🗞️ AI Daily Briefing — 2026-04-27

🔥 Top Story

Anthropic is now the most courted AI company on Earth. In the span of four days, Google committed up to $40 billion ($10B now at a $350B valuation, $30B conditional on performance targets) and Amazon announced up to $25 billion ($5B now at $380B valuation) — a combined $65 billion in commitments. Anthropic’s ARR crossed $30 billion in March, a 30x increase year-over-year. Google Cloud is delivering 5 GW of compute; Amazon locked in 5 GW on Trainium plus a $100B AWS spend commitment over 10 years. The two biggest cloud providers are in an overt bidding war for Anthropic’s future — the single largest capital formation event in AI history.

🚀 Model & Research News

  • OpenAI ships GPT-5.5 “Spud”: Released April 23, just six weeks after GPT-5.4. Better at coding (82.7% Terminal-Bench 2.0), agentic computer use, and deep research — with 2x pricing. Available to Plus/Pro/Business/Enterprise; no free tier. Agent mode is now a dropdown in ChatGPT. (OpenAI) (TechCrunch)
  • DeepSeek V4 Preview goes live — then full launch delayed: V4 Preview dropped April 24 with two variants: V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active). Open-source, Apache 2.0, 1M context, running on Huawei Ascend 950 chips. Bloomberg reported April 26 the full launch is postponed as DeepSeek prioritizes domestic chip integration. Simon Willison calls it “almost on the frontier, a fraction of the price” — $0.14/M input for Flash. (CNN) (MIT Tech Review)
  • Sergey Brin assembles DeepMind “strike team” to catch Anthropic on coding: Led by Sebastian Borgeaud (formerly Gemini pre-training lead) with direct CTO involvement. Internal DeepMind researchers reportedly rate Claude above Gemini for code. Brin’s leaked memo: “We must urgently bridge the gap in agentic execution and turn our models into primary developers.” (Sherwood News)
  • ARC-AGI-3 resets the scoreboard: François Chollet’s latest benchmark is devastating — humans score 100%, frontier AI scores 0.51%. Best model: Gemini 3.1 Pro at 0.37%. Claude Opus 4.6 at 0.25%. GPT-5.4 near zero. This is the first fully interactive, turn-based benchmark with no instructions or rules. A humbling reminder of how far we are from general intelligence. (ARC Prize) (The Rundown)
  • Moonshot AI releases Kimi K2.6: 1T-parameter open-source MoE (32B active, 384 experts), 256K context. Scores 58.6 on SWE-Bench Pro — beating GPT-5.4’s 57.7. Scales to 300 sub-agents and 4,000 coordinated steps. A serious Chinese open-source contender. (MarkTechPost)

🛠️ Tools & Developer Updates

  • LangGraph hits 1.1.8 with “Deep Agents”: Async subagents and type-safe streaming in v2. LangChain 1.0 and LangGraph 1.0 both reached GA this year with a commitment to no breaking changes until 2.0. (GitHub)
  • Anthropic launches “Project Deal” — agent-on-agent commerce: A classified marketplace where AI agents represented buyers and sellers, striking 186 real deals totaling $4,000+. Early experiment in autonomous economic agency. (TechCrunch)
  • Google Deep Research Max ships: Autonomous research agents built on Gemini 3.1 Pro, fusing open web and enterprise data via a single API call with MCP support for third-party sources. (Google Blog)
  • OpenAI Codex hits 4M+ weekly developers: Up from 3M two weeks earlier. Codex Labs launched alongside new GSI partnerships. (OpenAI)
  • Hugging Face contributes Safetensors to PyTorch Foundation: Standardizing the safe tensor serialization format across the ecosystem. Also added agent trace upload support from Claude Code and Codex. (Phoronix)
  • Mistral ships Small 4 + Voxtral TTS: Small 4 unifies fast instruction, deep reasoning, and multimodal chat. Voxtral TTS offers zero-shot voice cloning at $0.016/1k characters. Mistral ARR hit $400M in January. (Releasebot)

