Microsoft just called Copilot a toy

Mixed signals all weekend. Anthropic announced Friday that Claude Pro and Max subscribers can no longer use OpenClaw or other third-party tools without paying extra—135,000+ OpenClaw instances were running at the time, and the flat subscription was covering usage that cost Anthropic 5x or more what subscribers paid. Boris Cherny said “our subscriptions weren’t built for these usage patterns,” and Anthropic is offering a one-month credit plus 30% off pre-purchased usage bundles to soften the landing. The timing is notable: just days after the Claude Code quota burnout saga, Anthropic is drawing another line around agentic usage costs. Meanwhile, Microsoft’s own terms of service describe Copilot as “for entertainment purposes only”—a disclaimer updated in October 2025 that went viral this weekend. Microsoft called it “legacy language” that “will be altered with our next update” but offered no timeline, even as enterprise customers pay $30 per month for the product. Neither Google nor OpenAI have used that exact phrase for their flagship AI products. One company is telling developers its AI costs more than they thought; the other is telling lawyers its AI was never serious in the first place.

And while the incumbents sort out pricing and liability, Google is giving the models away. Gemma 4 shipped April 2 with four model sizes under Apache 2.0—a first for the Gemma family, which previously used a custom license. The 31B model ranks #3 on Arena AI’s text leaderboard; the 26B MoE variant activates only 3.8 billion parameters per token yet scores 88.3% on AIME 2026 and fits on a single RTX 4090. At the small end, the E2B model hits 133 prefill tokens per second on a Raspberry Pi 5—CPU only, no GPU. Google shipped the same day as Alibaba’s Qwen 3.6-Plus, and the Go and Java 1.0.0 SDKs for Google’s Agent Development Kit landed in late March—ADK itself has hit v1.28.1 with 50 releases in its first 12 months and over 18,700 GitHub stars. We dug into which Gemma 4 size actually makes sense for most developers—the biggest model is probably the wrong choice. The pattern is clear: Anthropic restricts, Microsoft disclaims, and Google floods the zone with open weights and permissive licensing.

The open-weight push matters most at the framework layer, where developers are choosing how to wire these models into production. Google ADK’s 50-release cadence puts it alongside LangGraph and CrewAI as a serious multi-agent option, and the SDK now supports SequentialAgent, ParallelAgent, and LoopAgent patterns natively in Python. We walked through the full ADK tutorial—session state, the four gotchas that trip up production systems, and why Google’s orchestration primitives differ from LangGraph’s graph-based approach. The 26B MoE model running 3.8 billion active parameters on consumer hardware changes the calculus for local agent development: a multi-agent pipeline that would have required cloud API calls six months ago now fits on a $1,600 GPU. Between Anthropic charging more for agentic access, Microsoft hedging on reliability, and Google handing out Apache-licensed models with a full agent toolkit, the cost of building AI systems is diverging fast—and the licensing terms may matter more than the benchmarks.

Latest from PulseMark

Build a Multi-Agent System with Google ADK in Python SequentialAgent, ParallelAgent, LoopAgent patterns, session state management, and the four production gotchas Google’s docs don’t warn you about. Plus how ADK’s orchestration compares to LangGraph’s graph-based routing. Read the tutorial →

Gemma 4 Model Sizes: Which One Should You Actually Run? Four sizes from 2GB edge to 62GB workstation—the 26B MoE scores 88.3% on AIME 2026 with just 3.8B active parameters. Why the biggest model is probably wrong for your workload, and what Apache 2.0 licensing means for commercial deployment. Read the analysis →