The AI price floor just moved

Most teams add a second database for analytics. Then manage sync, lag, and drift forever. TimescaleDB extends Postgres instead. Hypertables, 95% compression, aggregates. No pipeline. Get $1000 Credits To Start.

I spent Saturday morning reading a pricing table instead of shipping, which is not how I planned the weekend. Three of your AI invoices moved this week, one of them quietly, and the fix most people will reach for costs triple (about 4 minutes).

Here's the signal today:

• 🪤 Two DeepSeek model names die July 24, and the obvious swap costs 3x

• 🔒 Anthropic's free Fable 5 window closed July 19

• 📉 Google quietly cut Gemini's output price 17 percent

• 🧾 Kimi K3 prices frontier work at half of Opus, with a catch

• ⬇️ Inkling put 975B open weights on Hugging Face, downloadable now

• 🧊 New York froze big data centers, and it is not an AI rule

• ✂️ Google opens your app to rival stores July 22, and still takes its cut

• 🇪🇺 Brussels orders Google to open Android to rival assistants

• 📚 A judge put a $3,000 price on every pirated training book

• 🤖 The EU makes you tell people they are talking to a machine

• 🍁 Canada's new hate speech law, decoded for creators

⏰ The Meter: what your agents cost this week

Two DeepSeek names die July 24, and the obvious fix costs 3x. DeepSeek retires deepseek-chat and deepseek-reasoner at 15:59 UTC that day. Both are aliases today for deepseek-v4-flash, $0.14 in and $0.28 out per million. Here is where it gets expensive. Reasoner sounds like it should become the pro tier, so that is the swap most people will make, and deepseek-v4-pro runs $0.435 and $0.87. You just tripled your bill for a model you were not using before. The like for like move is flash with thinking mode set explicitly. Grep for both names before the 24th or you get a broken agent and a surprise invoice in the same hour.

Anthropic's free window already closed. Included Claude Fable 5 access on paid plans, extended twice, ended 11:59 PT on July 19. Every call now meters at $10 per million in and $50 out. Same model, same code you already shipped, now with an invoice attached. If you have agents calling it on a loop, the bill will tell you before the logs do. Worth an hour today sorting which jobs actually earn back $50 a million tokens and routing the rest to Sonnet.

The cache line nobody reads. A cache hit on deepseek-v4-flash costs $0.0028 per million against $0.14 on a miss. If your agent resends the same system prompt every call, and mine do, that is a fiftieth of the price for work you are already paying for. I found this at the bottom of the same pricing table that carries the deprecation notice, which tells you something about how vendors rank the news they give you.

Google quietly cut Gemini's output price 17 percent. Gemini 3.6 Flash shipped July 21 at $1.50 in and $7.50 out per million, down from $9.00 out on 3.5 Flash, with input unchanged. A lighter 3.5 Flash-Lite runs $0.30 and $2.50. DeepMind also confirmed on the record that pre-training has begun on Gemini 4. If you have batch jobs pinned to 3.5 Flash, the swap is one model string and a smaller invoice for identical work.

🏷️ Agent Watch: the frontier went on clearance

Kimi K3 landed, and the sticker price is not the price. Moonshot shipped a 2.8 trillion parameter model July 16 at $3 in and $15 out per million, half or less the per task cost of Claude Opus 4.8. Artificial Analysis puts it at 1668 Elo on GDPval, above Opus 4.8 at 1600 and below Fable 5 at 1760. Then Simon Willison actually ran it and found the catch: it ships with one reasoning effort, max, and burned 13,241 reasoning tokens to produce 3,417 tokens of answer. One SVG drawing cost him 25 cents. Cheap per token and expensive per task are the same model here. Weights land July 27.

Inkling is the one you can actually download. Mira Murati's Thinking Machines put 975 billion parameters on Hugging Face July 15 under Apache 2.0. Commercial use, no licensing fight, no waiting list. It is not the best model in the room: Sebastian Raschka's teardown has it losing to GLM-5.2 on agentic coding, 63.8 percent against 82.7 on Terminal-Bench. And nobody has published what it costs to actually serve 975 billion parameters, which is the number that decides whether owning beats renting. Everyone spent the week talking about K3, which you cannot host yet, while the model you can host shipped a day earlier to a fraction of the noise.

