Signal #76: Meta priced your data at 10 Cents

Three frontier models shipped inside 36 hours and landed within two points of each other. The real difference was in the rate card: Meta will sell you the same model for a twelfth of the price if you let it train on you.

Model weeks used to be about who won the benchmark. This one wasn't. Anthropic, Google and Meta all shipped between Tuesday and Wednesday, the scores clustered, and the only numbers that separated them were prices, expiry dates and eligibility rules.

Below: what Meta wrote into its SKU, why Romania's new AI factory is a 2027 answer to a 2026 question, what Broadcom's quarter says about buyers who stopped accepting terms, and the fortnight in which ChatGPT became both an ad network and a regulated search engine.

Anthropic shipped Claude Fable 5.1 on 1 September, holding base pricing at $10 and $50 per million tokens and cutting cache reads from $1.00 to $0.25, worth about 25% off a typical bill. Google released Gemini 3.8 Flash the next day at an introductory $0.75 and $3.75, which becomes $1.50 and $7.50 on 1 January 2027.

Meta launched Muse Spark 1.3 the same day, and on Artificial Analysis's Intelligence Index its xhigh setting ties GPT-5.6 Sol, Grok 4.6 and Claude Opus 5 at 61, with Opus reaching 63 at its ceiling. Two points separate the frontier.

What did not converge is the terms. Muse Spark 1.3 ships as two SKUs on the same million-token context: Standard at $1.25 and $4.25 per million, and Contributor at $0.10 and $0.20, where the listing states that prompts and outputs may be used to improve Meta's products. That is 12.5 times cheaper on input, 21 times cheaper on output, with the consideration printed on the price list instead of buried in a data processing addendum. Anthropic's best release this week, Mythos 5.1, is the same model with looser refusals and cannot be bought at all, only granted to vetted organisations.

So price your own answer before an engineer answers it for you in a config file. Run the Contributor rate against last month's token spend, then read what your customer contracts promise about training and sub-processors, because a twelvefold saving that breaches a DPA is not a saving.

Pioneer, the summit where CX leaders redefine what’s possible, is on October 7th.

Join leaders from Fin, Anthropic, Clay, and Kalshi for an insightful conversation on the state of AI transformation.

You’ll discover how some of the most innovative minds in CX have transformed their organizations, learn how they think about CX, and hear how they're planning for what's next.

Join the conversation in San Francisco, or tune in virtually.

Reserve your spot

On 1 September, RO AI Factory entered its implementation phase at ICI Bucharest, co-coordinated by Politehnica Bucharest and funded through EuroHPC. The consortium is unusually broad for a research programme: the Technical University of Cluj-Napoca, the Mihai Drăgănescu AI research institute, the Transilvania IT Association, the National Council of Romanian SMEs and the Romanian Digital Innovation Hubs Association. The stated mandate is to move startups and smaller companies from using someone else's technology to building their own, with training, mentoring and validation attached to the hardware. Operation is targeted for the end of 2027, and no budget or GPU count has been published.

Two days earlier, EuroHPC signed a €387.8m contract with Bull for LUMI-AI in Kajaani, on AMD Instinct MI430X accelerators, ten times the AI capacity of the current LUMI, reaching users in 2027. Half the money is Digital Europe Programme, half a consortium of Finland, Czechia, Denmark, Estonia, Norway and Poland. It is EuroHPC's sixth AI Factory contract. Allocation tracks contribution share, so the terms of access to European compute are being set now, in national budget lines, by people who are not asking founders what they need.

The practical route in is the consortium list, not the ministry: Transilvania IT and the SME council are the two doors worth knocking on this month, while allocation policy is still a draft rather than a rulebook. Keep the calendar honest though. A machine that reaches users at the end of 2027 does nothing about the pricing decision sitting on your desk in September 2026.

