AI Weekly, Oct 26: skills go open, Kimi nears the frontier

1. Mollick looks back over 30 months: capability up exponentially, cost down sharply

Ethan Mollick, an associate professor at Wharton, wrote on October 20, on the day his new book Co-Existence came out, a look back at the 30 months since his previous book in early 2024. On METR's benchmark, which measures the amount of human work an AI can complete from a single prompt with at least a 50% chance of success, he says the figure was a four-minute task in 2024 and 17 hours of human work when he turned in his draft in April 2026. He adds that METR has said its current methods can no longer measure this accurately. On cost, he says that getting one correct answer on GPQA, an exam of problems that stump graduate students, used to cost about 30 cents in AI usage and now costs four hundredths of a cent. He also writes that 180 models are now better than GPT-4 was at its release, and 19 of them can run on a modern smartphone (original article).

He says the way people work with AI has changed too. Instead of guiding a chatbot step by step, a person now summons agents when needed, assigns them work and evaluates what comes back, a mode he calls the conjurer. The charts in the article were made that way: a couple of sentences of instruction, and the AI finished the research, graphing, layout and citations in minutes. He also names a new risk, which is that people lose track of what exactly the AI is doing to produce the results (same article).

Our view If costs keep falling, the dividing line moves from unit price to whether a company can put the work it wants delegated into words. Take quotations, first replies to inquiries and product description drafts: if a competitor hands such routine tasks, which take people hours, to AI first and takes more orders with the same staff, we think the gap will show up in lead times and prices within about six months. A company that starts late may lose jobs while customers wait for its quotes. Begin by writing down, for one task, the inputs, the procedure and the criteria for an acceptable result. On the risk Mollick names, of losing track of what the AI is doing, check before adoption whether the setup keeps a record of the work as well as the result.

2. Agent Skills moves to a Linux Foundation foundation

On October 22, Anthropic announced that it is donating Agent Skills, the open standard that teaches AI agents specific procedures and domain knowledge, to the Agentic AI Foundation (AAIF) under the Linux Foundation. A skill is a folder containing a SKILL.md file, bundling a procedure with any scripts or reference material it needs. The company says that since the standard was published in December 2025, more than 70 agent products, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated it, 1 million skills have been published and nearly 280 million installed. The name, logo, GitHub organization and the agentskills.io domain move to AAIF, and Anthropic has proposed a roadmap covering versioned releases, a formal proposal process and security guidance for people who write and install skills (Anthropic announcement).

The same day, the Linux Foundation announced that AAIF had reached 305 members in less than a year since its launch. New top-tier members include Ant Group, Royal Bank of Canada and Vercel. Executive Director Mazin Gilbert said the member organizations are asking how agents authenticate, how their actions are verified, and who is accountable when one acts on their behalf (Linux Foundation announcement).

Our view For a small business the meaning is simple: once a work procedure is written down as something like a SKILL.md, it becomes more likely that you can change the agent you use without rewriting the procedure. A company whose procedures live only in one person's head, or in the chat settings of one tool, repeats the same write-up every time that person leaves or the tool changes. However, we think installing skills published by others as they are should wait until the security guidance exists, and that a business should begin with procedures it wrote itself. We will decide on using outside skills after checking whether the first versioned release includes that security guidance.

3. Kimi K3.1 Pro: a near-frontier open-weight model at $2 and $10

On October 23, China's Moonshot AI released Kimi K3.1 Pro through its API and other channels. It has a 1-million-token context window, and the company says full model weights will be released by November 4. The company reports 65.7% on Terminal-Bench 4.0, which in its own table is just behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. These are the company's own measurements (Moonshot AI announcement).

In a blog post the same day, Simon Willison notes that the API price is $2 per million input tokens and $10 per million output tokens, the same as GPT-6.1 Sol and half the price of Claude Opus 5.5. On the independent Artificial Analysis index, he reports Kimi at 55, ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (all at 53) and behind only Opus 5.5 (58), and writes that this is the first time a model with weights due to be released has scored this high against the frontier (original post). Artificial Analysis's own page also shows that its output speed is about 58 tokens per second, which it calls slow, that it generated 150M tokens in the index run against a median of 87M, which it calls verbose, and that the average cost per evaluation task is $2.11 (Artificial Analysis).

