ChatGPT for Teens, a $10M airline data sale, and Anthropic at $65B (2026)

OpenAI locked teenagers into Study Mode by default. Google spent $10 million on a bankrupt airline's customer files to train models. Anthropic's run rate jumped to $65 billion. Same 24 hours, three different bets on where AI value is stacking up.

ChatGPT for Teens ships with the guardrails on

OpenAI launched ChatGPT for Teens, a separate experience for people under 18.1 Study Mode is on by default. Filters are tighter around self-harm, emotional dependence, and romantic or sexual content. Age prediction decides who lands in the restricted mode; anyone the system flags as under 18 stays restricted until they verify with facial age estimation or an ID.

Parents get blackout hours, activity insights, and age-appropriate model behavior once the teen is linked to a family setup. The product rolls on Free, Go, and Plus. The United States is first; other countries follow.

Why it matters: the big consumer chat apps are no longer one product with a terms-of-service age line. They are splitting the product itself by age band, and verification is becoming the gate. Watch how schools and parents actually adopt the parental controls — that will decide whether this is a real product line or a policy press release.

Google paid $10M for Spirit Airlines' customer data

Google bought Spirit Airlines' customer data out of bankruptcy for $10 million, according to court filings covered by Axios and The Register.2 3 The package covers roughly 36 million customers: names, contact details, travel and loyalty history, account IDs, flight history, and baggage claims. Google told the court it wants the data "to train AI models and improve Google products."

Spirit filed for Chapter 11 again on August 29, 2025, after exiting an earlier bankruptcy on March 12, 2025. Passenger data became an asset on the block like aircraft and gates.

Why it matters: model training is no longer only web crawl and licensed books. Distressed companies are selling behavioral datasets that never sat in a public training dump. If you fly a lot, your itinerary history may already be someone else's fine-tuning set. Watch which other bankrupt consumer brands list "AI training rights" next to trademarks in the auction packet.

Anthropic's annualized revenue hits $65B

Anthropic's annualized revenue reached $65 billion, up from about $9 billion in February — a climb TechCrunch reported on August 17.4 For context on the same curve: the company crossed $1B ARR in March 2025 and $5B ARR in August 2025.

That is not a product launch. It is a demand signal. Enterprise and API spend is concentrating fast enough that a lab can 7x a nine-figure run rate inside half a year.

Why it matters: the commercial race is no longer "who has the cleverest demo." It is who can keep capacity, contracts, and trust under a demand curve this steep. Watch how OpenAI and Google answer on pricing, enterprise packaging, and capacity — revenue like this forces the rest of the field to move.

Snowflake makes model routing a product

Snowflake put Dynamic Model Routing into general availability inside Cortex AI.5 The feature sits in front of multiple model providers — OpenAI, Anthropic, Google, Meta, Mistral, xAI, DeepSeek, and others — and picks a model per request. Three modes ship: Intelligent, Adaptive, and Performance. Snowflake says customers can cut cost by up to 50% without dropping answer quality.

Why it matters: once routing is a platform feature, "which model is best" becomes an ops setting, not a standing architecture decision. Teams that still hard-code one frontier model into every workflow will look expensive next quarter. Watch whether cloud data platforms or the model labs themselves win the routing layer.

Open weights and cyber risk stay in the spotlight

WIRED's August 18 reporting kept pressure on Z.ai's GLM-5.3, an open-weight model the company says refuses most harmful cyber requests and succeeds on only about 1% of them in its own tests.6 Independent security researchers argue the refusal layer is thin once weights are public — jailbreaks and fine-tunes can strip it. Z.ai's own model card and blog frame the release as a coding and agent model with explicit cyber-risk controls.7

Why it matters: open weights and dual-use capability are no longer a side debate. Labs that ship weights now have to argue the security case in public the same week they post benchmarks. Watch for red-team results from outside labs, and for any US or EU guidance that treats high-capability open cyber models as a distinct risk class.

What to watch next

Teen product copycats: Google and Meta already face the same under-18 pressure. If ChatGPT for Teens holds engagement without a safety blowup, expect matching "study-first" modes elsewhere within weeks.
More data auctions: Spirit will not be the last carrier or retailer to sell training rights in bankruptcy. The legal language in those sale orders is the real precedent.
Routing defaults: if Snowflake's 50% cost claim holds up in customer case studies, multi-model routers become table stakes for any serious AI platform budget review.