Broadcom reported third quarter results on 2 September: $29.6bn in revenue, up 86% year on year, of which AI semiconductors accounted for $16.7bn, up 221% year on year and 54% sequentially. Fourth quarter guidance puts total revenue near $34.8bn with AI semiconductors at $21.7bn. Free cash flow was $13.7bn, 46% of revenue. Hock Tan's line on the call was that demand for custom AI accelerators and networking "continues to be very strong."

Sequential growth of 54% is not a ramp, it is a step. Custom accelerators exist because the largest buyers of compute decided that renting capability on someone else's rate card was the more expensive option, and spent years and billions designing around it. The same logic that makes Meta's Contributor tier tempting for a small team makes a bespoke XPU rational for a hyperscaler: at sufficient volume, you stop negotiating terms and start owning the thing that sets them. The gap is that a hyperscaler can afford the exit and you cannot, which leaves portability as the only bargaining chip you own.

For anyone building on top of a single model provider, the watch item is your own switching cost, measured honestly. Count the prompts, evals and tool definitions that would need rewriting to move providers, and if that number is large enough that the January price change does not actually change your behaviour, you have already accepted whatever terms arrive next.

On 31 August, OpenAI said ChatGPT Ads had reached a $1bn annualised revenue run rate in under 200 days, and opened self-serve access to Ads Manager across India, Europe, the Middle East and North Africa the same day. The platform now runs in more than 40 countries, and OpenAI frames the advertising business as what pays for free access for more than a billion weekly users.

Hours later, the European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act, alongside Reddit and Roblox as platforms, after each declared at least 45 million average monthly EU users. Obligations attach four months from notification, so January 2027: systemic risk assessment covering illegal content, effects on minors, physical and mental wellbeing, fundamental rights, electoral processes and public security, applied to the service and to its algorithmic systems. A general-purpose assistant is now regulated as search infrastructure, on a separate track from the AI Act, with its own enforcement and its own fines.

Two things to do with that. Marketers should treat the September window as the cheap part of an auction that will not stay cheap, because acquisition costs in an unsaturated channel are a temporary condition, not a strategy. Founders building assistant or agent surfaces at consumer scale now have a number to plan against: 45 million EU monthly users is the tripwire, and the compliance work it triggers takes longer to build than the feature that crosses it.

Dev: Kilo went native on JetBrains on 1 September, as a Kotlin plugin rather than a VS Code port. It runs several coding agents at once in isolated git worktrees inside one IDE window, adapts existing run configurations to each worktree, and supports Gateway and dev containers. Covers IntelliJ, WebStorm, PyCharm, GoLand, PhpStorm, Rider, CLion and RubyMine. Bring your own API keys. Worth an hour if your team lives in JetBrains and has been told agents mean switching editors.

Productivity: Perplexity shipped Hybrid Compute on Mac on 1 September. Cloud handles reasoning, search and planning while a local model handles private files and on-device actions, with an on-device gate that keeps credentials, card numbers and government IDs local, refuses the action, or rewrites the request. Runs Gemma 4 E4B or a Perplexity in-house model on any Apple silicon Mac with 24GB of unified memory on macOS 15 or later. Pro, Max and Enterprise.

Finance: The Computable GPU Index publishes an open, verifiable price for GPU compute in dollars per GPU-hour, with the methodology and data on GitHub so any published value can be reproduced from the public record. If you are writing a board slide or a model that assumes a compute price, this is the first citable reference point rather than a vendor quote.

Founders: Challenger's August report, out 3 September, counted 52,881 announced US job cuts, down 38% on August 2025 and the lowest August total since 2022, though up sharply on July's 33,429. Restructuring led the stated reasons and AI attribution fell back after five months at the top. Read it as a caution about narrative: what gets blamed for a layoff moves faster than what causes one.

Research: Google's Fairwind programme is now the access route to Gemini 3.8 Flash Cyber, which posts frontier results on vulnerability discovery and automated patching. It is open to government bodies, critical infrastructure operators and software maintainers. If you maintain a widely used open source package, you plausibly qualify, and applying costs an afternoon.

Next edition soon,

Çelik