GPT-6 shipped and the price of AI moved into the contract
The headline number stopped being what mattered this week. OpenAI shipped GPT-6 at the same list price as its closest rival and then rationed who could use it. Meta cut its price by more than 90% for customers who hand over their data. Even the regulator moved on terms: Europe now audits ChatGPT as a search engine. The condition is now the price.
OpenAI ships GPT-6 Astra, then rations access to it
On September 3 OpenAI released GPT-6 Astra, a model that operates apps and browsers as a person would. The list price is $10 per million input tokens, the units of text a model processes, and $50 per million output. OpenAI rates it "Critical" for cyber capability, a first, so vetted security customers got it before paying customers. Sam Altman apologized the next day for the rollout.
OpenAI now decides who gets the frontier first, and when, and the newest model is not automatically the best buy. An independent benchmark, Artificial Analysis, scored it below Claude Fable 5.1 on September 3. Plan any migration around the date you actually get access.
Source: OpenAI
Anthropic releases Claude Fable 5.1 and cuts cached-input prices by 75%
Anthropic released Claude Fable 5.1 on September 1 with its list price unchanged. What changed is the cost of re-reading context the model has already seen: cached input fell 75%, to $0.25 per million tokens. The company says typical bills drop by about a quarter and agent-heavy ones by close to half.
The discount rewards workloads that reuse the same context many times, which is what agents do. A single question pays the same as before. The savings figures are Anthropic's own; nobody outside the company has verified them yet.
Source: Anthropic
Meta cuts Muse Spark 1.3 prices over 90% for customers who share data
Meta released Muse Spark 1.3, a model for coding and multi-step agent work, on September 2 with two price lists. The standard tier costs $1.25 per million input tokens. A "contributor" tier costs $0.10 for the same model, with tighter rate limits, in exchange for letting Meta train on the customer's prompts and outputs.
This is the clearest price a major vendor has put on a customer's own traffic. A client memo, a sales pipeline, a draft contract: each now has a market value, and it is most of the bill.
Source: Meta
OpenAI ends Cursor's access to its models after the SpaceX takeover
OpenAI said on August 28 that it will end Cursor's access to its models on November 12. Cursor is a coding tool that SpaceX bought for $60 billion in August. A change-of-control clause let OpenAI cancel, and it says it cannot trust the new owner to honor its terms; Cursor's chief executive puts OpenAI at about 5% of its traffic.
The damage to Cursor is small. The clause is what matters: a supplier can withdraw when your company changes hands, on a trigger you never chose, and the same clause sits in many enterprise AI contracts. Any operator planning a sale or a fundraise that changes control now has an AI-vendor line on the diligence list.
Source: OpenAI
The EU puts ChatGPT under the same rules as Google Search
On August 31 the European Commission designated ChatGPT a "very large online search engine" under the Digital Services Act, the EU's platform law. The threshold is 45 million monthly users in the EU, which OpenAI's own declared figure for ChatGPT search clears several times over. The company has four months from the designation to complete a systemic risk assessment, submit to independent audits and share data with regulators.
An assistant is now regulated as search infrastructure in Europe, with risk audits, researcher access to data, and transparency on how it ranks what it shows. For anyone whose customers find them through ChatGPT, that is the first outside view into how the answers get made.
Source: European Commission
ChatGPT's ad business reaches a $1 billion annual run rate in 200 days
OpenAI said on August 31 that advertising in ChatGPT has passed $1 billion in annualized revenue, about 200 days after the first ad ran. Ads run in more than 40 countries, and self-serve buying opened the same day to advertisers in Europe.
Search budgets now have a second destination, in front of an audience that asks in full sentences. European agencies can buy it directly from this week.
Source: OpenAI
The move this week is a reading exercise. Pull the contracts behind every AI line item you pay for and find three things: the change-of-control clause, what the vendor may train on, and what any discount actually applies to. The list price will keep falling. The terms are where the next surprise is.
Meta and OpenAI Started Charging for AI by Who You Are
Meta and OpenAI priced AI access this week by who the buyer is, and the list price stopped being the price.
On September 2 Meta published two prices for the same model. Muse Spark 1.3, its model for coding and agent work, costs $1.25 per million input tokens on the standard tier and $0.10 on a "contributor" tier, per Meta's developer pricing, if the customer lets Meta train on their prompts and outputs. A day later OpenAI released GPT-6 Astra to security-vetted organizations before its paying subscribers. In one week, four vendors, a legislature and a regulator priced AI by an attribute of the buyer. The list price stopped being the price.
Four vendors, four attributes
Meta priced the customer's data. The contributor tier is the same model with tighter rate limits at roughly a twelfth of the standard input price: a market quote for training rights on a company's own traffic. On OpenRouter, the service developers use to reach many models through one interface, the contributor version is listed as a separate model name. The consent lives in a string an engineer types.
OpenAI priced the buyer's security posture. GPT-6 Astra is the first model OpenAI has rated "Critical" for cyber capability under its Preparedness Framework, the company's own risk rules. On September 3 it went first to vetted organizations in the company's security program, per OpenAI and Bloomberg. Paying ChatGPT subscribers waited, and Sam Altman apologized the next day, per TechRadar. Paying more did not move a customer up the queue; passing a vetting process did.
