This week's developments that matter directly for Estonian small and mid-sized companies planning AI work — pricing, security, liability and new tooling.
Anthropic and OpenAI shipped new models on the same day at roughly half the price of the previous generation
Claude Opus 5.5 plus GPT-6 Sol and Luna offer modest capability gains at clearly lower token costs. Why it matters: if you shelved an AI feature because the numbers didn't work, redo the maths — the same build may now cost half as much. simonwillison.net
The frontier model market has entered a comparison-shopping phase where quality is no longer the only differentiator
Vendors increasingly compete on price and speed rather than capability alone. Why it matters: avoid locking into one provider — keep model swapping easy in your codebase, because the landscape shifts every few months. arstechnica.com
Microsoft took down a criminal service platform that used AI to break into more than 12,000 accounts
It was a turnkey offering that makes mass account takeovers fast and cheap. Why it matters: phishing in fluent Estonian is getting both better and more frequent — two-factor authentication and a short staff briefing are still the cheapest defence. arstechnica.com
A serious flaw in Meta's new AI assistant lets an attacker fully hijack the agent
A simple user-tricking attack was enough to take over an assistant with broad permissions. Why it matters: before giving an AI agent access to email, files or payments, cut its permissions to the minimum and require human approval. arstechnica.com
A new class of "decision models" returns categories and confidence scores instead of text
TypeSafe AI's Jev answers yes/no and multiple-choice questions numerically, very fast and very cheaply. Why it matters: routine classification — routing customer emails, sorting invoices, validating inputs — becomes cheaper and more predictable than using a chat model. simonwillison.net
Cloudflare Python Workers reached general availability after a two-year preview
Python is now a fully supported language on Cloudflare's edge platform. Why it matters: a small team can ship an AI prototype with no server administration and very low fixed costs. simonwillison.net
British Columbia is suing OpenAI, seeking accountability for a chatbot's role in a tragic case
The claim also demands the release of the user's conversation logs. Why it matters: case law is starting to define who answers for AI output — document where and why you use AI in your business. arstechnica.com