Anthropic raises $65B at $965B valuation while shipping Opus 4.8
Anthropic disclosed a $65B Series H led by Altimeter, Dragoneer, Greenoaks, and Sequoia at a $965B post-money mark (@anthropicai), with run-rate revenue crossing $47B earlier in the month (@anthropicai). Simon Willison flagged Axios's Jim VandeHei calling it the fastest organic revenue scaling in any era, though the same thread suggests the line gets harder to defend the closer you look (@simonw). Hacker News surfaced the funding round itself as the day's top story (last30days, ycombinator.com).
Alongside the raise, Claude Opus 4.8 went live at the same price as 4.7 — Boris Cherny cited a SWE-bench Pro jump from 64.3 to 69.2 and ~4x less likely to let code flaws pass unremarked (@bcherny, @openrouter). Reception split: Jeremy Howard found it noticeably more cooperative and willing to pause for input (@jeremyphoward), while Peter Steigrad's bench testing flagged regressions on Vending Bench and Blueprint-Bench versus 4.7 and GPT-5.5 (@steipete). Steigrad also called out Anthropic for putting GPT-5.5 directly into its comparison charts — a strategic shift from past Anthropic-vs-Anthropic framing (@steipete).
Claude Code dynamic workflows and the agentic-coding toolchain
Claude Code shipped dynamic workflows in research preview: Claude writes an orchestration script on the fly and spawns coordinated subagent fleets (@bcherny, @claudedevs). Cherny pitched it specifically for migrations, refactors, and batch bug fixes — the kind of "tie up engineers for a quarter" work teams normally defer (@bcherny). The launch post landed on Hacker News the same day (last30days, claude.com).
Around the release, infra and eval tooling moved in parallel. Peter Steigrad open-sourced octopool, a Cloudflare Worker pooling GitHub PATs behind a shared read cache (@steipete), and pushed OpenClaw updates showing 2.9x faster cold agent turns and a 59% smaller tarball (@steipete). Hugging Face released Repo2RLEnv to turn any repo into a verifiable coding environment for agent eval or RL (@huggingface). SWE-rebench's March–May refresh showed GPT-5.5 xhigh leading at 62.7% resolved, with Cursor Composer 2.5 noted as cheap-and-strong (@steipete).
RL infrastructure: weight sync, snapshots, and the multi-turn trap
vLLM shipped native weight-syncing APIs with NCCL and CUDA IPC implementations plus pause/resume fixes for async RL, in collaboration with Anyscale, NovaSky, and Red Hat (@vllm_project). NVIDIA's Dynamo Snapshot — built on vLLM — cuts inference cold-start from minutes to under 5 seconds via cuda-checkpoint + CRIU + memfd parallel restore (@vllm_project). Hugging Face's science team made async RL weight sync ~100x cheaper on bandwidth by exploiting the fact that ~99% of bf16 weights are bit-identical between RL steps (@clementdelangue).
Clément Delangue separately warned that most teams training agentic LLMs with RL have a silently broken training loop — re-tokenizing after tool-call parsing drifts assistant-prefix tokens — and Hugging Face followed up arguing the multi-turn "tito" problem needs just one implementation rule and one chat-template property (@clementdelangue, @huggingface).
Model releases sweep: multimodal, on-device, decentralized
StepFun's Step 3.7 Flash (198B sparse MoE, ~11B active, 256K context, Apache 2.0) went live with day-0 vLLM support, claiming #1 on ClawEval-1.1 and #2 on SWE-PRO (@openrouter, @vllm_project). Liquid released LFM2.5-8B-A1B for phones, laptops, and robots — 8B MoE with 1.5B active, trained on 38T tokens (@_akhaliq). Google made Nano Banana 2 (Flash) and Nano Banana Pro GA at $0.045 and $0.134 per image, with Flash now accepting video input (@googledeepmind, @_philschmid).
Pluralis's Paris 2.0 claims the first decentralized-trained video generation model, beating a same-budget monolithic baseline ~2x on FVD (@clementdelangue). Mistral shipped Mistral Vibe as its long-horizon coding agent with a VS Code extension (@mistralai), fal hosted FLUX Virtual Try-On (@fal), and Surya OCR 2, RF-DETR, and NVIDIA's LocateAnything also landed on Hugging Face (@huggingface).
AI economics: ugly ROI math meets runaway token bills
Gary Marcus surfaced FT/Panmure Liberum's implied 2025–2030 hyperscaler AI returns: only Amazon clears positive (+7.2%), with Meta at -28.8% and Oracle at -35.6%, and that's assuming zero costs (@garymarcus). He paired it with a 30-day rundown — Microsoft canceling most Claude Code licenses on cost, Uber burning its 2026 AI budget in four months, H200 rentals collapsing from $7 to $4/hr (@garymarcus). Simon Willison and Marcus both questioned the viral "$500M in a single month on Claude" anecdote on sourcing grounds (@simonw, @garymarcus, @emollick).
Counter-signals: MongoDB reported $688M Q1 revenue (+25% YoY), Atlas at +29.4%, VoyageAI customers more than doubling QoQ (@mongodb). Clément Delangue flagged that open-model use in Factory has 3x'd against closed models in a month (@clementdelangue), and Larry Ellison conceded models are commoditizing — proprietary data is the only moat left (@garymarcus).
Cybersecurity: critical RCEs and AI-in-the-enterprise risk
A heavy CVE day: pre-auth FortiClient EMS exploitation pushing credential stealers (CVE-2026-35616, CVSS 9.1), an unpatched Gogs RCE via malicious branch names during rebase-before-merge (no CVE issued) with a Rapid7 Metasploit module already public, and a SharePoint deserialization RCE (CVE-2026-45659, CVSS 8.8) requiring only Site Member rights (@thehackersnews). Kimsuky is hitting South Korean targets with HTTPSpy RAT via spoofed Webex pages, and two new Android NFC relay families — DevilNFC and NFCMultiPay — are draining bank cards in Europe and LatAm (@thehackersnews).
LayerX's 2026 report says the top 5% of employees generate 144+ AI conversations each, nearly half use personal identities, and 6%+ contain sensitive data — most orgs lack visibility (@thehackersnews). Microsoft publicly criticized researcher disclosures of multiple Defender/BitLocker zero-days now under active exploitation, with GitHub removing the researcher's account (@thehackersnews).
The Bottom Line
The day was bookended by Anthropic's $65B raise and Opus 4.8 launch, while the developer surface filled with dynamic workflows, cheaper RL weight-sync, and a Step 3.7 Flash / LFM2.5 / Nano Banana model wave. Underneath the launches, ROI skepticism and runaway enterprise token bills kept pace with the growth narrative — and security teams had a brutal CVE day alongside it.