The claim: AI makes everything cheaper to produce, so more of it is always better.
The counter-fact: Google just stopped paying for open-source bug reports because the inbox filled with machine-written ones.
Picture a maintainer opening their security inbox on a Monday and finding a few hundred “critical” vulnerabilities, each confident, each formatted perfectly, nearly all of them invented. That is the week’s real story, and it sits underneath the model launches everyone else is covering. Five things crossed my desk. Here is the operator’s read, sorted by whether it matters, whether it is marketing, or whether it breaks something you already run.
Google Closes Its Open-Source Bug Bounty to New Reports
Matters. On October 1, Google stopped accepting new submissions to its Open Source Software Vulnerability Rewards Program. BleepingComputer’s report quotes Google directly:
This pause is due to a significant rise in automated submissions, the vast majority of which are not valid.
Supply chain reports, the Patch Rewards Program (up to $15,000) and the Cloud program for Google Cloud open-source repositories stay open. Google says it will share reforms in the first quarter of 2027. The same piece notes curl dropped its HackerOne bounty in January and Intel pulled cash rewards from its Intigriti program in September.
Think about the incentive. A bounty pays per accepted report, and a model can draft a thousand reports for the price of one coffee. The reward stays fixed while the cost of trying drops to zero, so the system floods. Nobody designed it to fail. It just never priced in a participant that does not get tired. If you run any intake that pays or ranks by volume (support forms, lead forms, job applications), you have the same exposure, and the fix is a cost to submit, not a better filter.
Gemini’s Free Tier Drops to One Model on October 9
Marketing, with a real change inside it. Per 9to5Google’s breakdown, free users get Flash-Lite only starting October 9. AI Plus ($4.99 a month) keeps Flash-Lite and Flash but loses Pro. AI Pro ($19.99) gets all models plus a new Deep Think mode, and AI Ultra runs $99.99 to $199.99.
The announcement leads with effort levels and Deep Think. The actual news is the squeeze: if your team prototypes on the free Gemini app, the answers you got last month came from a bigger model than the one you will get next week. Google also has not said where its new Gemini 4 Argon fits in the ladder, which tells you the ladder is still being decided.
OpenAI Dots Run All Day, and the Price Starts at $100
Matters, for anyone building agents. OpenAI’s DevDay brought GPT-6.1 Sol and Dots, always-on agents that keep working without a prompt. AlternativeTo’s rundown says Dots run 24/7, operate their own browser and connect to more than 4,000 services. It also reports OpenAI’s own claims for Sol: close to GPT-6 Astra at about one-fifth the token cost, and factual errors down from 11.4% to 7.7%. Those are vendor numbers, so I would test them on your own data before believing them.
Implicator’s coverage has the access details: Dots arrive on Pro plans from $100 a month and Business Premium at $125 per seat, and they are not available to free or $8 Go users or, for now, to subscribers in the EEA, Switzerland and the UK.
Breaks your stack: an agent that acts on its own browser at 3 a.m. needs permissions, logs and a kill switch before it needs a good prompt. I would not point one at anything that sends email or moves money until those exist.
OpenAI Raised the Ceiling and Trimmed the Floor
Marketing, and a budget problem. Implicator also reports a new $500 Pro 500 tier from October 30, with Ultrafast mode of up to 300 tokens a second in Codex. The same date cuts the $200 Pro plan: Work and Codex usage falls from 20x to 10x the Plus allowance, and GPT-6 Pro chat goes from 200 to 100 weekly messages. Existing Pro subscribers get a one-time $2,500 usage credit that expires at year end.
Read that as a pricing ladder doing its job. The $500 tier makes the $200 tier look like the compromise, and the compromise just got smaller. If your workflow leans on the old allowance, October 29 is your deadline to measure how much of it you really use.
Claude Sonnet 5.5 Holds the Price and Shrinks the Task Bill
Matters, quietly. Anthropic released Sonnet 5.5 on September 28. Unite.AI’s writeup confirms the list price is unchanged at $2 per million input tokens and $10 per million output tokens, and quotes Anthropic saying output is more than 30% faster at up to 30% less per task in its testing. The saving comes from fewer tool calls and shorter runs, not a cheaper rate card. One early user, Base44, reported 3.6 iterations per build against 7.7 on Opus 5 across 118 apps.
Notice that OpenAI’s Sol lists at the same $2 and $10. When two frontier-tier models share a rate card, the rate card stops being the comparison. Cost per finished task is. I run my own content engine on this logic: I track how many attempts a job takes, because a cheap model that needs four passes costs more than a pricier one that needs one.
What I Would Change This Week
Three moves, all small. Put a per-task cost number next to your per-token one. Write down which of your tools quietly depend on a free tier before October 9. And if you accept anything from the public, add friction that costs the sender something.
If you want this read in your inbox, the Scoy newsletter is where I send it.