Editions
The Same Model at a Different Price
Three AI vendors charge exactly 10 percent more to keep inference in one country. On a cloud marketplace the same model costs 56 percent more in London than in Virginia.
November 6, 2026The Per-Minute Transcription Price Is an Estimate
Five speech-to-text rate cards, five billing units. One vendor prints the assumptions behind its per-minute figure: 25 audio tokens a second in, 175 text tokens a minute out.
November 5, 2026Embedding Prices Come in Four Different Units
Per million tokens, per image, per billion pixels, per instance-hour. Four vendors price embeddings four ways, and one of them hides its rates behind a collapsed table row.
November 4, 2026The Fine-Tune You Keep Paying Rent On
Training a custom model is a one-off charge. Keeping it callable is a monthly one, and on one platform the vendor's own worked example puts a year of hosting at over fifteen thousand dollars a month.
November 3, 2026The Fifty Percent You Get for Waiting
Three AI platforms cut token prices by exactly half if you accept a 24-hour turnaround. The discount is identical on all three rate cards. What it covers, and what an unfinished job costs, is not.
November 2, 2026How to Size an AI API Account Before You Pay
Six things to read before the first API key: the free tier's data terms, the purchase that lifts a tier, the models a new project cannot use, and the tool calls that bill separately.
November 1, 2026Rate Limits Are the Real Tier Difference
Moving up a paid API tier buys throughput, not features. One vendor multiplies tokens per minute forty times between its first and last tier. Another has stopped printing the table.
October 31, 2026What a Million-Token Window Actually Holds
Three AI platforms advertise a context window of about one million tokens. Read the model pages and the number means three different things, with output caps from 65,536 to a full million.
October 30, 2026The Price Changes Partway Through the Prompt
Two of the three large AI platforms charge a higher per-token rate once a prompt crosses a published size. One charges five times more. The thresholds are 100,000 and 272,000 tokens.
October 29, 2026The Monthly Cap on a New API Account
Every AI platform puts a ceiling on how much a new account can spend in a month. Three vendors publish three different ceilings, and two of them stop at exactly the same number.
October 28, 2026The Trial That Needs a Sales Call
Six AI pricing pages read in one sitting. One of them has no dollar sign on it at all, and four others stop printing numbers at exactly the tier where the governance features start.
October 26, 2026The Add-On That Is Not in the Seat Price
Three unrelated vendors price extra AI usage at exactly one cent a credit. What differs by a factor of eighty is how many credits the seat you already pay for includes.
October 25, 2026The Introductory Price Is Not the Price
Four AI price lists carry an expiry date in the small print: a twelve month discount, a twenty seat cap, a dated rate rise and a discount that only exists on annual billing.
October 18, 2026Microsoft Renamed Copilot and the Licences Moved
The Microsoft 365 Copilot app is now Microsoft Copilot, the add-on is now Microsoft Copilot Business, and the admin pages use a third vocabulary. Here is what each name buys.
October 17, 2026The Model Shutdown Calendar
Three vendors publish retirement dates for the models they sell, with three different notice policies. The nearest date on the list is November 30, 2026. Here is the whole calendar.
October 16, 2026What the AI Vendors Say They Keep
Thirty days appears on all three big AI retention pages and means something different each time. The longest published clock runs to seven years. Here is what each vendor actually commits to.
October 15, 2026The Fifty Percent Discount You Have to Ask For
OpenAI, Anthropic and Google all cut model prices in half for work that can wait a day. The discount is identical; the envelope sizes, expiry rules and result windows are not.
October 14, 2026The Cached-Input Price Is the One That Matters
Two of the three big AI platforms publish a separate rate for input the model has already seen. It ranges from 2.5 percent to 50 percent of the headline price. Here is how to read it.
October 13, 2026Three Things in AI: Connected Apps, Secret Checks, Markdown
Gemini connects to Airtable, Linear, monday.com and PandaDoc from inside the chat, GitHub ships a purpose-built model for spotting leaked passwords and puts the new checks on an AI Credit meter, and Google Docs edits Markdown files without converting them. What shipped, what it costs, who it changes the week for.
October 10, 2026Three Things in AI: Haiku 5.5 Pricing and Two Free Upgrades
Anthropic prices Haiku 5.5 at ten cents per million input tokens, GitHub makes a local sandbox for Copilot's agent standard at no extra cost, and Google's Workspace Studio gains nested conditions. What shipped, what it costs, who it changes the week for.
October 10, 2026Three Things in AI: Who Pays When the Agent Does the Work
GitHub lets organisations pick up the bill for Copilot code review, Microsoft's Sales Development Agent enters public preview and starts consuming Copilot credits on November 16, and Anthropic's updated usage policy requires a human reviewer for high-stakes uses. What shipped, what it costs, who it changes the week for.
October 10, 2026What Gets Into This Brief, and What Gets Cut
Issue zero. Three slots every morning, one primary-source rule, four kinds of story that never make it, and the words you will never read here. Read it once and you will know exactly what this brief is promising.
October 7, 2026Three Things in AI for Wednesday, October 7, 2026
Anthropic opens three verified tiers for security work, GitHub makes stacked pull requests standard on every plan, and Slack turns Later into To-dos. What shipped, what it costs, and who it changes the week for.
October 7, 2026