Short version: Only one thing on this page has a legal date on it, and that date is today. The EU AI Act's Article 50 transparency rules are now in force — if EU users can reach your AI, you owe them disclosure and machine-readable marking of AI-generated content. Everything else that happened this week was the price of intelligence moving, not the frontier of it: OpenAI cut the cheap tier hard, Anthropic shipped a frontier-adjacent Claude at flat pricing, and the largest open-weight model ever went public. So do the compliance work today, spend an hour re-pricing your routing this week, and don't rebuild your agent around any one model — the tier you'd pick is a moving target this month.
The one deadline: EU AI-transparency duties are live as of today#
As of August 2, 2026, the EU AI Act's Article 50 transparency obligations apply. Three duties matter for a builder:
- Disclose the AI. When a person interacts with an AI system (a chatbot, a voice agent), you must tell them — unless it's obvious from context.
- Mark the output. Generative-AI outputs must be marked as artificial in a machine-readable format. That covers synthetic audio, image, and video, and text published to inform the public on matters of public interest.
- Label deepfakes. Content that resembles real people, places, or events and is AI-manipulated must be clearly labeled.
The teeth: non-compliance carries penalties up to €15 million or 3% of worldwide annual turnover, whichever is higher. The reach: it's extraterritorial — a non-EU startup is covered if its output is used in the EU. The mercy: it's not retroactive — content generated before today isn't covered, so this is a go-forward pipeline change, not a back-catalogue cleanup.
What it means: this is the only item this week with a date attached, and the date is now. If EU users can sign up, add interaction disclosure and content marking to your generation path this week. We have a founder compliance checklist and a what-to-actually-ship guide for the machine-readable part.
The week the cheap tier moved: OpenAI cuts Luna ~80%#
On July 30, OpenAI cut API prices on the lower tiers of the GPT-5.6 family. As reported by Axios and VentureBeat: Luna fell ~80% (from $1.00 to $0.20 input and $6.00 to $1.20 output per million tokens), Terra fell ~20% ($2.50→$2.00 input, $15→$12 output), and flagship Sol held at $5/$30.
What it means: the floor dropped and the ceiling didn't, so the spread between "cheap enough for bulk grunt work" and "good enough for the task you can't get wrong" got wider. That's a routing signal, not a switch-everything signal. Re-run your cost-per-completed-task by task class — not per token — and re-route only the classes where Luna's new number actually wins your total bill. The sticker-versus-bill breakdown shows where the naive per-token math misleads.
The frontier held its price: Claude Opus 5#
On July 24, Anthropic shipped Claude Opus 5 at $5 input / $25 output per million — unchanged from Opus 4.8 and roughly half Fable 5's input price — and made it the default on Claude Max. Prompt caching still buys up to ~90% off cached reads; batch, ~50%.
What it means: unlike the OpenAI move, nothing about your Anthropic bill changed. If you benchmark agent quality against the frontier, re-baseline against Opus 5 — but there's no pricing pressure forcing a migration this week. For the head-to-head on when the cheap open tier is enough versus when you still want the frontier default, see Kimi K3 vs Opus 5.
Open weights hit 2.8 trillion: Kimi K3 goes public#
Around July 26–27, Moonshot AI published open weights for Kimi K3 — a ~2.8-trillion-parameter mixture-of-experts model (~104B active per token) with a 1M-token context, described as the largest open-weight model released to date and tuned for coding and agents.
What it means: it's a genuinely near-frontier model you can, in principle, run yourself and keep your data in-house. The catch is the same one every large open weight carries: at 2.8T parameters, self-hosting is real infrastructure, not a side project, and the build-vs-buy line only crosses in your favor at high, steady utilization. Trial it on a hosted endpoint first; decide on self-hosting once the meter proves it out; and read the repository LICENSE before you ship commercially — Moonshot's earlier K2 line carried a name-attribution clause above large revenue and usage thresholds. Our self-host-vs-API cost math and K3 founder guide go deeper. And the direction of the open-weights policy debate — where Anthropic just drew its line — is decoded in Amodei's position, for founders.
The plumbing locked down: MCP goes stateless, inference engines ship together#
Two infrastructure moves that don't make headlines but change what you build on:
- MCP's 2026-07-28 spec finalized with a stateless core — it drops the
initialize/initializedhandshake and theMcp-Session-Id, adds header-based routing and cacheabletools/list, and puts Roots, Sampling, Logging, and Dynamic Client Registration on a 12-month deprecation offramp. Practically: a remote MCP tool server can now sit behind a plain load balancer with no sticky sessions. Audit your server for the deprecated features now — the clock is running. See what breaks in the stateless core. - vLLM 0.26.0 and SGLang 0.5.16 shipped the same day (July 25) — vLLM matured tiered KV-cache offload and per-KV-group attention backends; SGLang made its radix-tree prefix cache the default and added confidence-driven speculative decoding. If you self-host, these are the versions to target for throughput per dollar. The two engines' diverging bets are laid out in the memory-hierarchy fight.
What to do this week#
- Today: if you serve EU users, ship AI-interaction disclosure and machine-readable output marking. It's the only legal deadline here, and it's not retroactive — every unmarked output from today forward is exposure.
- This week: pull your token bill by task class, drop in the new Luna/Terra numbers, and re-route the classes where the cut flips the winner. Don't rebuild — a model-swappable router is the durable win, not a new vendor lock-in.
- On the backlog: audit any MCP server you run for the deprecated Roots/Sampling/Logging/DCR features, and re-baseline your frontier evals against Opus 5.
This is the same lesson the last month of releases keeps teaching: capability is leapfrogging weekly, so the thing that compounds isn't picking this week's winner — it's staying cheap to switch. For the wider read on how the model market got here, the week-of-July-20 edition is where this cycle's story starts.



