Key Highlights

  • Google Cloud Next ‘26 dominated the news cycle — Sundar Pichai unveiled eighth-generation TPUs (8i for inference, 8t for training), the Gemini Enterprise Agent Platform, Deep Research Max, and a $750M partner fund, positioning Google as the full-stack answer to Anthropic and OpenAI’s momentum.
  • Anthropic argues infrastructure noise is invalidating agentic benchmarks — resource-budget differences alone swing Terminal-Bench 2.0 scores by up to 6 percentage points, a provocative claim that calls much of the current leaderboard meta into question.
  • The Rundown reports Sergey Brin is personally running a DeepMind “strike team” to close Gemini’s internal coding gap with Claude — a rare signal that even Google’s founders think Anthropic has a meaningful lead in coding quality.
  • Meta will start recording employee keystrokes and mouse movements to train agent models — an early sign of how far big tech will reach for “real computer use” training data now that the public web is saturated.
  • Photonic computing moves from curiosity to credible contender (IEEE Spectrum) as energy constraints on GPU training force the industry to re-examine optical alternatives.

Analysis & Opinion

Sergey Brin commits DeepMind to a Claude catch-up — The Rundown

Internal DeepMind researchers reportedly rank Claude’s code-writing above Gemini’s, prompting Brin to assemble a dedicated team led by Sebastian Borgeaud to close the gap. The framing is notable: Brin reportedly sees coding parity as a prerequisite to self-improving AI, making this less a product war and more a bet on the fastest path to recursive capability gains.

Anthropic’s locked-down Mythos leaks — The Rundown

Anthropic’s cybersecurity-focused Mythos model — held back from public release for safety reasons — was reportedly accessed within days by a Discord group that pieced together deployment URLs from the Mercor breach and a contractor’s vendor credentials. The incident is a concrete case study in why “limited partner release” is not the same as “secured,” and undercuts a common containment assumption in AI safety plans.

How to Get Multiple Agents to Play Nice at Scale — Stack Overflow Blog

Intuit engineers describe abandoning a per-product multi-agent hierarchy in favor of a flatter “skills and tools” architecture with a central planner. Their lesson: customers don’t want to manage a zoo of agents — they want one interface that can answer cross-domain questions like “what if I raised salaries 5% and margins drop?” — and that demands a coordinator with access to every skill.

Mind the van Emden Gap — Fogus Blog

A sharp re-reading of M.H. van Emden’s 1982 vision of “Computer-Aided Thought” as an interlocutor that forces users to formalize fuzzy ideas. Today’s LLMs do the opposite: they absorb ambiguity, pick a plausible interpretation, and respond with unearned confidence — losing the “productive friction” that made the original idea valuable.

Meta Will Record Employees’ Keystrokes to Train Its AI Models — TechCrunch

Meta will capture mouse movements and keystrokes inside select internal applications to train computer-use agents. Coming on the heels of reports that defunct-startup Slack archives are being sold as training data, it signals the next training-data frontier is internal telemetry — and raises obvious consent questions even when the subjects are employees.

10 leading enterprises show why agents mean business — Google

Google’s customer roundup makes the case that agentic AI has crossed from sandbox to production line — engineering consoles, retail scanners, bank branches, power grids. Read alongside the 1,302-use-case tally, it’s a clear attempt to reframe the debate from “will agents work?” to “you’re already behind.”

New Products & Tools

Cloud Next ‘26: Momentum at Google scale — Google

Pichai disclosed Google’s first-party models now serve >16B tokens/minute via direct API (up from 10B last quarter) and that over half of 2026 ML compute investment will go to cloud customers — setting the stage for the rest of the Next announcements.

Google launches eighth-generation TPUs (8i and 8t) — Google / TechCrunch coverage

TPU 8i is optimized for low-latency agent inference; TPU 8t is built for training with a single massive memory pool. Google claims up to 3× faster training and 80% better perf/$, with clusters scaling to over 1M chips — though Google will still offer Nvidia Vera Rubin later this year.

