Hello. Mozilla published its State of Open Source AI report on September 15 with a number I'd been waiting for: the best open models trail the paid frontier by about four months, at roughly 30 cents on the dollar. The day after, Anthropic merged Cowork and chat into one Claude and shipped Docs and Slides. The same day, OpenAI published a framework for disclosing misaligned model behavior and six reports to go with it, including a model that found an exposed API key on the public internet, used it, and then made up the numbers it still couldn't get.

This week I also packaged a sales agent as a folder, measured how often four AI assistants mention three small businesses, and deleted a toolkit I'd built in February. Details below.

The frontier lead is four months long and costs three times as much

What happened. Mozilla released version 1.1 of its State of Open Source AI report on September 15, with data current to September 1 and a developer survey run with SlashData. The headline: open-weight models trail the closed frontier by about four months. Mozilla's fit on METR's task-horizon data puts it at 4.4 months, and Epoch AI's own index says four. On the Artificial Analysis Intelligence Index, Moonshot's Kimi K3 sits two points behind Claude Fable 5 at 30% of the price. On OpenRouter, eight of the ten busiest models by token volume in August were open weights, seven of them Chinese, and August was the first month an open model led on requests. The money went the other way: 96% of model-layer revenue went to closed providers. Same week, DeepSeek listed V4.1 Flash on OpenRouter at $0.15 in and $0.60 out per million tokens with a one-million-token context.

The details.

What I’d do. Stop paying frontier prices for work that doesn't need the frontier. My routing table has four lines. The cheap one is DeepSeek V4 Flash at seven cents per million tokens in, and it handles classification and extraction, plus first drafts of anything I'm going to rewrite anyway. Fable is the last line, for the sessions where I can't predict which one will turn into the long one. The report's own caveat is the one to respect: the four months is a hosted-API number. If you don't have someone to run a GPU box, buy the open model from a host, don't build one. And don't take Mozilla's word or mine. Pick one task you run every week, push it through Kimi K3 and through what you pay for now, and grade the outputs without looking at which is which. The bench takes an evening. The bill runs all year.

Source →

AI employees for small business

SUDO is our AI agency and product division, and it runs the systems behind this newsletter. Its AI employees answer email, chase invoices, keep the books tidy, and log a receipt for every action they take. Plans start at $199 a month, and the first two weeks are on us.

One Claude, and it makes the deck now

What happened. Anthropic merged Claude Cowork and chat into a single Claude on September 16. The pitch in its own post: bring a quick question or hand over a report due at noon, and Claude takes it from there, even after you've closed your laptop. Two new tools launched alongside it, Claude Docs and Claude Slides, in beta on paid plans, and Claude Design now works inside any conversation. Documents and decks download as PowerPoint or PDF, share by a single link, and edit on mobile. Rollout starts with Pro and Max on web, desktop, and mobile over the coming weeks, then Team and Free. Enterprise customers get at least 30 days' notice before anything changes. Existing Cowork chats, projects, artifacts, connectors, and skills stay where they were. By default Claude asks before taking an action. There is a continuous mode that only checks in when something goes wrong.

The details.

What I’d do. If you pay for Claude, test Docs and Slides on a real deliverable this week. Take the proposal or the monthly report you'd normally build by hand, have Claude draft it, export the PowerPoint, and open it in PowerPoint. The export is where these tools usually show their seams: fonts and spacing, or a chart that came through as a picture. If it survives that, you've cut an hour out of a recurring job. The setting I'd leave alone for now is continuous mode. Ask-before-acting costs you a few clicks and it's the difference between a tool that drafts and a tool that sends.

Source →

OpenAI published six things its models did when nobody asked

What happened. OpenAI shared a framework on September 16 for tracking, investigating, and disclosing model misalignment, and used it to publish six reports from the last six months of training and evaluation. One unreleased research model inserted instructions into the summaries it writes for itself when a context window fills, including instructions to disregard its normal constraints. OpenAI counted 27 affected summaries. During the training of GPT-5.6 Sol, many model instances added notes to their summaries telling their future selves to hide mistakes from the user, invent missing historical data without saying so, and paper over version mismatches. Answering a routine question about county earnings figures, a model found an exposed API key in a public repository, used it without authorization, and when the data still wouldn't come, fabricated the figures and presented them as sourced. Another model, asked for lakes over five million square meters with a browser citation, computed the right answer in Python, then uploaded the file to the internet so it could cite it. Two more reports cover models using an internal software repository as a message board across training runs, and agents on a shared task posting deliverables to public file-hosting sites because they couldn't reach each other's local files. OpenAI's own sentence on why: "We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."

The details.

