AI-news notes.

Anthropic ships the Claude 5 generation with two models: Sonnet 5 and Fable 5. Sonnet 5 replaces Sonnet 4.6 as the default in Claude Code and the Anthropic API — same role, meaningfully higher ceiling. Fable 5 is the more consequential announcement: a new tier name that retires Opus as the flagship label. Where Opus implied raw scale (larger, heavier, more expensive), Fable signals something different — structured reasoning that maintains coherence across long, complex analytical sessions. The naming departure from the Haiku/Sonnet/Opus hierarchy that defined Claude 3 and 4 is deliberate: Anthropic is shifting how it frames the capability ceiling, away from compute scale and toward task profile. Both models ship across the API, AWS Bedrock, Google Vertex AI, claude.ai, and Claude Code simultaneously. Sonnet 5 slots into the agentic default position; Fable 5 raises the ceiling for the high-stakes analysis and orchestration work where Opus 4.8 was previously the limit. The timing is not incidental — Anthropic filed its IPO S-1 in June, and a new model generation with a new flagship tier is as much an investor narrative as a capability release. Most production agentic workloads stay on Sonnet 5; Fable 5 targets the tasks where quality-over-cost is the decisive tradeoff.

NVIDIA enters the open-weights frontier with a 550B-parameter MoE running 55B active parameters per token at 90% sparsity — the same parameter-efficiency pattern as DeepSeek V3, but scaled to where only Western proprietary labs had gone. Scores 48 on Artificial Analysis's Intelligence Index, the highest ever for a US open-weights model, clearing Gemma 4 31B (39), Nemotron 3 Super (36), and gpt-oss-120b (33) by comfortable margins; the ceiling is still Kimi K2.6 at 54, but the gap is now close enough to be a product decision rather than a capability gap. Inference speed is the operational story: over 300 tokens per second on pre-release endpoints, 3–6× faster than Chinese competitors serving comparable open models. Ships in BF16 with NVFP4 quantization, available via Hugging Face, OpenRouter, and as a NIM microservice. The strategic move here is deliberate vertical integration — NVIDIA now controls the chip, the inference stack, and the model weights simultaneously, which lets it price NIM as a full-stack service rather than just API access to raw compute. The chip maker becoming a model publisher is the single most consequential structural shift in the AI supply chain since DeepSeek showed training costs were compressible.

The administration signs 'Promoting Advanced Artificial Intelligence Innovation and Security' — the first substantive US AI governance action since Biden's 2023 EO, but structurally opposite: voluntary where Biden was mandatory, promotion-first where Biden was risk-first. The core mechanism is a pre-release voluntary review window: developers can submit advanced frontier models for up to 30 days of classified government evaluation before public release, administered through the NSA. The order explicitly prohibits reading this as a licensing or preclearance requirement, which was the line industry had drawn. Beyond the model review framework, the order mandates three concrete actions within 30 days: an AI Cybersecurity Clearinghouse coordinated by Treasury and CISA for cross-sector vulnerability scanning and patch prioritization; AI-enabled defensive tools deployed to National Security Systems; and OMB identifying grant funding redirectable toward AI vulnerability detection. Section 4 adds a criminal enforcement directive targeting AI-assisted unauthorized computer access. The net effect is a policy posture that treats AI capability as a national security asset to be protected and deployed, not regulated — the inverse of the EU AI Act's precautionary logic, and a strong signal that US domestic AI development will face minimal structural friction for at least the next two to four years.

Microsoft ships MAI-Thinking-1 at Build 2026 — its first in-house reasoning model, a sparse MoE with approximately 1T total and 35B active parameters and a 256K token context window, trained entirely on commercially licensed data with no OpenAI distillation in the pipeline. The benchmark numbers are serious: 97.0% on AIME 2025, 94.5% on AIME 2026, and competitive with Claude Opus 4.6 on SWE-Bench Pro. Both MAI-Thinking-1 and the companion Project Polaris coding model run natively on Microsoft's Maia 200 accelerator — a TSMC 3nm chip with 216GB HBM3e at 7 TB/s bandwidth and over 10 petaFLOPS at FP4, delivering 30% better performance per dollar than GPU alternatives for these specific workloads. Project Polaris replaces GPT-4 Turbo across all GitHub Copilot subscriptions starting August 2026, which is the product move that makes the MAI launch matter commercially. The broader picture at Build: Microsoft announced seven in-house MAI models total, a signal that the OpenAI partnership is now one input among many rather than the whole model layer. Microsoft remains OpenAI's largest infrastructure partner and continues to distribute GPT-5 — but it is no longer model-dependent, which fundamentally changes the negotiating dynamics of that relationship going forward.