Our view Once the weights of a high-performing model are released, businesses gain bargaining material on price and, later, an option to run it on their own servers. But switching is not what a small business should do this week. A slow model with long outputs costs more in billing and waiting time even at the same unit price. And if customer lists or transaction records will be sent to it, the provider's terms on where data is stored and whether it is used for training must be read before signing. We think the sensible course is to wait until the weights are actually released after November 4, then compare cost and accuracy against other models on your own real documents before deciding.

4. Linux Foundation survey: AI halves some work time, but low-quality submissions rise

On October 20, the Linux Foundation published a research report on how generative AI is changing open source development. 59% of project maintainers say contributions have grown since generative AI spread, rising to 69% in large projects. At the same time, an estimated 25% of incoming pull requests, and 31% in large projects, are low-quality AI slop. Respondents report productivity gains of up to 50%, and code authoring time falls by 52% on average. For high-complexity tasks, however, 21% of respondents say generative AI made the work take longer. 53% of maintainers disagree that generative AI has improved community trust (Linux Foundation announcement).

Our view The survey's numbers come from open source, but the pattern applies to ordinary companies: when the people producing work get faster, work piles up for the people checking it. If an outside production firm or a trading partner delivers AI-made output such as site changes, quotations or translations, and the receiving side has no acceptance criteria and no one assigned to check, slow checking becomes slow operations. We think a business should set 3 to 5 acceptance criteria for deliverables before adopting AI. If in the first month any task takes more time to check than AI saved, it is reasonable to take that task out of scope.

Our outlook: companies that keep procedures and prices in machine-readable form get chosen first

Taken together, this week's announcements show the parts for agents to act on behalf of organizations starting to fall into place: a standard for handing over procedures (Agent Skills), a neutral foundation that manages it (305 member organizations, including a bank and payments-related companies), and models priced close to the top tier. The three questions AAIF's executive director named, authentication, verification and accountability, bear directly on business-to-business dealings. We expect that within 6 to 24 months, personal AI agents will ask company-side AI agents directly for quotes and stock and carry a deal from comparison to order without a person involved. This is our own projection. The companies that get onto the shortlist first will be those that publish prices, stock, lead times and scope of service in a form machines can read, and that have their inquiry procedures written down. A company whose procedures exist only in people's heads may be unable to answer an AI's inquiry and be dropped at the comparison stage. We think now is the time to lay the groundwork: write down the procedure for one frequently repeated task, and organize your product and price information.

Our overall view

The week of October 19 to 25 shows AI costs continuing to fall while the gap in preparation between those who use it well and those who do not widens. The drivers are the standardization of a portable format for procedures, and price competition from high-performing models whose weights are due out. For the next 2 to 4 weeks, we suggest putting priority on picking one recurring task and writing down its procedure, acceptance criteria and the scope of data sent out, rather than switching to a new model. Whether to widen the cost comparison depends on whether Kimi K3.1 Pro's weights are actually released on November 4. If the release slips, it is fine to carry on with the models you already compare. We will keep following the content of AAIF's security guidance as a condition for deciding how to treat outside skills.

Questions this article answers

Sources

1. Ethan Mollick, 30 Months on the Exponential (in Charts), One Useful Thing (2026-10-20)oneusefulthing.org
2. Anthropic, Donating Agent Skills to the Agentic AI Foundation (2026-10-22)anthropic.com
3. Linux Foundation, Agentic AI Foundation Surpasses 300 Members in Less Than One Year (2026-10-22)linuxfoundation.org
4. Moonshot AI, Kimi K3.1 Pro Tech Blog (2026-10-23)kimi.ai
5. Simon Willison, Introducing Kimi K3.1 Pro (2026-10-23)simonwillison.net
6. Artificial Analysis, Kimi K3.1 Pro model pageartificialanalysis.ai
7. Linux Foundation, New LF Research Survey on generative AI and open source (2026-10-20)linuxfoundation.org

Related Q&A

• Will AI leak our internal documents?
• How much does AI cost a small business?

Related services

• AI office assistant
• Custom development
• Pricing

• PHP2Shell: AI-found PHP flaws and what small sites must do
• AI Weekly, Oct 19: usage gap widens as model prices fall