Five verified moves this week, and a team of one can act on each before Monday. A cheaper workhorse model dropped the floor on token cost again. A Chinese regulation quietly deleted companion agents for hundreds of millions of users — and the EU's version of the same rule starts August 2. Databricks raised into a $188B valuation while nine-figure rounds landed all month. Apple finally shipped AI in China by routing through domestic models. And the MCP betas started a four-day clock. Every item below is dated and sourced; each carries the one line that changes what you do next.

1. Gemini 3.6 Flash undercuts token prices again — re-benchmark your default#

Google announced Gemini 3.6 Flash around July 22, pitched as a cheaper "workhorse" tier. The claim that matters for anyone running agents: it uses roughly 17% fewer output tokens than 3.5 Flash on average, and up to 65% fewer on long-horizon coding, at a reported $1.50 per million input / $7.50 per million output (Google; VentureBeat). It's live via the Gemini API, the consumer app, and Search; 3.5 Pro is flagged as coming.

What it means: Token efficiency compounds on long agent runs, where output volume — not the per-token headline — is the bill. But cheaper-per-token is not cheaper-per-result. Before you swap a default, re-benchmark on cost per completed task for your actual workload; a model that shaves output tokens but needs an extra retry can cost more, which is exactly why one tokens-per-second number lies to you.

2. China's persona law starts biting — and the EU's version lands August 2#

China's Interim Measures for AI Anthropomorphic Interaction Services took effect July 15, and this was the week the effect showed up: ByteDance's Doubao and Alibaba's Qwen disabled personalized companion agents used by hundreds of millions rather than risk non-compliance (IAPP; SCMP). The measures require disclosing AI status at session start, verified parental consent for minors, and anti-addiction and self-harm safeguards. We called this shape when the law passed; now it's operational.

What it means: If you don't ship into China you aren't directly bound — but the EU AI Act's chatbot-disclosure obligation begins August 2, so "tell the user they're talking to an AI" is becoming a global default rather than a regional quirk. The cheap, correct move is to add a plain AI-status disclosure and age gating to any persona or companion product now, before a market forces it on your timeline instead of yours.

3. Databricks hits $188B — data-infra capital keeps concentrating#

Databricks announced a round led by Coatue on July 17, reported at roughly $3B, valuing the company at $188B — up from $134B just five months earlier (TechCrunch; CNBC). It capped a July of nine-figure AI and security rounds — SambaNova (~$1B at $11B), Together AI (~$800M at $8.3B), and Glow out of stealth at $1.2B for AI-era endpoint security among them.

What it means: The capital at the platform layer is consolidating around a few incumbents, which is bad news if your plan is to out-fund Databricks and good news as a demand signal: enterprises are still paying for data and AI infrastructure. For a bootstrapped team the read is to pick a wedge the incumbent won't chase — a specific vertical, a specific workflow — rather than a general platform play.

4. Apple shipped AI in China by routing through Qwen and Baidu#

After roughly 22 months, Apple Intelligence cleared Chinese regulators by integrating Alibaba's Qwen for language and image understanding and Baidu for search (TechCrunch). Alibaba shares rose on the news.

What it means: China requires generative features to run on an approved domestic model — so localizing AI there isn't a data-residency checkbox, it's a model swap. If China is on your roadmap, budget for an architecture where the underlying model is pluggable per region, not hard-wired to one provider. The same modularity that lets you A/B a cheaper model (see item 1) is what lets you clear a regulator.

5. The MCP v2 SDK betas are out — four days to migrate#

The v2 Model Context Protocol SDK betas shipped for all four Tier-1 languages — Python mcp[cli]==2.0.0b1, TypeScript @modelcontextprotocol/server@beta, Go go-sdk@v1.7.0-pre.1, C# --prerelease — ahead of the stateless 2026-07-28 spec that publishes July 28 (MCP blog).

What it means: Existing servers keep working — new clients fall back to the old handshake — but new clients speak stateless by default, and the deploy win (horizontal scaling behind a plain load balancer) is worth claiming before you have real traffic. We walked through installing the v2 betas and flipping a server stateless, with old-vs-new code; it's a one-afternoon job on most servers.

Also this week: Anthropic added model choice to Claude's voice mode (pick Opus, Sonnet, or Haiku for voice), matching the industry's move to make voice table-stakes (TechCrunch). And the US Treasury said sanctions "remain on the table" for Chinese AI firms after a White House official accused Moonshot AI of improperly distilling Anthropic's Fable model (TechCrunch) — a compliance signal worth watching if your stack leans on Chinese open-weight models.