Orca runs five coding agents at once. An open source orchestrator that runs Codex, Claude Code, OpenCode and Pi side by side, each in its own git worktree. Fan one prompt across all five and compare what comes back. MIT licensed, v1.4.146 shipped July 18. I have been doing this by hand in parallel worktrees for months, so watching someone package it properly is a small relief.

🔌 Continental: your hosting decisions are now political

New York froze the data center, sort of. Every headline I saw called this a ban on AI data centers. Read Executive Order 62 and it is neither. Not a construction ban, not an AI rule. State environmental permits pause for any facility that can draw 50 megawatts or more, until a grid and water study finishes. A 40 megawatt AI training site is untouched. A 60 megawatt streaming warehouse is caught. Local permits are carved out entirely. The tell is who is unhappy: the Data Center Coalition says the investment will now flow elsewhere rather than to New York, while Food and Water Watch, which campaigned for the pause, says the 50 megawatt line leaves communities facing clusters of smaller proposals unprotected. Too broad and too narrow, from opposite sides, about the same paragraph.

Google opens your app to rival stores July 22, and still takes its cut. Every US Play listing becomes available to third party Android stores as of July 22, a remedy out of the Epic case, and the opt-out deadline is the same day the program starts. Here is the part the announcement buries: Google Play's service fee still applies to downloads made through those stores, and the stores themselves pay $5,000 a year for catalog access. Distribution outside Play does not escape Google's cut. The setting is in Play Console under Settings, then Catalog Settings. When a company writes the default, doing nothing is still a decision, just not yours.

Brussels told Google to open the door. Two binding decisions July 16 order Google to give rival AI assistants the same Android access its own gets, and to share anonymized Search query data with competing engines. If you have ever built something that Android quietly kept at arm's length, that is no longer just your suspicion. It is a documented obligation with a regulator behind it.

🛡️ The Shield: what you are allowed to say and ship

Training on pirated books now costs $3,000 a book. A federal judge gave final approval July 20 to Anthropic's $1.5 billion settlement with authors, roughly $3,000 per work across about 500,000 books, the largest copyright recovery in US history. The judge also cut the requested legal fees from $187.5 million to $101.5 million. The strategic detail nobody is saying out loud: because Anthropic settled instead of appealing, last year's ruling that training on lawfully obtained text is fair use never becomes binding precedent. The piracy, not the training, is what cost $1.5 billion. If you fine-tune on scraped data, that distinction is your entire risk model.

Tell people they are talking to a machine. EU AI Act Article 50 applies August 2. If your bot, AI email responder or SMS agent reaches anyone in the EU, it has to disclose that it is AI, and separately, generated output needs machine readable marking. The disclosure duty has no grace period. This catches solo operators, not just labs, and the disclosure is the easy half. The marking requirement is the one that will take real work.

Canada redrew the hate speech line. Bill C-9 created a new offence: wilfully promoting hatred by publicly displaying listed terror symbols. Up to two years. I watched creators start pulling archive footage the day it passed. Read the text first: journalism, education and art are expressly carved out, hatred is defined strictly as vilification and detestation, and speech does not qualify merely because it offends. Prosecutions need the Attorney General's sign off, so one angry complaint cannot drag a creator into court.

The part worth watching. The old good faith religious opinion defences were repealed, and the symbols list rides on a registry cabinet can update by regulation, a design the Canadian Bar Association flagged. A list that can grow without going back through Parliament is a different kind of law than one that cannot. The criminal line still sits far above the platform moderation line, and knowing that gap is the whole edge.

Hit reply and tell me which one of these cost you money this week. I read every response, and it decides what leads tomorrow. Forward this to one builder who would use it. They can sign up at nomadsignal.ai.

Chase & Kobe 🐕