Gemini Enterprise Agent Platform — Google

A unified developer platform for building, governing, and operating autonomous agents, bundling Gemini 3.1 Pro, Nano Banana 2, Lyria 3, and — notably — third-party Anthropic Claude Opus/Sonnet/Haiku. Positioned directly against AWS Bedrock AgentCore and Microsoft Foundry.

Deep Research and Deep Research Max — Google

Two Gemini 3.1 Pro-based research agents: one tuned for interactive speed, one for exhaustive background analysis. Both support MCP for proprietary data, native chart generation, and integrations with FactSet, S&P Global, and PitchBook.

Chrome “auto browse” for Workspace — TechCrunch

Gemini gains the ability to act across open Chrome tabs — entering CRM data, comparing vendor pricing, summarizing portfolios — with a required human-in-the-loop approval step before any final action. Rolling out to US Workspace users first, with reusable “Skills” shortcuts.

Gemini Embedding 2 GA — Google

Natively multimodal embedding model now production-ready on the Gemini API and Vertex AI, replacing fragmented per-modality pipelines.

Claude Code auto mode — Anthropic

A middle ground between manual approvals and full YOLO mode. Model-based classifiers screen commands at input (prompt-injection probe) and output (transcript classifier), aiming to block dangerous actions while letting safe ones proceed — motivated by the data point that users approve 93% of prompts anyway.

Managed Agents — Anthropic

Anthropic’s hosted agent service, framed around an OS-style insight: harnesses encode assumptions about current model weaknesses that rot as models improve (e.g., context-reset hacks that became unnecessary with Opus 4.5). The product bet is on durable abstractions over bespoke scaffolding.

ChatGPT Images 2.0 — OpenAI (via The Rundown)

OpenAI’s new image model adds planning, web search, and self-checking before generation, supports 2K resolution and 8 concurrent images, and has taken the #1 slot on Arena’s text-to-image leaderboard. Altman is calling it a “GPT-3 → GPT-5” leap for imagery.

Workspace agents in ChatGPT and ChatGPT for clinicians — OpenAI

Two vertical product moves: collaborative agents in Workspace, and clinician-tuned ChatGPT for medical workflows.

OpenAI + Infosys partnership — TechCrunch

Codex and other OpenAI tools get embedded into Infosys’ Topaz platform for legacy-modernization and DevOps engagements — mirroring Infosys’ earlier Anthropic deal and OpenAI’s HCLTech partnership.

NeoCognition launches with $40M seed — TechCrunch

Ohio State’s Yu Su emerges from stealth to build agents that specialize in new domains the way humans do, attacking the “today’s agents succeed ~50% of the time” reliability ceiling.

Thinking Machines × Google Cloud multi-billion deal — TechCrunch

Mira Murati’s lab expands onto Google Cloud’s GB300-based infrastructure — her first cloud-provider deal, alongside existing Nvidia ties. Part of Google’s push to lock in frontier labs as anchor tenants.

NVIDIA × Google Cloud: Vera Rubin A5X — NVIDIA

New A5X instances claim 10× lower inference cost per token and scale to nearly 1M Rubin GPUs. Notable: Gemini models running on Blackwell via Google Distributed Cloud with confidential computing, and NVIDIA Nemotron models showing up on Google’s agent platform.

Gemini 3.1 Flash TTS — Google DeepMind

Text-to-speech with audio tags — natural-language control over vocal style, pace, and delivery — hitting 1,211 Elo on Artificial Analysis’ TTS leaderboard.

Research

Quantifying infrastructure noise in agentic coding evals — Anthropic

Container resource settings alone cause up to 6-point swings on Terminal-Bench 2.0, and infrastructure error rates can hit 6% when CPU/memory specs are enforced as strict ceilings. The takeaway: many leaderboard differences fall inside the infrastructure noise floor — agentic benchmarks are not static tests, and the runtime is part of the evaluation.

Advancing Emerging Optimizers with NVIDIA Megatron — NVIDIA Developer

Muon (MomentUm Orthogonalized by Newton-Schulz) and related higher-order optimizers — already used for Kimi K2 and GLM-5 — now match AdamW throughput on GB300 NVL72 via layer-wise distributed optimization and three Newton-Schulz parallelization strategies.