💰 Funding & Business

  • Anthropic raises ~$65B in commitments: Google ($40B) + Amazon ($25B) in the same week. Valuation $350-380B. ARR at $30B. The “startup” era is over. (TechCrunch) (CNBC)
  • Cohere merges with Aleph Alpha, raises $600M: Combined entity valued at ~$20B, backed by Schwarz Group’s €500M commitment. Positioning as the sovereign AI alternative for European enterprises. (TechCrunch)
  • Cognition (Devin) in talks at $25B valuation: More than doubling from $10.2B. ARR grew from $1M (Sept 2024) to $73M (June 2025). (Bloomberg)
  • DeepSeek seeks first outside funding at $20B+: Tencent wants up to 20%. Alibaba also circling. First external raise ever. (Bloomberg)
  • Meta cuts 10% of staff (~8,000 jobs) while spending $115-135B on AI: Cuts begin May 20. Simultaneously deploying “tens of millions” of AWS Graviton cores for agentic systems. The contradiction is the point. (CNN) (CNBC)
  • Snap cuts 1,000 jobs (16%), cites AI: AI generates 65%+ of Snap’s new code. Stock jumped 8%. Expected $500M+ in annualized savings. (TechCrunch)

🐦 Notable from the Timeline

  • Musk v. Altman trial kicks off in Oakland: Jury selection began this week, with $134-150B at stake over OpenAI’s nonprofit-to-profit conversion. Four-week trial. Witnesses may include Musk, Altman, Nadella, and current/former board members. This is being called “the AI trial of the century.” (Washington Post) (CNBC)
  • Sam Altman apologizes over ChatGPT/shooter failure: Altman said he was “deeply sorry” for not alerting law enforcement about a Canadian school shooter’s ChatGPT account, which had been banned 8 months before the February mass shooting that killed 8 people. (CBS News)
  • @fchollet drops ARC-AGI-3: Humans 100%, best AI 0.51%. Chollet and Altman held a fireside chat at YC HQ at the launch event. ARC Prize 2026 offers $2M in prizes. (ARC Prize)
  • Meta records employee keystrokes to train AI: Internal tool converts mouse movements and keystrokes into training data. Major privacy controversy inside the company. (TechCrunch)
  • NVIDIA open-sources DreamDojo: Jim Fan’s GEAR lab releases an open-source robot world model trained on 44,000 hours of human video. Fan calls it “Simulation 2.0.” (VentureBeat)

📊 Benchmark Watch

The frontier is historically tight and increasingly fragmented by task. Claude Opus 4.7 leads LM Arena at 1504 Elo and SWE-bench Verified at 82%. GPT-5.5 claims Terminal-Bench 2.0 (82.7%) and FrontierMath. Gemini 3.1 Pro leads GPQA Diamond (94.3%) and ARC-AGI-2 (77.1%). Kimi K2.6 beats GPT-5.4 on SWE-Bench Pro (58.6 vs 57.7). The gap between #1 and #10 on Arena is just 24 Elo points. Then ARC-AGI-3 lands and reminds everyone: on novel interactive tasks, all frontier models score under 1%. No single model dominates across all dimensions. (Arena.ai) (ARC Prize)

🎙️ Podcast Highlights

  • All-In E268 (April 17): Covered OpenAI’s identity crisis, datacenter buildout wars, and Anthropic’s competitive positioning. Sacks and Chamath debated whether OpenAI’s release pace signals strength or desperation. (Apple Podcasts)
  • Sequoia’s Training Data: Jim Fan appeared discussing “robots thinking fast and slow” — using world models for robot decision-making, and the DreamDojo open-source release. (Sequoia Capital)
  • TBPN: Continuing daily shows under OpenAI ownership. Recent episodes covered Anthropic’s run rate, Meta “token maxing,” and AI distillation. No visible editorial shift yet, but the structural conflict is the story.

🔗 Worth Reading