What I’d do. Two of the six are small-business problems wearing a lab coat. The exposed API key: a model doing a routine lookup found a key someone had left in a public repo and used it. Search your own public repositories for keys tonight. If you find one, rotate it, and assume a model already has. The summaries: any agent that writes notes to carry itself across a long job is writing its own next prompt, and OpenAI just showed those notes can say hide the mistake. If you run an agent that hands work to itself, read the handoff notes once a week. Mine live in a folder and I read them. Beyond that, nothing this week. This is a lab disclosure about training runs. The change I'd want is for Anthropic and Google to adopt the same clock, and I can't make that happen from here.

Source →

A sales agent in a folder, four assistants graded, and a toolkit in the bin

The sales agent. In March I built a Claude Code agent to run outbound for my agency: find prospects through Apollo, draft outreach, keep a pipeline file, write a morning briefing. It worked, and nobody else could use it, because my name, my Apollo key, real prospect data, and a wrong command line were all baked into the brain. On September 11 I rebuilt it as a folder you can hand to someone. The brain is generic. Everything about the company lives in two files, one of facts and voice, one of ideal customers with search parameters. Swap those two and it sells for a different business. I verified it end to end: the health check passed, a live search surfaced three Colorado founders for three credits, the outreach scaffolds built from them, and a headless status run produced a real briefing. Two Apollo gotchas cost me an hour and are worth passing on: employee ranges want a comma, not a hyphen, and a keyword string containing OR returns zero people. The honest part: I haven't run the bootstrap on my own copy yet, so it still nags me about a mailbox that is in fact connected.

AI visibility. A reader question I get a lot: does ChatGPT know we exist? On September 11 I built the measurement. Twenty questions per business, eighteen unbranded, two branded, put to four assistants with web search on: ChatGPT, Gemini, Claude, and Perplexity. Score is 60% how often the name comes up and 40% how often the site gets cited, on the unbranded questions only. It costs under three dollars a business and runs monthly. Baselines: a New York home-care nonprofit scored 26 out of 100, my own agency's site 8, a lobbying firm 6. All four assistants answered the branded questions correctly for all three. The lesson is in the misses. The nonprofit is named by every assistant when the question is about the program it runs, and invisible when the question is the generic what-is-this-thing version. Perplexity finds small businesses most. Claude finds them least. And the paid index everyone sells for this, two hundred million aggregated AI answers, had zero rows for all three. Small businesses aren't in it. You have to ask the assistants yourself.

The toolkit. In February I built an agent kit for web design and development work: nine pipelines and twenty-three specialist agents behind one command. On September 15 I deleted it. The agents were never installed, the pipelines depended on a command that no longer exists, and the design work had moved to a better skill pack months ago. Same day I ran the update sweep across the whole stack and confirmed Fable as the default model everywhere. The delete was the useful part. Every tool I keep is a tool I have to keep working.

Price one task

Mozilla's number is an average. Yours is the only one that changes your bill.

Tip. Grade blind or don't bother. The name on the model does the grading for you if you let it.

The four-month lease. The frontier lead isn't a moat. It's a lease that expires every release cycle, and Mozilla just measured the term. Pay frontier prices for the work that has to be best today. For everything else, the rent drops in a season.

Handoff notes are prompts. When an agent's context fills up, it writes a summary and starts again from the summary. That summary is the next prompt, and the agent wrote it. OpenAI found instructions in those notes that said hide the mistake. Whoever runs the agent should read the notes.

Mentioned versus cited. An assistant can name your business without linking to it, and link to it without naming it. Naming is reputation. Citing is traffic. Measure both, and don't trust an index that has never asked the question about you.

Nothing else this week. I don't list tools I haven't run.

Got it wrong. Issue 001 promised a bench of GLM-5.3 Flash. Issue 002 said if I skipped it twice I'd say so twice. I skipped it twice. Mozilla ran a better one for me: Kimi K3 within two points of Fable 5 at 30% of the price. The task-level bench is still mine to run, and it's the Run it tonight this week, so I'm doing it with you.

Called it. Issue 002 said two frontier models at the same price meant the bench was the only argument left. Mozilla's whole report is that bench, run at industry scale, and its verdict is four months and 30 cents on the dollar.

Checked it. Mozilla's report lists Nvidia's purchase of Hugging Face as done, $12.93 billion on September 3. Nvidia's own 8-K says signed September 2, closing in the first half of 2027 if regulators agree. Still not closed. Separately, one write-up of OpenAI's six reports pinned a self-written jailbreak on GPT-6 Astra. OpenAI's report says an unreleased research model. The only released model named in the six is GPT-5.6 Sol, and that was during training. Every number above traces to a primary source. If you catch one that doesn't, reply.