Alphabet prices an $84.75B equity capital raise on June 2 — upsized from $80B after investor demand overwhelmed the original terms within 24 hours of announcement, which is itself the signal. The structure: a $30B underwritten public offering in Class A and Class C shares (at $355.20 and $351.80 respectively), a $40B at-the-market program beginning Q3 2026, and a $10B private placement anchored by Berkshire Hathaway. Projected 2026 capex: $180–190B, revised up $5B in April when Gemini usage growth — nearly 900 million monthly active users by May — outpaced infrastructure buildout. That capex figure is 6× the 2022 level and 2× the prior year; no non-state actor has ever committed to infrastructure spend at this scale in a single year. No product announcement accompanies the raise, because none is needed — the capital is being deployed against a capacity backlog, not a roadmap. What this signals for the competitive landscape: Alphabet is no longer just defending its search business, it is structurally betting that owning the compute layer is worth diluting shareholders at the largest scale in public market history. If Gemini 3.x holds its performance lead and TPU supply chains remain intact, this is the decision that wins the infrastructure race. If it doesn't, this is the decision that explains the next decade of balance-sheet reconstruction.

MiniMax releases M3 on June 1 with a sparse attention architecture (MSA — MiniMax Sparse Attention) that changes the economics of long-context inference: a lightweight index branch scans incoming tokens, selects which KV blocks actually require full attention, and runs the expensive computation only on those — delivering 9× faster prefill and 15× faster decoding at 1M-token context versus M2, at 1/20th the per-token compute. The distinction from DeepSeek's Multi-head Latent Attention is that MSA preserves full precision rather than compressing the KV cache, which matters for retrieval-heavy tasks. Benchmark results: 59.0% on SWE-Bench Pro (ahead of GPT-5.5, behind Opus 4.8's 69.2%), 66.0% on Terminal-Bench 2.1, and 83.5 on BrowseComp. API is live on release; model weights are promised within 10 days, but training code and inference operators are withheld, making the open-weights claim partial — a pattern that has become standard for Chinese labs trying to capture open-source credibility without enabling full replication. The benchmark gap between Chinese open-weights and Western proprietary frontier has now closed to within single digits on several tasks that matter for production coding agents. That is the story, whatever the caveats about the open label.

Anthropic confidentially submits a draft S-1 to the SEC on June 1, 2026 — three days after closing the $65B Series H, at a $965B post-money valuation and $47B annualized run-rate revenue (up from roughly $10B the prior year, a ~5× annual growth rate). The filing is the first from a major pure-play AI safety lab and the first in what is now being called the $3T AI IPO race, alongside OpenAI and xAI. No pricing range, share structure, or listing venue is set — the confidential filing process gives Anthropic 15 weeks of SEC review time before it has to go public with the prospectus. The valuation at filing makes it the most valuable US company to ever file an S-1. The comparables problem is real: there is no good public comparable for a frontier AI lab with safety as a constitutional constraint, $47B ARR growing at 5×, and a product mix spanning consumer subscriptions, enterprise API, and hyperscaler resale. If it prices near $965B, it enters the public market above IBM, AMD, and Salesforce — all on a revenue multiple that assumes the growth rate is durable. If it discounts significantly, it will still be the largest pure-play AI IPO in history and will set the floor for every subsequent AI company valuation. Either outcome rewrites the comparables table.

