🎯 Top 3 Things to Know
1. Google I/O 2026 opens Tuesday with a Gemini reveal that will define the week. The Monday morning keynote covers, in Google's words, "the latest Gemini model updates" and "agentic coding," widely read as a Gemini 4.0 plus Android XR glasses lineup. The friction the company is trying to address is straightforward. It has been more than two months since Google last shipped a frontier-tier model, while Anthropic dropped Claude Mythos Preview and OpenAI rotated GPT-5.5 Instant in as the default. Relevant for anyone tracking the competitive frontier and for builders sizing up Gemini's agentic coding pitch against Claude Code. Worth checking the I/O livestream for whether Google publishes head-to-head numbers against Claude Mythos Preview's 94.6% GPQA result, or quietly avoids the comparison. I/O 2026
2. PwC will roll out Claude to hundreds of thousands of professionals and certify 30,000 staff. The expanded alliance announced May 14 puts Claude Code and Cowork in front of PwC's full global workforce, with a new finance practice (the "Office of the CFO") built entirely on top of Claude. PwC cites a 10-week-to-10-day reduction in insurance underwriting and overall delivery improvements of up to 70% in current production work. The friction this addresses is the consulting-industry bottleneck. Big Four firms have historically been the slowest to operationalize new tooling, so a commitment of this size is a leading indicator that enterprise AI is past pilot phase and into operating-system territory. Worth watching for client case studies once the joint Center of Excellence ships its first reference architectures. Anthropic
3. Anthropic's $30B round at a $900B-plus valuation is tracking to close by month-end. No term sheet is signed yet, but Sequoia, Dragoneer, Greenoaks, and Altimeter are co-leading. The valuation rests on Q1 2026 ARR above $44B, up 80x year-over-year, with more than 1,000 customers spending $1M annually. If the round closes at the rumored mark, Anthropic passes OpenAI's $852B March valuation for the first time. The capital is earmarked for compute, not growth: AWS and Google Cloud commitments through 2027, plus the SpaceX Colossus 1 rental that doubled Claude Code rate limits earlier this month. Worth watching whether the round prices below $900B once Google reveals Gemini 4.0 on Tuesday and shifts the comparison set. Build Fast (Bloomberg summary)
🚀 Frontier Models & Features
- Claude for Small Business shipped May 13. A toggle inside Claude Cowork that connects to QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, and Microsoft 365, with 15 ready-to-run agentic workflows covering payroll, invoice chasing, campaign management, and month-end close. Every action requires user approval before executing. Anthropic
- GPT-5.5 Instant has been the default ChatGPT model across Free, Plus, and Pro since May 5. Scores 81.2 on AIME 2025 math (versus 65.4 prior) and integrates persistent memory across past conversations, uploaded files, and Gmail. TechCrunch
- Meta Avocado remains silent. Reuters sources had pointed to a May or June window. With I/O dominating the news cycle, a June reveal is the more likely path.
🔬 Research Worth Reading
Agentic Systems as Boosting Weak Reasoning Models (Sunkaraneni, Beneventano, Neumarker, Poggio, Galanti / MIT). arXiv
- TL;DR: Reframes inference-time agent committees as a classical boosting problem. A pool of weak proposers plus a verifier and a comparator can close the gap to much stronger reasoning models, when each component does its job well. The theory separates proposal coverage, local identifiability, progress, and diversity, so it is clear which weak link kills the lift.
- Stat: On verifier-backed tasks (code repair, theorem proving, program synthesis), a committee of weak calls matches frontier-tier accuracy at comparable or lower per-query cost.
- Apply it: Next time a reasoning eval comes back below target, before swapping in a bigger model, try a small committee of cheap calls behind a verifier on one hard slice and compare quality per dollar to the frontier baseline.
Case-Based Calibration of Adaptive Reasoning and Execution for LLM Tool Use (Pang et al., authors — see arXiv link). arXiv
- TL;DR: Introduces CAST, which retrieves similar past tool-use cases to calibrate how deeply the model should reason before each call, instead of burning a fixed deliberation budget every time.
- Stat: Reports gains in schema-faithful execution and task-level tool-use success while cutting unnecessary deliberation.
- Apply it: Audit recent agent traces for steps where the model over-deliberated on a routine API call. A case-based shortcut on the most repetitive tools is the cheapest latency win.
🏢 Enterprise in the Wild
- Anthropic and the Gates Foundation announced a $200M, four-year partnership combining grant funding, Claude credits, and engineering support across global health, life sciences, education, and economic mobility. First focus areas are polio, HPV, and preeclampsia: outbreak detection, vaccine candidate screening, and supply chain support for health ministries. Anthropic
- Isomorphic Labs closed a $2.1B Series B led by Thrive Capital, with Alphabet, GV, MGX, Temasek, and the UK Sovereign AI Fund participating. The accompanying technical report claims its Drug Design Engine (IsoDDE) reaches roughly 76% accuracy on antibody-antigen interfaces on FoldBench, against AlphaFold 3's 48%. PR Newswire
🛠️ Tooling & Ecosystem
- Salesforce shipped the Data 360 MCP Server (developer preview) May 15. Uses a facade-tool architecture to expose roughly 200 Data 360 API operations to coding agents without blowing past model context windows. Currently local-only and single-user; multitenant is not yet in scope. Salesforce Developers
- The MCP ecosystem now reports more than 14,000 servers, with governance under the Linux Foundation's AAIF.
⚖️ Policy & Regulation
- CAISI now has pre-deployment review agreements with all five US frontier labs. Google DeepMind, Microsoft, and xAI signed May 5, joining OpenAI and Anthropic, who have been in the program since 2024. Reviews cover cyber, biosecurity, and chemical risks. CAISI has completed more than 40 model evaluations to date. Practical effect: a longer gap between announcement and API availability for frontier releases. CIO
- EU AI Act amendments agreed May 7. Up to 16 months of additional time before high-risk-system rules apply, an SME compliance carve-out extended to firms with up to 750 employees and €150M revenue, and new prohibitions on AI-generated intimate content effective December 2, 2026. Council of the EU
📌 Watch List
- Agentic committee inference: verifier-backed weak-model loops as an alternative to scaling the base model.
- Pre-deployment review windows lengthening the gap between model announcement and API availability.
- Life sciences and healthcare as the next enterprise vertical after legal AI.
- The frontier talent diaspora: Sutskever's SSI, Igor Babuschkin's rumored $1B raise, Cohere's independent path.