Find the best AI tools — without the hype.

Each category links to our tested-and-ranked review. Use the question that matches your workflow. We re-test the recommended pick in every category once a week.

Ten current frontier models grouped by category. Star ratings reflect editorial assessment based on published benchmarks and our hands-on testing.

A short daily brief on what changed in AI yesterday — written by our editorial desk, not auto-aggregated. We pull from 15+ sources, pick the 5 stories that actually matter, and add our take. Each brief is fact-checked against the original source before it ships.

• 📅 June 1, 2026
  NVIDIA opens Cosmos 3 — what it calls the first open omni-model for physical AI — on Hugging Face, folding world generation, reasoning, and action into one Mixture-of-Transformers model shipped in 8B Nano and 32B Super sizes; at GTC Taipei NVIDIA unveils RTX Spark, a class of Windows PCs purpose-built for personal AI agents (up to 1 petaflop and 128GB unified memory), plus the OpenShell runtime that runs agents locally under user-defined policy; DuckDuckGo's "no-AI" search traffic roughly triples to a new high as it ships Chrome and Firefox extensions to set the mode as default — a measurable, durable opt-out market emerging against AI-saturated search; Strava overhauls its API agreement, replacing free tiered access with a flat paid subscription (reported at $11.99/month with a 90-day transition) and adding Model Context Protocol support to gate AI scraping ahead of a possible IPO; and a preview of what to watch as Microsoft heads into Build, where the Windows-as-AI-platform contest meets NVIDIA's on-device agent push
• 📅 May 31, 2026
  SoftBank commits up to €75 billion (about $87 billion) to build as much as 5 gigawatts of AI data-center capacity in France, with a first phase delivering 3.1GW to the Hauts-de-France region by 2031 — its largest AI infrastructure push in Europe and a marker that the buildout has entered its sovereign-infrastructure phase; two hands-on reviews of Google's Gemini Spark agent land on opposite verdicts, with TechCrunch calling it "actually pretty useful" for chores while Wired finds it misses what matters — the 2026 split between mechanical competence and real understanding; TechCrunch surfaces researchers warning that developers refusing to work without AI risk faster code that isn't better code, a maintainability debt that surfaces months later; the 2026 browser wars reopen with a wave of AI-first challengers reframing the browser as an agent that acts on the web for you; and The Verge issues a blunt retail-investor warning on the SpaceX IPO, flagging the cross-financing across Musk's AI-adjacent ventures
• 📅 May 30, 2026
  At I/O 2026, Google showcases Gemini 3.5 and a new Gemini Omni model line with a run of demo videos, alongside a companion "vibe coded" quiz built in Google AI Studio — the framing is capability-via-demo over spec sheet; Wired's hands-on with Gemini Spark finds a capable permissioned agent that did the mechanical party-planning work but missed the person most important to the reviewer — the 2026 reliability gap the demos don't show; Microsoft ships a speed-focused redesign of Microsoft 365 Copilot that it says loads twice as fast with more structured, scannable responses across desktop and mobile — a surface-and-latency update for the largest enterprise AI footprint; OpenAI publishes a shared playbook for trustworthy third-party evaluations of frontier models, landing the same week Illinois mandated third-party safety audits — the labs shaping the eval standard before regulators impose one; and Cognition's Scott Wu, who builds Devin, tells TechCrunch that coding agents aren't designed to replace human programmers — augmentation framing from the maker of the category's most aggressive product
• 📅 May 29, 2026
  Anthropic closes a $65 billion Series H at a $965 billion post-money valuation — likely its last private round before an IPO — and ships Claude Opus 4.8 the same day, with chip partners Samsung, SK Hynix, and Micron joining alongside Altimeter, Sequoia, and other co-leads; Illinois passes SB 315, the first US law to mandate third-party safety audits of frontier models, with a $500M-revenue threshold, 72-hour incident reporting, and a January 2028 effective date — and the major labs signaling support; OpenAI launches Rosalind Biodefense, opening its GPT-Rosalind model to vetted developers and US government partners under a gated-access-plus-partnership model; Asana acquires no-code agent builder Stack AI to push its "operating system for human-agent teams" pitch, betting integration context beats model quality; and Apple is reported to be distilling Google's Gemini into on-device models for a rebuilt, privacy-framed Siri expected later in 2026
• 📅 May 28, 2026
  OpenAI publishes a customer story on Warp using GPT-5.5 and other OpenAI models to coordinate coding agents across local, cloud, and open-source workflows — the cleanest statement yet that OpenAI sees the multi-agent orchestration terminal as a category, not a one-off product; IBM Research and Artificial Analysis ship ITBench-AA, the first benchmark dedicated to agentic enterprise IT work — frontier models score below 50%, the conservative planning baseline for autonomous-agent throughput on real IT operations this year; Wired covers Trajectory, a startup from former Google and Apple researchers betting that the vibe-coding iteration cycle can be packaged as a continuous-learning loop for every AI product — the second customer-facing example this week of "self-improving in production" moving from research vocabulary to a buyable thing; YouTube starts rolling out an AI-prompted custom video feed pinnable to the home screen — the first major consumer feed platform to put a prompt bar on home as a first-class navigation primitive, with real second-order effects for niche creators and YouTube's ad targeting unit; and Vertu launches a $6,880 AI foldable for CEOs built on the open-source Hermes project — the cleanest data point yet that AI hardware is going to fragment by buyer segment (executive, developer, consumer) before it converges
• 📅 May 27, 2026
  NVIDIA publishes the first public Phoronix benchmark results for its Vera CPU built for the agentic-AI factory — the host-CPU spec line in an AI-rack quote is now a variable, not a constant; a critical "BadHost" vulnerability in Starlette puts 325 million weekly downloads of agent-runtime infrastructure at risk, with same-day patch action on every team running an agent service on FastAPI; OpenAI publishes a joint case study with Thrive and Crete on a self-improving tax-filing agent built with Codex — the first named-customer template of an agent that generates its own evaluations and trains on them in production; MIT Technology Review quantifies the agentic-AI readiness gap at 85% of organizations want to be agentic in three years but 76% say their operations and infrastructure can't support it — the next year of enterprise adoption priced on org design, not model release notes; and The Verge keeps the Uber AI-spend story on the front page after president Andrew Macdonald told Rapid Response the spend is "harder to justify"
• 📅 May 26, 2026
  The Sober Turn — MIT Technology Review's two-part counter-narrative argues that the white-collar collapse story doesn't match the aggregate labor data but the first rung of the career ladder is quietly breaking; TechCrunch reports nine-year-old ClickUp is laying off hundreds and replacing them with "thousands of AI agents" in the cleanest "replace, not augment" SaaS case study to date; The Verge quotes Uber president Andrew Macdonald telling Rapid Response that the company's AI spend is "harder to justify" after exhausting its 2026 budget four months in — the first Fortune 100 operator to put the ROI question on the record this directly; Wired publishes the definitive year-one retrospective on how Claude Code and OpenClaw kicked off the agent era for engineering and reframes the rest of the brief as one continuous chain of consequence; and a new arXiv preprint, "Chain-of-Thought Hijacking," finds that over-extended reasoning in large reasoning models weakens safety behavior rather than strengthening it — the working assumption of two years of safety research, now empirically inverted

View full daily-brief archive →