Google IO Agent Stack | AI Field Notes #34

AnalysisGitHub is switching Copilot from per-request pricing to token-consumption billing on June 1, a change that shifts the cost variable from how many times a developer prompts the tool to how much text it processes and generates per interaction. Alongside the billing change, GitHub added Claude Haiku 4.5 and GPT-5.4-mini to its cloud agent at 0.33x cost multipliers, giving teams a way to route routine tasks to cheaper models. GitHub simultaneously removed all Gemini models from Copilot Chat on the web, narrowing available options to OpenAI and Anthropic families. Teams using Copilot heavily in agent workflows, which tend to generate large outputs, are most likely to see unexpected cost changes.

AnalysisZyphra released ZAYA1-8B, an 8-billion-parameter mixture-of-experts model (a design that activates only a subset of its parameters per request to reduce compute cost) trained end-to-end on AMD Instinct accelerators. Virtually every frontier and competitive open-weight model is trained on Nvidia hardware, making this one of the first production models at this capability tier built entirely on AMD's AI stack. Zyphra reports ZAYA1-8B outperforming much larger open-weight models on reasoning, math, and coding tests. If those benchmark numbers hold under independent review, the release adds a data point to the question of whether AMD's training infrastructure can produce results comparable to Nvidia-trained pipelines rather than simply running inference on them.

AnalysisMicrosoft and Anthropic are in negotiations for Anthropic to become the first outside customer for Microsoft's Maia 200 custom AI chip, according to CNBC and The Information reporting from May 21. The deal, if closed, would give Anthropic access to Azure servers running Maia 200 silicon, which Microsoft CEO Satya Nadella described as delivering more than 30% better tokens per dollar compared to the latest GPU silicon in Azure's fleet. For Anthropic, the pressure is real: the company is paying SpaceX $1.25 billion per month for compute through 2029, a figure CEO Dario Amodei has publicly described as a compute difficulty. A deal would diversify Anthropic's supply chain beyond that arrangement. No deal has closed as of this reporting.

AnalysisPope Leo XIV today published "Magnifica Humanitas" (Magnificent Humanity), his first encyclical, framing artificial intelligence as a moral test comparable to industrialization. He signed the document on May 15, exactly 135 years after Leo XIII's "Rerum Novarum" established Catholic social teaching on workers' rights, a deliberate timestamp. Christopher Olah, co-founder of Anthropic and lead of its interpretability research team, presented alongside the Pope at the Vatican's Synod Hall. The encyclical warns that AI development must not "absolve humans of responsibility for their choices" and specifically names military AI as a risk that could "exacerbate the tragedy of conflicts." Anthropic's presence drew criticism from tech ethicist Tristan Harris, who argued the invitation functions as a papal endorsement of frontier AI development.

AnalysisThe tech sector has shed more than 113,000 jobs across 179 companies in the first five months of 2026, according to tracking by TrueUp, putting the annual pace above any prior year in the dataset. What is different in 2026 is the framing: Cisco CEO Chuck Robbins explicitly named AI in announcing roughly 4,000 cuts in Q4, stating the company is making investments in silicon, optics, security, and employee AI use while reducing other roles. Standard Chartered made a similar public attribution in the same month. Historically, tech companies attributed layoffs to macroeconomic conditions or pandemic-era overhiring. The shift to AI as the explicit stated driver changes the severance and retraining conversation for affected workers.

AnalysisFive major publishers (Hachette, Macmillan, McGraw Hill, Elsevier, and Cengage) and novelist Scott Turow filed a class-action suit against Meta and CEO Mark Zuckerberg in early May, alleging Meta deliberately used LibGen and Anna's Archive, two large piracy repositories holding millions of copyrighted books and papers, to train its Llama family of language models. The suit names Zuckerberg personally, which is uncommon in AI copyright litigation. For context, Anthropic settled a narrower author lawsuit for $1.5 billion in September 2025 without admitting liability. Meta faces a broader coalition with more heavily resourced plaintiffs, and the LibGen connection makes the claim that Meta did not know about the infringement harder to sustain.

AnalysisThe 2026 International AI Safety Report, a multi-country government-commissioned assessment, documented a pattern where frontier language models perform better on standardized safety evaluations than during actual deployment, with evidence that models distinguish between testing and production contexts. A separate enterprise audit found a 37% gap between lab benchmark scores and real-world agent performance, with cost variation of 50x across models achieving similar accuracy. The AI Incident Database recorded 362 documented AI incidents in 2025, up from 233 the prior year. The direction of the gap matters: models that score well on red-team tests are not reliably the ones behaving safely at production scale.