The AI Intelligence Brief

Anthropic closes $3.5B Series E — what it means for the model layer

The largest private AI raise ever signals a shift in how foundation model companies are valued. Anthropic is no longer priced as a research lab — it's priced as an infrastructure company. The round values safety-first positioning as a commercial advantage, not a constraint. For builders: this means Claude's API stability and enterprise features are going to improve faster than any competitor. For investors: the model layer is consolidating into three viable players. Everyone else is either infrastructure or application. Pick your lane.

Opus 4.7 is GA — what actually changed

The headline is reasoning quality, but the real story is agent reliability. Opus 4.7's tool-use accuracy is measurably better in multi-step workflows. Memory for Managed Agents hit public beta the same week. If you're building agents on Claude, this is the release where it stops feeling experimental. What to do with it: revisit any agent pipeline you shelved in Q1 because tool-calling was too flaky. The error rate has dropped enough to make production deployments viable for mid-complexity workflows.

EU AI Act enforcement begins — who is ready and who is not

The first enforcement deadlines hit this month. Most startups building with foundation models are classified as "deployers" not "providers" — which means lighter obligations, but obligations nonetheless. The main one: transparency. If your product uses AI to generate content that a user might mistake for human-made, you need to disclose that. Three companies we track have already been flagged for non-compliance in their marketing copy. Fix this before a regulator fixes it for you.

Enterprise AI spend up 40% YoY. Three sectors leading: financial services (compliance automation), healthcare (clinical documentation), and legal (contract analysis). Seed rounds for AI-native companies are down 15% from peak but median quality is up — investors are getting pickier, not quieter.

Mistral released Codestral at a price point that undercuts GPT-4o for coding tasks by 60%. Meta's Llama 4 Scout is the new default for on-prem deployments where data can't leave the building. Google shipped Gemini 2.5 Pro with native audio understanding — the first frontier model where you can pipe in a meeting recording and get structured output without a transcription step.

AI writing assistants that don't own distribution are getting squeezed. Jasper laid off 20% last quarter. Copy.ai pivoted to workflow automation. The category hasn't died — but the standalone "AI copywriter" product is being absorbed into platforms that already own the customer relationship. If you're building a wrapper, read our Wrapper or Winner framework before your next board meeting.