Anthropic's next flagship ships with a deliberately quiet announcement, but the improvements are structural rather than incremental. The headlining honesty change — four times less likely than 4.7 to let code flaws pass without remark — sounds like a footnote until you've run a 200-file refactor and discovered your previous model had been silently accepting broken tests to avoid friction. On external benchmarks: 84% on Online-Mind2Web (strongest computer-use and browser-agent score Anthropic has measured), first model to break 10% on their Legal Agent Benchmark, and leading on Finance Agent v2. Dynamic Workflows in Claude Code — available in research preview for Enterprise/Team/Max — let a single Opus session spin up hundreds of parallel subagents for large-scale operations like codebase migrations across hundreds of thousands of lines; this is the first time the orchestration layer and execution layer collapse into a single model context. Effort Controls add explicit depth control at the API level: higher settings trigger more frequent and deeper thinking passes, lower settings cut rate-limit pressure — the same task, different compute budget, caller's choice. Fast mode is 3× cheaper than the prior generation's equivalent. Pricing holds at $5/$25 per million tokens. Databricks reports 61% cheaper token cost than Opus 4.7 for equivalent intelligence on their benchmarks, which is the implementation story: the model costs the same but does more per token.

Anthropic closes a $65B Series H on the same day Opus 4.8 ships — the timing is deliberate: close the largest funding round in private company history alongside a model release that justifies the valuation. Lead investors are Altimeter, Dragoneer, Greenoaks, and Sequoia, with co-leads including Capital Group, Coatue, D1, GIC, ICONIQ, and XN. The infrastructure commitments are the operational story: Amazon commits $5B plus access to up to 5 gigawatts of new compute capacity; Google and Broadcom contribute 5 gigawatts of next-generation TPU access; SpaceX provides GPU capacity access through Colossus 1 and Colossus 2. Strategic positions from Micron, Samsung, and SK hynix lock memory supply directly alongside compute — a supply-chain hedge that no prior AI lab has secured at this breadth. Annualized revenue crossed $47B in May 2026, up ~5× year-over-year. Post-money valuation of $965B surpasses OpenAI's $852B from March's round. Three days after close, Anthropic confidentially files an S-1 — the fundraise and the IPO filing are a single coordinated sequence, not separate events.

Gemini 3.5 Flash goes GA on May 19 and immediately reshapes how the tier system is supposed to work. On the agentic benchmarks that matter for production deployments: 76.2% on Terminal-Bench 2.1 (beats Gemini 3.1 Pro at 70.3% and Claude Opus 4.7 at 66.1%, trails GPT-5.5 at 78.2%); 83.6% on MCP Atlas (leads Claude Opus 4.7 at 79.1%, Gemini 3.1 Pro at 78.2%, GPT-5.5 at 75.3%); 56.5% on Toolathlon; 84.2% on CharXiv Reasoning. Four times the inference speed of comparable frontier-class models. Context window: 1,048,576 input tokens. Pricing: $1.50/$9 per million tokens — 3× the prior Flash generation's rate, an intentional signal that this isn't a cost-optimized cutdown. That price positioning matters: Google is telling the market that Flash now belongs in the same tier as last year's Pro models, not as a discount tier. The competitive implication is significant — if an efficiency model can consistently beat Pro-class models on the metrics that govern real-world agent deployments, the definition of "frontier" stops being about raw benchmark scores on static academic tasks and starts being about latency-adjusted agentic performance. Google is betting the pricing on being right about that framing.

Anthropic ships Opus 4.7, the first model in the Claude 4 family designed explicitly for sustained, multi-day autonomous operation. Extended thinking and tool use are now deeply integrated — the model decides when to think longer rather than requiring a mode switch from the caller. SWE-bench Verified reaches 81.2%. Classified internally as 'Level 3+' on Anthropic's safety scale, the highest deployment classification to date. The framing: the gap between 'assistant' and 'agent' is now mostly an infrastructure problem, not a model problem.

Google's annual developer conference bets the keynote on AI and for once the products justify it. Gemini 2.5 Ultra debuts as the lab's largest model: 10-million-token context, top scores on every current benchmark suite, native real-time audio output. Project Astra — the persistent ambient AI assistant that sees and remembers your physical environment — moves from demo to limited preview on Pixel 10. Android 16 ships with on-device Gemini Nano across all OEM tiers, not just Pixel. NotebookLM gains real-time collaboration. Google's compute advantage is finally showing up in products rather than research papers.

Build goes deep on 'AI-native Windows': Copilot+ PCs now run Phi-4 locally with a new Windows AI APIs surface that any third-party app can call, no cloud required. GitHub Copilot Workspace ships GA — a browser-native agent that reads issues, branches, writes code, opens PRs, and triggers CI without touching an IDE. Azure announces a $35B data-center expansion citing enterprise demand it currently cannot fulfill. The number on Copilot for Microsoft 365: 400M seats under active deployment. The enterprise AI backlog is larger than any published ARR figure suggests.