Microsoft priced geography. From September 1, per Microsoft's own pricing announcement, running a model in its EU Data Zone, the deployment option that keeps processing inside the European Union, costs 9% more than the global pool, and the premium applies only to models launched on or after that date. A European team that wants the newest model and its data kept home now pays for both.
Google priced time. Gemini 3.8 Flash shipped on September 2 at $0.75 per million input tokens, with both input and output prices set to double on January 1, 2027, per Google's launch pricing. The introductory rate is a number with an expiry date, four months out.
The state did the same thing
Brazil's Senate approved a data-center tax regime on September 1 and sent it to the president for signature, per Agência Senado, the Senate's own news service. It suspends import and production taxes on data-center equipment for five years, on conditions: renewable or low-emission electricity, a water-efficiency ceiling for cooling, in-country research investment, and at least 10% of installed computing capacity reserved for the Brazilian market. That last clause is a legislature doing what the model vendors did: the discount is real, and the price is an attribute of the buyer.
The European Commission moved on August 31, designating ChatGPT a very large online search engine under the Digital Services Act, the EU's platform law. OpenAI's own declared figure put ChatGPT search at about 159 million monthly users in the EU, more than three times the 45 million threshold. Within four months the company owes a risk assessment, independent audits, and a public repository of every advertisement it shows. Access to that audience now comes with disclosure attached.
The reading this argues against
The strongest case for the other side is that this is commoditization and nothing more. Headline rates are converging: OpenAI's newest model and Anthropic's, released two days apart, both list at $10 per million input tokens and $50 per million output, per the two companies' published pricing, and two budget models released the same week undercut them roughly tenfold. On this reading the contributor tier is a launch promotion, the vetting queue is a rollout problem, and the EU premium is a rounding error. Prices fall and the buyer wins.
The convergence is real. It is also the reason the tiers exist. When rivals match each other at the headline rate, that rate stops carrying information. Artificial Analysis, an independent benchmarking firm, published its scores on September 3. Anthropic's Claude Fable 5.1, released two days earlier, ranked above GPT-6 Astra. A vendor that cannot win on the headline number differentiates on terms, as cloud providers did years ago with regional pricing. The commoditization reading predicts one price; the week produced one published price and a spread of conditions around it.
One strand of this is thinner than the others. Meta's tier is four days old and its uptake is unknown. Google has extended introductory pricing before. Microsoft's premium is modest. What is verified is the direction: every one of these vendors chose to price a buyer attribute rather than cut the headline rate.
Ownership is an attribute too
The clearest example is the one the Newsletter, GPT-6 shipped and the price of AI moved into the contract, led with. OpenAI said on August 28 that it will end Cursor's access to its models in November, invoking a change-of-control clause. The trigger was SpaceX's purchase of the coding-tool company for $60 billion in August, per CNBC. Cursor's chief executive puts OpenAI at about 5% of its traffic, per the same report, so the damage is contained. Who owns the buyer is now a term the supplier prices, and the price of a change in ownership can be the supply itself.
In July, in The model's edges got a price, we argued that the model's inputs, outputs and actions had acquired external prices. This week the pricing crossed to the other side of the contract. The buyer's data, security standing, jurisdiction, timing and ownership each got a number.
Who should be uncomfortable
Anyone whose 2027 AI budget is a list price multiplied by a volume forecast. The list price is now the least stable input in that spreadsheet. Around it sit two prices on two different things: a 9% surcharge for keeping data in Europe, per Microsoft, and a 92% discount for handing over traffic, per Meta's price list.
Any company with a sale, a control-shifting fundraise, or a group reorganization scheduled in the next twelve months, because one AI supplier has just treated that as a trigger.
European teams that assumed parity with US deployments, and consultancies whose pitch is the cheapest token, now that the cheapest token carries a condition.
The move
Put a price on your own attributes before a vendor does. Meta's gap gives you the market's number for your traffic: roughly ninety cents of every dollar on that model, per its September 2 price list. Decide in writing whether that trade is allowed, and check your usage logs for the contributor model name, because someone may already have picked it. Read the change-of-control clause in every AI agreement before your diligence list is written by a buyer. For European deployments, model the EU premium against the cost of staying one model behind.
What would make this read wrong is visible within two quarters. If the contributor tier is withdrawn for lack of takers, the introductory Flash price is extended indefinitely, and OpenAI returns to spend-based access, the week was launch noise and the commoditization case wins. Watch instead whether Anthropic or OpenAI publish a training-for-discount tier of their own by year-end. If they do, buyer-attribute pricing is the model, and the published price is a marketing document.
The position
Utilities charge everyone the same price for the same unit, and for two years AI was sold that way. This week the vendors, a legislature and a regulator stopped. A discount for your data, a queue that reads your security file, a surcharge for your jurisdiction, an expiry date on your rate, a clause on your ownership: each one prices something the buyer has and the vendor wants. That is how platforms behave, and platforms are harder to leave than utilities. The operator who wins the next twelve months is the one who knows what their own attributes are worth before signing, because the vendor already has a number.