The State
of AI.

OpenAI had a dense day: GPT-5.6 is now the live default model inside Microsoft 365 Copilot, meaning it's already deployed to hundreds of millions of enterprise seats — not a beta, not a preview. Separately, ChatGPT Work reframes ChatGPT itself as a long-running autonomous agent, capable of executing multi-hour goal-directed tasks across apps and files. For builders, the M365 Copilot positioning matters most: if you're building on the Microsoft graph or M365 APIs, your assumptions about underlying model behavior just changed. The benchmark reliability post on SWE-Bench Pro is also worth reading — OpenAI is actively challenging how coding model quality gets measured, which has downstream implications for how you evaluate AI coding tools in your own stack.

Anthropic is executing on distribution as aggressively as on models today. Claude Sonnet 5 drops as a new mid-tier option, Claude Cowork goes mobile and web (expanding the surface area where users can compare it directly against ChatGPT Work), and the Claude Apps Gateway makes Claude natively accessible via Amazon Bedrock and Google Cloud — with Microsoft Foundry already GA. For enterprise builders, the multi-cloud story is now complete: you can deploy Claude entirely within your existing cloud provider relationship. The centralized MCP connector auth update is quieter but operationally important — it means platform teams can govern all Claude agent tool connections from a single control plane rather than managing credentials per-agent.

Google's headline contribution today is the managed agents expansion in the Gemini API, specifically the addition of remote MCP support alongside background task execution. This is a meaningful architectural signal — by adopting MCP at the managed agent layer, Google is endorsing it as an interoperability standard rather than pushing a proprietary tool-calling format. Builders building on Gemini's agent infrastructure can now connect external MCP-compatible tool servers directly, which dramatically expands what Gemini agents can do without requiring Google to build every integration. The practical guidance: if you haven't yet exposed your internal tools as MCP servers, this is another reason to prioritize that work.

Meta's builder-relevant news today is Muse Image — the first image generation model from Meta Superintelligence Labs, now live in Meta AI. Whether the model becomes available via API for third-party builders is the key unanswered question; until then it's a consumer product launch to watch. The Canada data center groundbreaking (1GW, AI-optimized) is infrastructure news that signals Meta is betting its compute future on AI workloads at a scale few companies can match. For builders, the AI glasses FAQ and the engineering post on ultra-narrow batteries are worth skimming if you're building wearable or edge AI applications — Meta is further ahead in AI-integrated wearables than most people track.

The open source ecosystem had a strong day anchored by two funding announcements with immediate builder implications. Together AI's $800M Series C came paired with Provisioned Throughput — reserved inference capacity for frontier open models with a real SLA and token-based pricing up to 90% cheaper than proprietary APIs. This is the most credible enterprise-grade open model infrastructure offering to date. Meanwhile Ollama confirmed 8.9M developers and $88M raised, and shipped a genuinely significant performance update: Gemma 4 on Apple Silicon via MLX is up to 90% faster for coding agent tasks with multi-token prediction enabled in version 0.31. Mistral's Studio prompt management system is the quieter but operationally useful release — version-controlled prompts with ownership and traceability is the kind of boring-but-necessary infrastructure that prevents production prompt drift.