AI Weekly - 2026-05-20

One-line summary

Opus 4.7 makes 1M context usable—not the “fits in the window” kind of 1M, the 96%-recall kind ^[1][2]. That opens a “stuff it all in” product path alongside RAG, with the biggest implications for document-heavy enterprise scenarios.

Models

Anthropic shipped Claude Opus 4.7, headlining a 1M context window with retained recall ^[1]. Cohere’s lab notes independently reproduced RULER 1M at 96.3% needle-in-haystack accuracy ^[2]—the first public model to cross the 95% mark, which has historically been the divide between “marketed context” and “actually usable context”.

The Information published screenshots of internal OpenAI o4 benchmarks; AIME-25 (math olympiad) is up 17 pp over o3 ^[3]. If verified at release, o4 reopens the frontier-reasoning gap. Caveat: internal benchmark ≠ shipping model, and the leak omits chain-of-thought budget.

DeepSeek-V4 briefly topped the LMSYS Chatbot Arena reasoning sub-leaderboard under the anonymous handle bouvardia-r2 ^[6]. The Chinese community has all but confirmed it’s the new DeepSeek release.

Regulation and compliance

Mistral passed the EU AI Office’s GPAI Tier-2 review, becoming the first European frontier lab to clear that compliance bar ^[4]. Tier-2 demands complete model cards, third-party red-team reports, and training-data provenance disclosure ^[5]. Later this year Tier-2 becomes mandatory, which will shape US labs’ EU launch cadence.

Compute

SemiAnalysis reports that NVIDIA Blackwell-2 volume shipments have slipped from Q2 to Q3, blamed on CoWoS-L yield ^[7]. The hyperscaler trio is rebalancing 2026 training-cluster racking; AWS and Azure are leaning into more aggressive H200 renewals.

Applications

Anthropic opened Computer Use beta to enterprise customers ^[8]—the first time a major lab has shipped “agentic screen control” as an SLA-backed enterprise product. Limited to Bedrock or direct-API accounts spending ≥ USD 5k/month.

Capex vs revenue

Sequoia updated their AI capex-revenue gap analysis: on an annualised basis, capex now exceeds attributable AI revenue by roughly USD 600B ^[9]. Their take is more optimistic than last year’s, though: “the gap is real but recoverable in 36–48 months if inference unit-cost continues compounding at -45%/yr”.

Ecosystem

Hugging Face published a mid-2026 commercial-use tally for open-weight licences ^[10]: clearly commercial-permissive includes Llama, Mistral, Qwen, DeepSeek; ambiguous (gated by an Acceptable Use Policy) includes Gemma and Falcon.