The Investor

The Information says OpenAI's run-rate is up about 70% since the start of Q3, going from roughly $41B to nearly $70B, and the report credits enterprise sales and Codex for the climb, with aggressive price cuts helping things along. This column has been suspicious of run-rate figures for as long as companies have been quoting them, because a run-rate is a good month multiplied by twelve and handed to a reporter. Even with that discount applied, this is a big number. I think the growth is the least interesting part of it, because the named drivers pull in different directions. Enterprise sales and Codex are the kind of revenue that tends to sit in a contract and renew. Price cuts are a decision to sell more units for less money each, which means OpenAI is choosing volume over price per customer, and that choice usually lasts longer than whoever made it planned. Discounts show up in gross margin, not in the run-rate. The report does not include margin. There are a few plausible readings. The cuts could be a one-time reset that pulled in customers who stay at the new price, which would make nearly $70B a floor. They could be the opening move in a longer price fight, in which case the run-rate keeps rising while each dollar of it earns less. Or (and this is the version I would most like to see confirmed) Codex and the enterprise book are carrying most of the load, and the discounting matters less than the headline suggests. This is probably wrong, but I lean toward the first reading with an asterisk, mostly because enterprise buyers who sign at a lower price rarely leave in the same quarter. The test comes next quarter. A run-rate that holds near $70B without another round of cuts was a floor, and one that needs fresh discounts to move was bought with them.

The task-cost gap comes from output volume, not list price. Per AINews, Argon emits 62K output tokens per task against Astra's 27K. That makes Argon about 39% cheaper per task only while Google's 50% discount lasts. At its $4/$20 standard price it is 22% more expensive than Astra. The discount has no end date. OpenAI went the other way: it cut its own cost per task 31% between generations (GPT-6 Sol at $1.04, GPT-6.1 Sol at $0.72) without changing list prices, by using fewer turns and cheaper cache reads.

The answer depends on the workload. On Vals' tasks, AINews notes, Argon uses roughly a quarter of Sonnet 5.5's output tokens, the opposite of the Artificial Analysis result. On Terminal-Bench Science, AI Breakfast reports Sol at $5.47 per task against $23.21 for Opus 5.5. Simplifying AI flags a hidden cost: Sol now reasons before every reply and has no 'none' mode. That means high-volume, simple calls can cost more than before, not less. Many of these benchmark figures come from the vendors themselves.

The clone clock

Latent.Space captured the second shift. OpenAI's Decisions API was not on the roadmap four weeks before launch. Two engineers built it in about a week on existing Luna weights, with no new training. It is priced at Luna rates, after Luna's ~80% price cut, and it includes vision, which Jev lacks. About 100 thin clones appeared within two weeks. If a lab can rebuild an API pattern in days, the interface itself is worth almost nothing.

The moat that remains is the one OpenAI admits it hasn't solved: calibration, meaning how well a model's confidence matches its accuracy. Separate evidence in TLDR IT points the same way. In one author's test across five organizations, Jev scored 0.948 F1 on identity resolution at $0.62 per 1,000 accounts. Claude Haiku 4.5 scored 0.894 at 19x the cost. That is one test, and the source doesn't say whether the tester is independent of Jev's maker.

When frontier models tie on intelligence, margin goes to the company that measures cost per finished task, and the API wrapper loses its moat.

What it means for your book

Model-agnostic app companies that route across vendors and track cost per task should expand gross margin. Single-model wrappers and products whose moat is the shape of their API should be priced as features. Last Week in AI adds that Sonnet 5.5 beats Opus 5.5 on agentic coding by running multiple agents within cost limits. That moves value further toward orchestration and routing.

Any browser agent that can run scripts on a page can read what gets autofilled into it. Instinct is where that showed up first. Yuechen Zhao of Informed Ventures demonstrated it with a fake card loaded into Instinct's Vault, and Instinct's spokesperson answered that users 'are always in control of their data and credentials,' which is a sentence about something else. Meta says Muse sidesteps the problem with Stripe single-use virtual cards capped at the approved purchase. Meta is the only one saying so, and nobody has tested it. The fix is a credential that is worthless if it leaks.

The rails are writing the rules

The value is pooling one layer down, under the agents. Mastercard's verifiable intent records what the user authorized and whether the agent complied, and Sherri Haymond says that data is available to issuers 'in our regular charge-back system,' the least glamorous phrase in payments and probably the most important. Visa and Mastercard also run a joint agent registry showing merchants who is behind an agent. The dispute rules decide how losses split among banks, merchants and shoppers. Merchants keep a blunter option, which is to refuse access. Amazon banned Muse, citing undisclosed third parties accessing customer accounts.

Unit economics favor B2B

At Newcomer's summit, Sapiom's Ilan Zerbib estimated an always-on assistant costs about $10 a day to run. Against a $20 monthly subscription, Newcomer works out, that is roughly 15x underwater, which is a consumer business only in the sense that consumers are involved. Beside a knowledge worker's salary the same cost rounds to nothing, so real ticket sizes are turning up in B2B (one Browserbase customer's agent made a $300,000 card purchase). The hard enterprise problem stays open. ByteByteGo reports that DoorDash, which runs agent traffic at scale, has not yet shipped cryptographic identity for each agent or short-lived credentials scoped to a single task. National Cyber Director Sean Cairncross has framed the buyer's test, per CyberScoop, around who is authorized and what they can access, plus where they run.

In agentic commerce the moat sits with the credential holder and the dispute settler, along with whoever verifies the agent. The chat window on top rents it.

Incumbents are hedging the same way. The Information reports DoorDash claims its assistant lifts grocery baskets nearly 50%, against a baseline nobody has specified, so treat the figure as likely selection-biased. DoorDash is also building its own assistant while supplying ChatGPT and talking with Muse. This is probably wrong in the details, but the logic is plain: the owner of the customer's intent eventually charges rent, and DoorDash would prefer not to be the tenant.