TypeSafe's Jev is a System One model: it returns typed, probability-backed answers instead of generated text, at $0.042 per million input tokens with free output. That makes it a cheap fit for the classification and routing work teams have been quietly handing to small LLMs. Coolhand ingests Jev calls today.
Read more →
Cache write, cache read, refresh, breakpoint, TTL, storage fee. Six words that mean different things at different providers — and one of them costs money while you sleep. A working glossary for engineers and the people who approve their invoices.
Read more →
Three providers, three incompatible cost shapes, and three rules of thumb that survive the next repricing. Includes the side-by-side chart and the arithmetic behind each rule.
Read more →
Cache writes now cost 1.25x on OpenAI's flagship line, and the default breakpoint lands on your most volatile content. There's no TTL to tune — the whole optimisation is boundary placement and key discipline. Worked four ways.
Read more →
Anthropic is the only provider where the TTL you pick changes your write price. Here's the break-even math for an orchestrator, a sub-agent, a chatbot, and a high-volume extraction pipeline — plus the clock-start rule that makes most cadence estimates wrong.
Read more →
Google charges no write premium and no read penalty — it charges rent. That inverts the whole decision: the question isn't cadence, it's reads per hour. Below roughly three, explicit caching costs you money. Worked four ways.
Read more →
A working proof of concept and a funded, trusted production agent are different problems. Here's the three-part case — the cost you're replacing, the actual cost of running it, and evidence it's improving — that gets an agent like an automated bug-fixer taken seriously by the people who fund it.
Read more →
Silent failures don't crash — they silently frustrate your customers while you are none the wiser. We rank the five best strategies and open-source tools for catching agents that quietly fail, each scored on signal caught per hour of effort.
Read more →
We compared gateway overhead across GoModel, Bifrost, Portkey, and LiteLLM — hot-path latency, memory, and cold starts that matter once you're routing real production traffic. Here's why we lean toward GoModel, first in a new series on AI tool bakeoffs.
Read more →95% of enterprise AI projects fail. Most never close the loop between what the model produces and what users actually want. We're open sourcing everything we've learned about fixing that.
Read more →On June 11 we observed elevated HTTP 431 errors from the OpenAI API across client accounts. OpenAI has since resolved the incident, but our logs show the issue beginning roughly 90 minutes before their official start time.
Read more →