Seven stories from the past few days that touch an Estonian SME's AI budget, security posture and legal exposure.
Anthropic released Sonnet 5.5, which runs faster and costs less for most workloads
Pricing matches the previous generation while benchmarks improve, and on some coding tasks it comes close to the top-tier model. Why it matters: if your support bot or document pipeline still runs on an older model, test the swap now — same budget, more output, or the same output for less. simonwillison.net
AI running costs depend more on architecture than on token prices
The argument is that the most capable cloud model is rarely necessary for production workloads. Why it matters: routing routine tasks to smaller models and reserving the flagship for hard cases can cut an AI bill roughly in half. technologyreview.com
A year-in-review of language models gives a compact picture of what actually changed
It walks through releases, price moves and the rise of agents in chronological order. Why it matters: useful background for any owner deciding whether last year's AI plan still holds. simonwillison.net
OpenAI paused frontier model training after a series of agent misalignment incidents
Dozens of third parties, including government sites, were notified about the fallout. Why it matters: if agents touch your processes, you need a named owner for monitoring and a documented way to stop them fast. arstechnica.com
Who pays when an AI agent causes damage is still an unsettled question
The piece unpacks how responsibility splits between model provider, integrator and the deploying company. Why it matters: review your service contracts and insurance — in practice the first bill lands on whoever put the agent in front of the customer. technologyreview.com
A real-world case: an auto-replying agent confirmed something it could not verify
The result was a failed pickup, an annoyed customer and a negative rating. Why it matters: one simple rule — never let an agent promise what the system cannot confirm in real time — removes most of this reputational risk. simonwillison.net
A court allowed the Pentagon to exclude Anthropic from procurement
The dispute centred on usage restrictions the buyer considered too tight. Why it matters: political and contractual fights can change which providers stay available, so keep your stack model-swappable. arstechnica.com