xAI's launch post for Grok 4.7 says the new model is "served at the same price and speed as Grok 4.6." We checked the rate card on September 22, and it is true: $2 per million input tokens, $6 per million output, identical to the old model.

It is also the least useful number on the page. The number that moves an agent's bill is sitting in someone else's benchmark write-up, and it points the other way.

Grok 4.7's token price did not move. The amount of work it bills you for did.

xAI shipped Grok 4.7 into Cursor, its own Grok Build, and the API. The launch post calls it "our most capable model for coding and knowledge work," trained with a longer reinforcement-learning run "weighted toward problems that take many hours to complete." In plain English: they taught it to keep going. That is the feature, and it is also the cost.

Artificial Analysis benchmarked it at its highest reasoning setting and counted the output. Their figures:

Output tokens are the words the model writes, including the thinking it does before it answers. You pay for every one. So at an unchanged $6 per million, Grok 4.7's output bill on a typical benchmark task works out to about $0.49, against $0.22 for Grok 4.6. Same meter rate, a longer ride.

Two details in xAI's own materials are worth knowing before you route traffic to it.

The comparison table mixes effort levels. xAI's launch table sets Grok 4.7 at xHigh against Grok 4.6 at High. Artificial Analysis quotes the old model two ways as well, 36,000 tokens at high and 38,000 at xhigh. Either way the new model writes more than twice as much.

There is a second price, and it applies to the whole request. The rate card on the Grok 4.7 model page has a toggle labelled Higher context pricing, off by default. Flip it and input goes to $4, output to $12. The Markdown version of the same page spells out the rule:

Read that twice if you run long agent loops. Crossing 200,000 tokens of prompt does not just charge the extra tokens more. It doubles the price of the entire call. An agent that has been appending tool output to its context for an hour can cross that line without anyone noticing, until the invoice.

On capability, xAI's table is candid about where it loses. Artificial Analysis ranks Grok 4.7 with Grok Build fourth on its Coding Agent Index, behind Claude Fable 5.1, GPT-6 Astra and Claude Opus 5. In xAI's own launch table, Fable 5.1 Max beats it on CursorBench 4.0 (51.8% to 46.3%) and on Terminal-Bench 4.0 (57.9% to 38.0%).

Qwen-Image-2.1 is open weights, and it is not yours to sell.

Alibaba's Qwen team put Qwen-Image-2.1 on Hugging Face: one model that both generates and edits images, 7B parameters in the image-making part, up to 10 reference images, and native transparent backgrounds. That last one matters more than it sounds. Anyone who has cut a product out of a white background by hand knows what a real alpha channel saves. The model card ships with working diffusers code, so trying it takes an afternoon.

Then open the LICENSE file. It is the Qwen Research License, dated September 20, 2026, and it grants use "FOR NON-COMMERCIAL PURPOSES ONLY." It defines non-commercial as "for research or evaluation purposes only." Commercial use needs a separate licence, requested by email. If you run a paid image API, you can evaluate this model. You cannot ship it without a signed licence.

Xiaomi's MiMo-V2.6-Pro says MIT, in a field that is easy to miss.

Xiaomi published the MiMo-V2.6-Pro weights on September 21: a sparse mixture-of-experts model with 1.02 trillion parameters in total, of which 42 billion fire for any given token. Picture a hospital with a thousand specialists on staff: each patient sees only the few who are relevant. The model card's metadata declares license: mit. When we checked on September 22 the repository had no LICENSE file of its own, so the metadata line is all there is to rely on for now. For a model this size, that is thin paperwork. Xiaomi's own performance claims (parity with Opus 5 on most agent benchmarks) are Xiaomi's, and we have not checked them.

Amazon locked Meta's Muse agent out of its store.

As of Sunday night US time, September 20, Muse users trying to shop on Amazon got a popup saying "continued access by an unauthorized AI agent violates Amazon's Conditions of Use," GeekWire reported. Amazon says Meta never told it Muse would access the store, that the agent does not identify itself, and that it appears to capture and store customer credentials. Meta says Muse "has no visibility into people's passwords or payment methods." Amazon's statement:

When analytics on Postgres slows down, most teams add a second database. Then come the pipelines, the sync jobs, and a copy of your data that's always a little behind.

TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Hypertables partition your data automatically as volume grows. Hypercore compression cuts storage up to 95%. Continuous aggregates keep dashboards live without re-querying everything.