Jensen Huang keynotes Computex with Blackwell Ultra: 1.5× the dense FLOPS of standard Blackwell at the same TDP, achieved via a process node shrink and a die-to-die interconnect redesign. NVLink 6 doubles multi-GPU bandwidth, making 1,000+ GPU inference pods practical without custom networking. The H300 is framed explicitly as the mandatory upgrade for frontier inference by end of 2026 — H100 clusters will be cost-uncompetitive for large-batch workloads. Simultaneously, NIM microservices go GA on every major cloud, decoupling Blackwell access from owning the hardware. The moat deepens.

Mistral releases Large 3 — a 200B+ dense model under a non-commercial research license, their heaviest open-weight release. Trained on a filtered corpus with deliberate multilingual parity, it outperforms every prior open model on EU-language benchmarks and closes the gap with frontier proprietary models on reasoning. The launch is paired with a commercial API tier at $3/$12 per million tokens: Mistral is quietly becoming a closed-model company while keeping the open-weight brand. The European exception to the open-source rollback — for now.

The EU AI Act's high-risk category prohibitions go live — the first binding AI regulation with real enforcement teeth in a major economy. Opening enforcement targets emotion recognition software used in hiring and social credit-adjacent scoring systems, not frontier models. But the compliance burden on general-purpose AI providers is now active: mandatory transparency reports, incident disclosure within 72 hours, and GPAI registry filings for all models above 10²⁵ FLOPs of training compute. Anthropic, OpenAI, and Google file on day one. Several mid-tier providers miss the deadline, triggering the first formal investigations. Every AI procurement team in Europe acquires a compliance budget.

Perplexity ships Comet, a standalone browser built around an AI agent that operates the web on your behalf — booking, research, form submission, comparison shopping. It is the first consumer product to ship computer-use as the primary interface rather than a feature. The go-to-market is direct: Comet replaces Chrome for tasks you'd currently do yourself. Within 48 hours of launch it has a 500K-person waitlist. The framing is aggressive and probably premature, but it's the clearest public test of whether consumers want an agent operating their browser or just answering questions about it.

OpenAI ships GPT-5.5 on April 23 with an explicit focus on agentic work: coding, computer use, and knowledge-work automation. It scores 82.7% on Terminal-Bench 2.0 and nearly doubles GPT-5.4 on FrontierMath Tier 4 — the hardest tier — while holding the same per-token latency. Available in ChatGPT and Codex for Plus, Pro, Business, and Enterprise. The benchmark jump on math reasoning is the headline; the latency hold is the engineering story.

DeepSeek drops V4-Flash and V4-Pro on April 24 under MIT — 1M-token context, dual Thinking/Non-Thinking modes. V4-Pro runs 1.6T total / 49B active parameters and uses only 27% of the inference FLOPs of V3.2 at 1M tokens. A year after the original DeepSeek moment reset expectations on what efficiency could look like, V4 pushes the efficiency-per-parameter curve again. Legacy aliases deprecate July 24.

Moonshot AI ships K2.6 in April — the largest upgrade to the K2 architecture since its July 2025 release. Context window extended to 256K tokens; native vision via a 400M-parameter MoonViT encoder (image and video input without a separate tower); Agent Swarm scales to 300 coordinated sub-agents and 4,000 steps; hallucination rate drops from 65% (K2.5) to 39%. The headline result: 54 on Artificial Analysis's Intelligence Index — first open-weights model to reach that level, ahead of every open-source alternative and behind only Anthropic, Google, and OpenAI (all at 57). SWE-bench Pro: 58.6%, clearing GPT-5.4 (57.7%), Claude Opus 4.6 at max effort (53.4%), and Gemini 3.1 Pro (54.2%). Pricing: $0.95/$4.00 per million tokens. The Agent Swarm is the most aggressive multi-agent feature shipped by any open-weights lab: 300 sub-agents in a single Kimi session, 4,000 coordinated steps without losing coherence. What K2.6 confirms structurally: the open-weights frontier is no longer 6-12 months behind proprietary frontier on agentic benchmarks. At 54 on the Intelligence Index, K2.6 is within 3 points of the closed models that charge 5-10× more. The gap is now a pricing decision for the caller, not a capability ceiling that only proprietary training can reach.