Two Labs Clear the 32-Step Range, Anthropic Locks the Compute Stack, DeepSeek Goes 1M, AlphaEvolve Lands in Production: Monday Briefing, May 11, 2026

The biggest capability datapoint of the weekend is that OpenAI’s GPT-5.5reportedly cleared the same 32-step end-to-end cyber-attack range that Anthropic’s Claude Mythos preview cleared three weeks earlier. Two frontier models from two different labs now sit above a long-horizon agentic-offense bar that was a research curiosity in February. The 32-step characterization is meaningful because it spans reconnaissance, initial access, privilege escalation, lateral movement, and objective-action chains end-to-end — not isolated capability probes. Anthropic’s previous result already underpinned its Project Glasswing selective-preview model and the recursive-alignment thesis (Claude proposing, running, and helping design next-gen alignment experiments); GPT-5.5’s result removes any remaining argument that the Mythos number was a one-lab anomaly.

The downstream implication for defenders is non-optional: if two frontier models can chain a full 32-step intrusion, the gap between offense-capable AI and defender-deployed AI is the window every CISO is now budgeting against. That intuition is corroborated by ground-level telemetry — The Hacker News and corroborating reports put malicious packages on public repos up ~75% and cloud intrusions up ~35% in 2025, and AI-generated phishing has reportedly outperformed human red teams on standardized templates. Combined with last week’s Mandiant 2026 M-Trends figure that 28.3% of CVEs are exploited inside 24 hours of disclosure, the time-to-exploit curve has crossed below the time-to-patch curve in median cases.

The headline capital story is structural. Amazon put another $5B into Anthropic, taking its total to $13B; Anthropic in turn committed to >$100B in AWS spend over the next decade and secured up to 5 GW of new compute capacity. In parallel, Google added another $40B on top of its previous stake, and a chip-supply agreement with Google and Broadcom is reportedly worth hundreds of billions of dollars in committed silicon over the contract life. A joint enterprise AI services venture with Blackstone, Hellman & Friedman, and Goldman Sachs was announced alongside, putting a financial-sponsor distribution layer underneath Anthropic’s enterprise push.

The strategic reading is that frontier-lab compute supply is no longer demand-constrained on paper for the labs that have already cleared the capital bar — it is contract-locked for the rest of the cycle. Google’s TPU 8t “Superpod”scaling to 9,600 chips, 2 PB of shared HBM, and 121 ExaFlops is the physical embodiment of that posture: a credible #2 training-chip ecosystem behind NVIDIA, with Anthropic as the anchor tenant. The April–May arc — OpenAI multi-cloud, Anthropic on Bedrock and Vertex, Google’s $40B and Broadcom silicon, Amazon’s top-up — closes out the “single-cloud frontier lab” era. Every major frontier model is now reachable on at least two hyperscalers, and every major hyperscaler is committed to at least two frontier labs.

DeepSeek V4 released as an open-weight Mixture-of-Experts model with a >1M-token context window — roughly a 10× context expansion over V3 — and strong coding performance against the Western frontier on agentic-coding benchmarks. It joins a tight cluster of recent open-weight launches (DeepSeek V4, GLM-5.1, MiniMax M2.7, Kimi K2.6, Google Gemma 4) that have collectively closed the open-source capability gap on coding and tool-use while keeping inference unit-cost meaningfully below closed-weight list pricing. The 1M-token mark specifically pulls long-context workloads — codebase-wide agent runs, multi-day research traces, full-document review — into a self-hostable budget for the first time from an open-weight stack.

The competitive arithmetic for closed-weight labs is now visible: every long-context premium in their pricing has an open-weight floor below it, and that floor is a contract month or two of fine-tuning away from production parity on the most common enterprise tasks. Combined with Cloudflare’s newly-shipped TurboQuant + Unweight serving path (covered in yesterday’s digest), the long-context inference economics for the rest of Q2 look materially cheaper than what was baked into March budgets.

Google DeepMind’s AlphaEvolve — a system that pairs Gemini with an evolutionary algorithm to discover new algorithms for unsolved problems — is already deployed inside Google to optimize data-center power management and TPU scheduling. This is the cleanest AI-for-science datapoint of the year so far: not a paper result on a closed benchmark, but a self-improving system whose outputs are running in production against the cost line of Google’s most expensive infrastructure. The combination with the TPU 8t Superpod buildout is notable — the same organization that ships the second-largest training-chip ecosystem on earth is also using its frontier model to optimize how that ecosystem is scheduled.

Adjacent research datapoints round out the picture. Google TurboQuant(ICLR 2026) is the KV-cache compression algorithm that just hit Cloudflare’s production serving path; UPenn’s Mollifier Layers stabilize inverse-PDE inference for scientific ML; UH Mānoa’s physics-informed ML algorithmimproves accuracy in fluid dynamics and climate modeling; and a Science study reports that an OpenAI reasoning model outperformed experienced physicians on diagnosis and care management using EHR data from a Boston emergency department. The shape of the cycle is visible: AI systems that improve their own training, serving, and scientific frontier in parallel, on the same calendar.

Four threads to track this week. First, whether Google DeepMind, Microsoft, or xAI publish their CAISI pre-release eval cards in the next ten business days — OpenAI set the baseline on Sunday and the second filing will set the ceiling on disclosure detail. Second, follow-on long-context serving paths from Bedrock, Vertex, and Azure AI Foundry on TurboQuant or equivalent KV-cache compression — expect at least one before the end of May. Third, the Colorado AI Act repeal floor vote— pass or fail, the result reshapes how every other state legislature drafts its 2026 AI bill. Fourth, the next vertical-agent scoreboard: IBM watsonx Orchestrate (post-Think 2026) and Microsoft Agent 365 (now GA) are the most-likely next subjects, and the per-task rubric Anthropic’s financial-services agents were graded on yesterday is a credible template for both.