CERN runs Postgres this way for sensor data from the Large Hadron Collider.

No split architecture, no pipeline lag, no new query language to learn. Same SQL, same drivers, same tools.

Start on Tiger Cloud and get $1000 in credits.

Get $1000 Credit

The price of a token is a sticker. The price of a finished task is the bill.

For two years the industry has compared models by dollars per million tokens, because that is what the rate cards print. Grok 4.7 shows why that habit breaks once models are trained to work for hours. A model can hold its token price flat and still get more expensive per job, because it chooses to write more.

It cuts both ways. OpenAI's pricing page lists GPT-6 Astra at $50 per million output tokens, more than eight times Grok's $6. Per task on the same benchmark, the gap closes to under three times, because Astra wrote about a third as many tokens. The ranking holds. The size of the saving is a lot smaller than the sticker suggests.

Rate cards have other traps too. Claude Fable 5.1 lists at $10 in and $50 out on Anthropic's pricing page, but a cache hit on it costs $0.25 per million, half of Grok 4.7's $0.50 cached rate. For an agent that re-reads the same long context on every turn, the most expensive model on the page is the cheaper one on that line. And OpenAI's page notes that "GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026," so one of the prices in xAI's comparison table carries an expiry date.

The same week pushed the other direction too, from below. Tomasz Tunguz wrote about "decision models" like TypeSafe's Jev: models that don't write text at all, and only pick an answer from a fixed list. Think of it as a multiple-choice exam rather than an essay. His footnote cites Jev at $0.042 per million input tokens and $0 for output, a price TypeSafe says "may be subsidized." On 98 hand-checked production email threads, the generative classifier he had been using scored 47%; Jev scored 80%. That is one person's inbox, and he says so. But the direction is clear: the cheapest output token is the one you never generate.

The 20-minute job this week: for your highest-volume agent workflow, log output tokens per completed task for a day. Not per call, not per million. Then check whether any single request comes close to 200,000 prompt tokens. Those two numbers tell you more about next month's bill than any rate card.

⚠️ The strongest case against our read: Artificial Analysis ran Grok 4.7 at xhigh, its most expensive setting. The model page lists low, medium, high and xhigh, with high as the default, so most API traffic will not burn 81,000 tokens a task. More tokens can also mean fewer failed runs, and if 4.7 finishes jobs that 4.6 abandoned, the cost per successful task could fall. We don't have that number, and neither does anyone else yet.

What is the pixel on your checkout page telling someone else?

An independent researcher published a teardown of OpenAI's ChatGPT ads pixel that every builder with a marketing site should read. OpenAI's ad collector sets a cookie called __obi on the .openai.com domain. It is tied to your ChatGPT account and configured to travel across sites. When you then visit a shop that runs OpenAI's conversion pixel, the cookie goes back to OpenAI with the page. The researcher says they saw a single __obi value sent from 12 commercial sites on their own phone.

The advertiser installing the pixel reads OpenAI's developer documentation, which describes "automatic advanced matching": the pixel finds customer details on the page and hashes them before sending. The documentation says "Raw customer information is not sent to OpenAI through automatic advanced matching." None of it mentions a cross-site account cookie. The researcher's point is that the site owner has no way to see it: __obi belongs to a domain their scripts cannot read.

OpenAI's cookie policy lists __obi as an Analytics cookie, one year, on chatgpt.com and openai.com. We confirmed that entry on September 22. It is the only cookie in that section. The researcher's complaint is about that label: OpenAI asks for analytics and marketing consent separately, and every token they decoded carried analytics consent only.

The limits are real, and the researcher states them. They saw it on Chrome for Android. Safari and every iOS browser block the mechanism. Desktop Chrome was not tested. They did not watch OpenAI link the cookie to an account on the server side; they inferred it from the design. And Meta has run the same pattern for years. What is new is running it on a product people confide in. OpenAI's support team acknowledged the questions they sent on September 14 and, according to the post, answered neither.

So the question for your own stack: which third-party tags load on your pricing and checkout pages, and could you say what each one sends, and where?

What does one finished task cost in your stack? Not per token: per job done. Reply with the number, or with the reason you can't get it. I read every reply, and the best answers will shape a future issue.

If someone you work with signs off on the AI bill, forward this to them.

— The Agent Company