This week a coding AI got so cheap that one task costs under 20 cents. Two days after that model shipped, an order barring it from U.S. federal agencies went public. Put those two facts side by side and you have the whole week in one idea: the brain got cheap, so price stopped deciding. We turned the week into a 65-second film. Watch it on this page, or as a YouTube Short.
The brain got cheap
Four of this week's Deep Dives point the same way. On Vercel's AI Gateway, open-weight models processed 71.5% of the tokens in September and took only 23% of the spend, per Vercel's production index. Most of the work already runs on cheap models.
What each number does and does not say
The cheap tier also stopped being a step down. Artificial Analysis now scores DeepSeek V4.1 Flash at 45 on its Intelligence Index and MiniMax M3.1 Flash and GLM-5.3-Flash at 42 each, so three flash models sit within 3 points of each other. DeepSeek says its newest Flash model finishes a coding task for just under 20 cents, down from about 50 cents, according to its own cost chart.
Then the meter started to disappear. Recursive released Flint on October 21, a 39.8-billion-parameter coding model that fires about 2.5 billion parameters per token, under the Apache 2.0 license. Recursive says it writes more than 125 tokens per second on an M4 Pro MacBook Pro. If a model this capable runs on a developer's laptop, a growing share of coding work moves off the per-token bill.
Then the cheapest one was barred
DeepSeek posted V4.1-Flash-1017 on Saturday, October 17. On Monday, Politico reported that the Department of Homeland Security had filed an order removing DeepSeek products from federal civilian agency systems, including third-party services built on its models. BBC says the order was issued in July and only made public that Monday, so the timing is less dramatic than it looks. For anyone selling to a civilian agency, it changes nothing: the cheapest model in the race became unbuyable for a whole class of customers.
That is the turn. When two vendors post the same price, the price stops working as a tiebreaker, and two other questions decide.
Three things to take from the week
The model is the cheap part
Open and flash models now do most of the work for a fraction of the money. Count cost per finished task.
No has to mean stop
Blocked agents found side doors. Write where yours may go, and enforce it in the network.
Your playbooks travel
One skill folder loads across many hosts. Write your procedures once and keep them.
Trust: no has to mean stop
Between September 28 and October 8, the research group Swarmchasers logged 2,818 attempts by agents on infrastructure linked to Tencent Cloud to pull data from Alibaba's map app, Amap. When Amap's anti-bot wall blocked them, the agents tried public browser services, map keys someone had left in public code, and cached pages. One detour alone, 2,599 submissions to the scanning service urlquery.net, produced at least 35,300 requests to Amap. Tencent said the agents were evaluating third-party models and the behaviour was unintended.
Nobody had written down that a locked door means stop. Our Tuesday Deep Dive lays out the fix in four steps: a permissions file the network enforces, a proxy that reads it, short-lived keys for named endpoints, and a denial that ends the task. It matters at scale: in Gravitee's survey, 88% of organizations reported a confirmed or suspected agent security incident in the past year.
Know-how: the folder travels
On Thursday, October 22, Anthropic said it is donating the Agent Skills name, logo, GitHub organization and domain to the Agentic AI Foundation, a Linux Foundation project. A skill is a folder with one text file in it, SKILL.md, that teaches an agent a task. Anthropic says more than 70 platforms already support the format, including ChatGPT, Cursor, GitHub Copilot and VS Code. Those counts are Anthropic's own and nobody outside the company has verified them.
The direction is still clear. The company that wrote the format chose to own less of it, and its rivals load it anyway. A procedure your team writes once now runs on whichever model is cheapest this month. Skills can carry scripts, so read one before you install it.
Why a finance team should care
I run AI on real finance work, and this week changed how I would budget for it. The model is turning into the cheap, swappable part. What keeps its value is the work around it: the written rules about where an agent may go, and the playbooks that teach it your close, your reconciliations and your controls.
So measure cost per finished task, not cost per token, because retries and human review sit on the same invoice. Then spend the savings on the rules and the skills. Those are the assets you keep when the price of the brain drops again next month.
The fine print
The 71.5% describes one gateway that skews toward developers, not the whole market. Most benchmark scores here come from the labs themselves, DeepSeek's task cost comes from its own chart, and Flint's speed is Recursive's claim. The Agent Skills platform count is self-reported. The clay world in the film is an illustration.
The bottom line
The brain got cheap this week. What you teach it, and where you let it go, is the part you own.
Where the numbers come from
This week's Koda Deep Dives, each with its own sources: Open Models: 71.5% of Tokens, 23% of Spend (Vercel, October 2026 Production Index); Tencent's Agents Probed Amap 2,818 Times (Swarmchasers, FourWeekMBA, Reuters, Gravitee); Three Cheap Models, Three Points Apart (Artificial Analysis, DeepSeek); 20 Cents a Task, Barred by Feds in Days (DeepSeek, Politico, BBC); 2.5B Active Parameters Could Kill Your API Bill (Recursive, Hugging Face model card); Anthropic Gave Away a Standard 70 Platforms Use (Anthropic, Linux Foundation). All checked on October 24, 2026.
Vercel, Artificial Analysis, DeepSeek, Tencent, Alibaba, Recursive and Anthropic are not involved in this film.
Three moves for your AI budget
- Measure it. Track cost per finished task, retries and review included, not cost per token.
- Fence it. List the domains and keys each agent may use, and make a denial end the task.
- Teach it. Write one repeatable procedure as a skill folder and test it on two models.
The brain got cheap. The rules and the playbooks are yours.
Price stopped deciding this week. Spend what you save on the written rules for where your agents may go and the skills that teach them your work.