AI Weekly Recap: Anthropic's $65B Raise & a cost Collapse

This AI industry weekly recap covers May 25-31, 2026 — one of the densest seven-day stretches of the year. Five stories defined the week, and Anthropic was behind two of the biggest: a new flagship model and a funding round that pushed its valuation toward $1 trillion on the same day.

What ties the week together is a single theme: capability is no longer the scarce resource — cost and access are. Two new open-weight models claimed frontier-adjacent coding scores at a fraction of closed-model pricing, NVIDIA reframed compute as a manufactured commodity, and Anthropic's raise priced the lab higher than OpenAI was at its last fundraise. For builders and marketers, the practical question shifts from "which model is smartest" to "which model is cheap enough to run the workload I care about."

Below we walk each story in order, hedge every vendor-stated benchmark honestly, and assemble two proprietary tables you won't find in other recaps: a cost-per-task comparison across the week's models and Anthropic's full valuation arc. If you missed the prior installment, start with last week's AI recap.

On May 28, Anthropic released Claude Opus 4.8 with pricing unchanged from Opus 4.7 — $5 per million input tokens and $25 per million output tokens. Holding price flat on a flagship upgrade is itself the story: the competitive pressure from OpenAI Codex and Gemini Flash has made frozen list rates the new normal. The release also arrived just 41 days after Opus 4.7, an unusually compressed upgrade cycle.

The headline feature is Dynamic Workflows, launched in research preview. It lets Claude Code orchestrate hundreds of parallel subagents to complete codebase-scale migrations spanning hundreds of thousands of lines of code. It's on by default for Max, Team, and API plans, but off by default for Enterprise accounts, where an admin must enable it. Alongside it, Effort Control shipped to claude.ai and Cowork across all plans — a selector beside the model picker that trades reasoning depth against speed and rate-limit draw.

The same day Opus 4.8 shipped, Anthropic announced it had raised $65 billion in Series H funding at a $965 billion post-money valuation. That figure puts Anthropic above the $852 billion valuation OpenAI carried at its last fundraise in March 2026 — making Anthropic, at the moment of this round, the most highly valued AI startup. TechCrunch, CNBC, and others independently confirmed the round.

The round was co-led by Altimeter Capital, Dragoneer, Greenoaks, Sequoia Capital, Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN, with additional investors including Blackstone, Brookfield, Fidelity, Baillie Gifford, DST Global, and Temasek. It also folds in $15 billion of previously committed hyperscaler investment — $5 billion of it from Amazon — and strategic infrastructure participation from Micron, Samsung, and SK hynix.

The financial backdrop is what makes the valuation legible. Anthropic's run-rate revenue crossed approximately $47 billion in May 2026, up from roughly $10 billion a year earlier, with the company projecting 130% revenue growth and its first operating profit. A valuation that nearly tripled — from $380 billion in February to $965 billion — in about three months is unusual even by AI-cycle standards, and it is the single most consequential business datapoint of the week.

Two open-weight releases landed in the back half of the week. On May 29, StepFun released Step 3.7 Flash— a 198B sparse Mixture-of-Experts vision-language model (a 196B language backbone plus a 1.8B ViT encoder) that activates roughly 11B parameters per token, shipped under an Apache 2.0 license. It supports a 256K-token context window, three selectable reasoning levels, and operates across ten languages. StepFun reports throughput up to 400 tokens per second, a vendor-stated figure. It succeeds the lab's earlier Step 3.5 Flash, which is no longer the current model.

On May 31, MiniMax brought M3to its API. Be precise about status here: the API is live, but the open-weight release and technical report were planned within roughly ten days — the weights were not yet out as of this post's date. M3 is a 229.9B-parameter MoE with 9.8B active per token across 256 fine-grained experts, built on a new MiniMax Sparse Attention (MSA) mechanism and a native 1-million-token context window. MiniMax describes it as the first open model to fuse frontier-level coding, a 1M-token context, and native multimodality simultaneously — a vendor-stated first-mover claim.

Read the three model releases together and a pattern emerges that no single announcement makes obvious: the price of frontier-adjacent agentic coding is collapsing for teams willing to self-host or route to open-weight APIs. StepFun's own framing is the sharpest illustration — it claims Step 3.7 Flash in Advisor Mode reaches roughly 97% of an Opus-class model's coding performance at about $0.19 per task versus $1.76, a near-ninefold cost reduction. That specific ratio is vendor-stated and unreplicated, so treat the headline number as directional rather than precise; but the direction is unambiguous, and three releases in one week reinforce it.

NVIDIA closed the week with its GTC Taipei keynote at COMPUTEX 2026 on May 31, where Jensen Huang announced that Vera Rubin is in full production. NVIDIA states Vera Rubin delivers 10x agent throughput at scale versus the previous-generation Grace Blackwell platform, with five purpose-built MGX racks operating as one unified AI supercomputer. That throughput multiple is a vendor-stated claim, and production shipments are slated to begin in fall 2026 — so this is a production-ramp announcement, not customer delivery.

The other headline was the RTX Spark superchip: a 20-core NVIDIA Grace CPU co-developed with MediaTek paired with a Blackwell GPU carrying 6,144 CUDA cores and up to 128GB of unified memory, delivering one petaflop of AI performance in slim Windows laptops and compact desktops. It arrives in fall 2026 from ASUS, Dell, HP, Lenovo, Microsoft Surface, MSI, Acer, and GIGABYTE. The supply chain behind Vera Rubin spans 150-plus Taiwanese partners across 350-plus factories in 30 countries. For the broader infrastructure picture, our earlier piece on NVIDIA's AI factory infrastructure sets useful context.

For marketers, the week's most directly actionable story was algorithmic, not model-related. The Google May 2026 core update began rolling out May 21 and was expected to take up to two weeks — meaning it was still in motion across our coverage window. Google described it as a regular update designed to better surface relevant, satisfying content, the same framing it applies to every broad core update.

The genuinely notable element is timing. This is Google's second core update of 2026, following the March update that completed April 8. Only about 43 days separated that completion from the May launch — well inside the three-to-four-month spacing that characterized core updates in recent years. A compressing cadence means sites have less time to recover between updates, which raises the premium on durable, genuinely useful content over quick tactical fixes. If you operate content programs, this is the SEO signal to internalize from the week. Our agentic SEO service is built around exactly this volatility.

Strip the headlines down and the week hands builders and marketers a short list of decisions worth making now. The cost curve moved, the funding cycle reset expectations, and the SEO ground shifted — here is how we'd act on each.

The thread connecting all five stories is leverage shifting toward whoever can run capable models cheaply and at scale. Anthropic priced that leverage at nearly a trillion dollars; StepFun and MiniMax attacked it from the open-weight side; NVIDIA is building the factories that make it physical. Our read for the rest of 2026: the competitive edge belongs less to teams chasing the single smartest model and more to teams that route work intelligently across a mix of open and closed models by task, cost, and risk.

Concretely, the move we'd make is to stand up a small multi-model routing experiment this quarter — Opus 4.8 for high-stakes generation, an open-weight model like Step 3.7 Flash for high-volume agentic loops, and an honest per-task cost ledger to settle the trade-offs with data rather than vendor benchmarks. The economics that emerged this week make that experiment cheaper to run than it was even a month ago.