier Models: The Race for Scale and Speed
Two labs moved the frontier this week for very different reasons: one shipped, one stalled.
- Moonshot Ships Kimi K3, the Largest Open Weight Model Yet. Moonshot’s new 2.8 trillion parameter model lands today with a 1 million token context window, with open weights following on July 27.
- Alphabet Stock Drops 4 Percent on Gemini 3.5 Pro Delay Report. Bloomberg reported the flagship model is running months behind because its coding output trails internal targets, just as OpenAI and Meta ship faster coding models.
The Harness Layer: Where Coding Agents Actually Live
The infrastructure around coding agents, not the models themselves, is where this week’s real engineering news sits.
- LM Studio Launches Bionic, a Local First Coding Agent. Bionic pairs local and cloud open weight models with coding, document and voice tools, betting privacy minded builders will pick control over frontier capability.
- A New Manual Maps What Coding Agents Actually Do. Harness Handbook reorganizes agent codebases by behavior instead of file structure, and its authors say the approach cuts a planner’s search cost while improving accuracy.
- Anthropic Publishes Its Playbook for AI Code Migrations. The six step, multi agent process comes from a Zig to Rust rewrite of the Bun runtime and a Python to TypeScript port.
- OpenAI Splits Codex Into Three GPT-5.6 Difficulty Tiers. Sol, Terra and Luna route coding work by difficulty, turning prompt writing into a cost and latency decision.
Agents in the Wild: Benchmarks, Routing and Real Deployments
Beyond the lab, agents are being benchmarked, routed and put to work inside real companies and real products.
- New Harness Pushes Opus 4.8 to 99 Percent on ARC-AGI-3. Schema has frontier models write executable theories of unfamiliar games, reaching 99 percent RHAE with Opus 4.8 and Fable 5 on the benchmark.
- Why Task Specific Routers Beat Panel Style Model Routing. OpenRouter’s Fusion and Cognition’s new router both chase cheaper frontier quality answers, but the real savings come from routing built around one workflow at a time.
- Replit Says Internal AI Agents Tripled Engineer Output. Replit reports internal agents nearly tripled per engineer code output in six months while incident and review metrics held flat, though the data is entirely self reported.
- Google Wires Instacart, Canva and YouTube Music Into AI Mode. AI Mode can now fill an Instacart cart, pull Canva templates and save playlists to YouTube Music, turning Search into a task completion layer.
The Economics of AI: Who Pays and Who Profits
The money underneath the model layer is moving fast, from inference valuations to enterprise spend controls.
- Fireworks Hits $17.5 Billion Valuation on Cheap Model Demand. The Nvidia backed inference startup topped $1 billion in annualized revenue and raised $1.5 billion, evidence open models are eating into frontier lab budgets.
- Ramp Expands AI Token Spend Tracking Across Major Providers. Ramp’s expanded tool centralizes AI costs from OpenAI, Anthropic and Google, giving finance teams provider level visibility they have lacked until now.
Today’s Quick Hits
- NVIDIA Nemotron 3 Embed Tops RTEB, Ships Open. NVIDIA’s 8B flagship embedding model ranked first on RTEB, and the company open sourced it plus two smaller variants for RAG and agent memory.
- GitHub Ships an Official SDK for Embedding Copilot Agents. The Copilot SDK lets developers invoke GitHub’s production agent runtime programmatically across six languages, competing directly with Anthropic and OpenAI’s agent frameworks.
- Google Renames NotebookLM to Gemini Notebook, Adds Cloud Compute. Google folded its research tool into the Gemini brand and gave every notebook a secure cloud computer that can write and run code.