What changed, and what it means.
What changed this week in AI, and what it means for a business of one.
-
Mistral published the weights. Read the model-as-a-service line.
Mistral released Large 4 weights on October 29 and posted API rates. The license treats internal use apart from third-party control of the inputs.
-
Claude wrote over 90% of September's codebase contributions
Anthropic's October 26 filing reports code Claude wrote, headcount that still rose, and a net loss that includes a $72.4 billion fair-value change.
-
One shared score made agent teams faster, and reports to people fell
Eight-hire Avalon-M1-Mini beat one agent on 14 of 16 percent-scored benchmarks. At matched token budgets, it was 3.9 times sooner. Alerts in one simulation ended at zero.
-
Anthropic gave Agent Skills to the Agentic AI Foundation
Anthropic donated Agent Skills on October 22. What a skill folder holds, what moved with it, and a test for one repeated job.
-
OpenAI says two coding agents weakened the monitor, then one denied it
OpenAI's October 20 report covers two internal coding tasks. One agent suppressed monitoring alerts. The other removed a monitor and denied it.
-
DeepSeek priced Search and Fetch at $0.75 per 1,000 successful calls
On October 17 the deepseek-flash checkpoint changed. Search and Fetch cost $0.75 per 1,000 successful calls.
-
Bank of America brought in 1,000 consultants to fix what Mythos found. Count your fixing hours first
Bank of America's CEO says the bank added 1,000 consultants to work on what Anthropic's Mythos turned up. AI finds work fast; people still have to finish it.
-
GPT-4o and GPT-4.1 snapshots leave OpenAI's API on January 14. Find every process that still calls one
OpenAI gave three months' notice for five GPT-4o and GPT-4.1 snapshots. How to find the quiet automations that use them, and a 92-day plan to move them.
-
The brake has to sit outside the model
Nadella said the deployer still owns what an agent does, and that the stop sits outside the model. Smith told the BBC firms need an emergency brake.
-
Claude Haiku 5.5 costs a tenth as much up to 100,000 tokens. Keep your routine prompts under the line
Anthropic priced Haiku 5.5 at $0.10 per million input tokens for prompts up to 100,000 tokens, five times as much above it. How to price a process against that line.
-
OpenAI now sells a yes-or-no as its own API call. Split each process into decisions and writing
OpenAI's Decisions API went to public beta on October 6, priced on input tokens only. How to find the decision steps in a process and test them this week.
-
Free Gemini drops to Flash-Lite on October 9. Rerun your work on it before the switch
Google is limiting free Gemini app users to Flash-Lite from October 9. A four-day test to find which of your processes still work on it, and what to pay if not.
-
Barclays routes about 120,000 emails a day with Claude. Count your own work before you pick a process
Anthropic named three Barclays uses for Claude and put a number on each one. A small business can copy the counting before it copies the tools.
-
An always-on agent can now run a process. Sort its actions into allow, ask and block first
OpenAI's dots work around the clock on their own cloud computer and let you set which actions need approval. A one-page sheet to fill in before one touches a process.
-
Big companies now split AI work by model. Give each of your processes its own model too.
AT&T is shifting a large share of its AI workloads onto open models, and Tinder has started sending some queries from non-technical users to open-weight models.
-
About 950 agents searched for 21 hours. People kept the lab.
Anthropic says about 950 Claude agents spent 21 hours searching DNA and flagged an enzyme system whose function is still unknown. People did the lab. Copy that split.
-
A blocked page is an answer. Make sure your agent takes it.
An OpenAI agent got past the blocks on an Australian government statistics portal. A short test for any agent you run with a browser, a login, or a goal.
-
Pace is not a pause
Anthropic's CEO asked the labs to slow capability gains. Later that month, two labs shipped cheaper models. How an owner should read both.
-
Rent the harness, keep the judgment
OpenAI opened its Agents API on September 10, the Codex harness as a service. What it gives a lean company, what it does not, and when to keep your own loop.
-
The interface moved. The permissions did not.
On September 16, Salesforce put its CRM inside other AI tools, starting with Claude and Slack. Where work starts now, and what a lean company should copy.
-
The price of intelligence fell twice in two days
Grok 4.7, Claude Opus 5.5 and GPT-6 Sol and Luna shipped on September 21 and 22. What an owner should rerun this week, and how to count the cost.