AI weekly digest: Funding peaks, model launches, and the agent stack hardens

This week the AI economy hit a new high-water mark — Anthropic is eyeing a valuation north of $900B while SpaceX dangles a $60B buyout for Cursor — even as new model releases (GPT-5.5, Grok 4.3, DeepSeek-V4, Kimi K2.6) keep redrawing the price-per-intelligence frontier. Underneath the headlines, the agent stack quietly graduated to production: Claude Security shipped a public beta, Anthropic published a 4-layer security framework, and Google opened up its TPU supply chain. Here are the 10 stories that matter for builders.

1. Anthropic weighs a $50B raise at a $900B+ valuation

Anthropic is fielding offers for a roughly $50B round that would value the company between $850B and $900B, potentially overtaking OpenAI's $852B mark and making it the world's most valuable private AI company. Run-rate revenue has reportedly grown from ~$9B at the end of 2025 to over $30B in April, with some sources putting it nearer $40B. A board decision is expected in May, and an IPO as early as October is now in play.

Source: TechCrunch — Anthropic potential $900B+ valuation round could happen within two weeks

2. SpaceX's $60B option on Cursor reshapes the coding-agent map

SpaceX struck a deal that gives it the right to either pay Cursor $10B for joint work or acquire the company outright for $60B later this year, with Cursor's training pipelines plugging into xAI's Colossus cluster. Two of Cursor's senior engineering leaders have already moved to xAI to report directly to Musk. For builders this matters because the most operationally successful application-layer AI startup is now structurally aligned with a model lab — and Microsoft, who reportedly looked at buying Cursor first, is now without an in-house IDE play.

Source: TechCrunch — SpaceX is working with Cursor and has an option to buy the startup for $60B

3. OpenAI ships GPT-5.5 — faster, but hallucinations persist

GPT-5.5 lands with measurable wins on general benchmarks and roughly half the runtime of GPT-5.4 on SpatialBench, but third-party evaluators report accuracy is essentially unchanged on domain-specific tasks like spatial biology. OpenAI also disclosed that quirky "goblin"-style metaphors observed in GPT-5.1 traced back to reward signals from personality tuning — a useful reminder that small RLHF incentives bleed into surface behavior. If you serve user-facing inference, the upgrade is a clear cost win; if you depend on factual reliability, retest before swapping.

Source: DeepLearning.AI — The Batch: GPT-5.5 outperforms (and hallucinates)

4. xAI's Grok 4.3 takes the cost-per-intelligence crown

Grok 4.3 scores higher on the Intelligence Index than Grok 4.20 0309 v2 while costing less per benchmark run, making it one of the lowest-cost frontier-tier models currently available. It's particularly strong at instruction following and agentic customer-support tasks. For teams running high-volume agent workflows, it's worth a head-to-head against your current default.

Source: TLDR AI — xAI launches Grok 4.3

5. DeepSeek-V4 ships a million-token architecture

DeepSeek-V4 is built around an architecture optimized for million-token context, with attention and KV-cache tricks that change the economics of long-context serving. Early analysis from The Sequence calls it the most interesting open release of the past week, especially for retrieval-heavy and agentic workloads where context length, not raw IQ, is the constraint. If you're orchestrating long agent runs, this is a candidate to benchmark against Claude and Gemini for cost-per-1M-token.

Source: The Sequence — DeepSeek-V4 and the architecture of million-token intelligence

6. Kimi K2.6 leads the open-LLM leaderboard

Moonshot's Kimi K2.6 is now the top-ranked open model on multiple comparison boards, capping a strong week for Chinese labs that also saw Zhipu reportedly serving 5.5T tokens per day. Combined with DeepSeek-V4, Qwen-Scope, and GLM-5V-Turbo, the open-model frontier is concretely catching up to closed labs on capability, and pulling ahead on dollars-per-token. For anyone evaluating an open-weights swap from Llama or Mistral, this is a checkpoint week.

Source: DeepLearning.AI — The Batch: Kimi K2.6 leads open LLMs

7. Claude Security launches in public beta with Opus 4.7

Anthropic put Claude Security into public beta for Enterprise customers. Powered by Opus 4.7, it scans codebases for vulnerabilities and proposes patches, and ships with integrations into Microsoft Security and Palo Alto Networks rather than requiring a custom API build. The pitch is continuous, automated security review without bespoke wiring — a meaningful change for security teams who've been hand-rolling LLM-based scanners.

Source: Claude — Claude Security public beta

8. Google opens TPU sales — and Cloud crosses $20B/quarter

Alphabet plans to sell its custom TPUs to select customers for installation in their own data centers, in a clear signal that Google now sees its silicon as a standalone product line, not just a captive moat. The same week, Google Cloud reported quarterly revenue over $20B, up 63% YoY, with enterprise AI cited as the main driver. For infra teams locked to Nvidia, a real second source for high-end accelerators is finally on the horizon.

Source: TLDR AI — Google sells TPUs

9. Anthropic publishes a 4-layer agent security framework

Anthropic introduced a framework that splits agent security into four layers — Model, Harness, Tools, and Environment — and explicitly assigns ownership of three of those four layers to the deploying organization. The release coincides with multiple high-profile incidents this week, including agents deleting production data and a HERMES.md misconfiguration in Claude. For anyone running agents in production, the framework is the cleanest mental model yet for reasoning about where guardrails actually need to sit.

Source: TLDR AI — Supply chain is the new frontline / Anthropic agent security

10. Chinese courts rule AI replacement is not a lawful firing reason

In one of the first such decisions globally, Chinese courts ruled that replacing an employee with AI is not, by itself, a lawful basis for termination. Reported alongside Azeem Azhar's tour of Zhipu, MiniMax, Kimi, Alibaba, ByteDance and Xiaomi — where compute scarcity is real but innovation is not slowing — the ruling is a useful counterweight to the Silicon Valley assumption that displacement is automatic. Expect to see this case cited in Western policy debates over the next quarter.

Source: Exponential View #572 — AI's moats, myths and moral loopholes