The Future of Deep Learning Is Photonic — IEEE Spectrum

Argues optical matrix-multiply hardware could collapse the energy cost of neural-network inference, with photons replacing electrons for the multiply-and-accumulate operations that dominate deep learning.

Reversing SynthID — Hacker Factor

Security analysis of Google’s SynthID AI-content watermarking system — relevant for anyone relying on provenance signals as a defense against synthetic media.

References

  1. Anthropic, “Quantifying infrastructure noise in agentic coding evals,” Anthropic Engineering, 2026
  2. Anthropic, “Scaling Managed Agents: Decoupling the brain from the hands,” Anthropic Engineering, 2026
  3. Anthropic, “Claude Code auto mode: a safer way to skip permissions,” Anthropic Engineering, 2026-03-25
  4. OpenAI, “Making ChatGPT better for clinicians,” OpenAI, 2026-04-22
  5. OpenAI, “Introducing workspace agents in ChatGPT,” OpenAI, 2026-04-22
  6. Sundar Pichai, “Cloud Next ‘26: Momentum and innovation at Google scale,” Google, 2026-04-22
  7. Google, “We’re launching two specialized TPUs for the agentic era,” Google, 2026-04-22
  8. Google, “Gemini Enterprise Agent Platform lets you build, govern and optimize your agents,” Google, 2026-04-22
  9. Google, “Deep Research Max: a step change for autonomous research agents,” Google, 2026-04-21
  10. Google, “Gemini Embedding 2 is now generally available,” Google, 2026-04-22
  11. Google, “10 leading enterprises show why agents mean business,” Google, 2026-04-22
  12. Google, “1,302 real-world gen AI use cases from the world’s leading organizations,” Google, 2026-04-22
  13. Google DeepMind, “Gemini 3.1 Flash TTS: the next generation of expressive AI speech,” Google, 2026-04-15
  14. Google DeepMind, “Partnering with industry leaders to accelerate AI transformation,” Google DeepMind, 2026-04-21
  15. The Rundown, “Anthropic’s locked-down Mythos leaks,” The Rundown AI, 2026-04-23
  16. The Rundown, “OpenAI reclaims the image crown,” The Rundown AI, 2026-04-22
  17. The Rundown, “Sergey Brin commits DeepMind to a Claude catch-up,” The Rundown AI, 2026-04-21
  18. NVIDIA, “NVIDIA and Google Cloud Collaborate to Advance Agentic and Physical AI,” NVIDIA Blog, 2026-04-22
  19. NVIDIA, “Advancing Emerging Optimizers for Accelerated LLM Training with NVIDIA Megatron,” NVIDIA Developer, 2026-04-22
  20. Stack Overflow, “How to Get Multiple Agents to Play Nice at Scale,” Stack Overflow Blog, 2026-04-22
  21. Michael Fogus, “Mind the van Emden Gap,” Fogus Blog, 2026-04-21
  22. IEEE Spectrum, “The Future of Deep Learning Is Photonic,” IEEE Spectrum, 2026-04-22
  23. Hacker Factor, “Reversing SynthID,” Hacker Factor Blog, 2026-04-22
  24. Maxwell Zeff, “Google Cloud launches two new AI chips to compete with Nvidia,” TechCrunch, 2026-04-22
  25. TechCrunch, “Google makes an interesting choice with its new agent-building tool for enterprises,” TechCrunch, 2026-04-22
  26. TechCrunch, “Google turns Chrome into an AI co-worker for the workplace,” TechCrunch, 2026-04-22
  27. TechCrunch, “Meta will record employees’ keystrokes and use it to train its AI models,” TechCrunch, 2026-04-21
  28. TechCrunch, “AI research lab NeoCognition lands $40M seed to build agents that learn like humans,” TechCrunch, 2026-04-21
  29. TechCrunch, “OpenAI teams up with Infosys to bring AI tools to more businesses,” TechCrunch, 2026-04-22
  30. TechCrunch, “Exclusive: Google deepens Thinking Machines Lab ties with new multi-billion-dollar deal,” TechCrunch, 2026-04-22