Google cut image prices roughly in half with Nano Banana 2.1 on Tuesday, and nearly every writeup stopped at that sentence. Google’s own price sheet says 4K images fell by a quarter.
The engine that publishes this roundup makes a hero image every weekday on Google’s models, so I read the sheet instead of the coverage. The rest of the day had the same shape: the headline number is true, and your bill sits in the line under it.
Matters: Nano Banana 2.1 and OpenAI’s Decisions API
Nano Banana 2.1. Google shipped it on October 6 with room for up to 14 reference images and a lower price at every size.
At 1K and 2K, “half” holds. On Google’s Gemini API pricing page, a 1K image drops from $0.067 to $0.0336 and a 2K image from $0.101 to $0.0504.
At 4K the sheet says $0.113, down from $0.151. That is a 25% cut, and the $0.0756 figure going around is almost exactly what Nano Banana 2 already charged for 4K on the batch tier.
Run the division and you can see why. The image token rate halved from $60 to $30 per million, but a 4K image now bills about 3,770 output tokens against roughly 2,500 on the old model.
My verdict: matters. If you generate at 1K or 2K, your image bill just halved. If you generate native 4K, budget for three quarters of today’s spend, or move anything that can wait onto batch, where 2.1’s 4K price is $0.0567.
OpenAI’s Decisions API. You hand it a question, the answers you allow, and some text or an image. It hands back a probability, a choice or a score.
OpenAI’s documentation says it answers about 10 times faster than GPT-6 Luna through the Responses API, at $0.10 per million input tokens with no output charge.
The price is not the news, since Luna already bills $0.10 per million input tokens through Responses. Caching is the catch.
An OpenAI staffer confirmed in the public beta thread that Decisions has no input caching. Cached Luna input on Responses costs $0.01 per million.
Say your router sends a 2,000-token rubric with every 200-token ticket. With the rubric cached on Responses, my math puts a million calls at about $42.50. On Decisions, the same million calls cost $220.
My verdict: matters, for the right job. Use it where speed is the product, like routing a live chat. Keep a batch classifier with a long rubric on Responses until caching lands.
Breaks Your Stack: Anthropic and Mistral Put Cyber Work Behind a Gate
Picture a one-person security shop running authorized pen tests on Claude. On Tuesday, two labs told that person exactly where the line is.
Anthropic: prove who you are.
Anthropic folded Project Glasswing into a three-tier Cyber Verification Program. Its general models, Opus 5.5, Fable 5.1 and Sonnet 5.5, “block most cyber work.”
Defense Access takes a few days and admits individual researchers with a disclosure record. Red Team Access takes a few weeks and admits no individuals at all.
Mistral: wait for the weights.
Mistral opened a public preview of Large 4: 1 trillion parameters, 49 billion active, $1.36 per million input tokens and $4.18 output.
Weights land “by the end of the month.” Until then, vetted partners and state authorities get “reduced moderation and expanded cyber capabilities.”
So the solo pen tester is shut out of Claude red-teaming by design. Defensive work still runs: Anthropic says its general models still handle code review, patching known issues and vulnerability finding in code you own.
Mistral’s door opens at month end, but a trillion parameters needs roughly a terabyte of memory just to hold the weights at 8-bit precision. That is nobody’s one-person server, and until the download exists, “open-weight” on Large 4 is a promise.
I think Anthropic’s gate is the right call. Its Glasswing partners found at least 129,000 verified vulnerabilities between April and July, and more than 33,000 of the verified bugs so far are rated critical or high.
This is the same class of model that wrote a working exploit in three hours last month. If you run security work on Claude, apply for Defense Access this week. If you need red-team capability, you need a firm behind you and a few weeks of lead time.
Marketing: HubSpot’s Memo and the Personal Agent Protocol
HubSpot. CEO Yamini Rangan told nearly 660 people, about 7% of staff, that their roles are going, in a memo HubSpot filed with the SEC on October 6. The 8-K itself puts the cost at $65 million to $75 million, mostly cash for severance and transition payments.
The memo says the cuts are “not driven by AI-related efficiencies.” The same memo reorganizes product teams around customer outcomes instead of Hubs, with HubSpot “delivering outcomes for them with AI.”
My verdict: marketing. Both sentences can be true at once, and the first is still written for customers and investors as much as for the people leaving.
If you run your business on HubSpot, the Hubs line is the one to watch: the teams behind the tools you learned are being reorganized around outcomes.
Personal Agent Protocol. Sierra and Meta announced an open standard for how personal AI agents sign in to businesses, with Walmart, Shopify and Stripe among the partners.
Sierra’s announcement lets the customer choose read-only or write access, and lists limits on specific actions as a later idea. The v0.1 spec is due later this month, without payments.
Stripe and Shopify also sit in Visa’s rival Trusted Agent Protocol, as The Next Web pointed out.
My verdict: marketing until there is a spec to read. Cloudflare already showed how you bill an agent per request; this protocol is about letting one sign in, and in v0.1 “write” means all of write.
What I would do with today: open the price sheet before you trust the headline, file a cyber access application before you need one, and reread the HubSpot memo and the agent protocol in a month to see which sentences survived.
If today’s math lands on something you run, book a free intro call. It is thirty minutes, so bring the numbers.