A new model tops the benchmarks. Should you switch?
On 2,000 PII-redaction requests, the benchmark favorite cost four and a half times our control. GPT-5.6 Terra matched its quality gain at an eighth of the price.
Test it on your own use case first:
three.dev/blog/your-use-case…
Anthropic cut Sonnet's price by a third. But on the same 191 requests, our bill went up 4%.
The tokenizer, answer length, reasoning defaults and settings can move a bill as much as the list price. You only find out by testing your own requests.
three.dev/blog/the-free-upgr…
Two coding agents got the same bug in a test support bot.
One had the code and a short bug report. It cut the failure from 6.8% of responses to 4.2%.
The other was connected to three.dev and could read all 99 failing responses. It found the cause. 0%.
three.dev/blog/detect-and-fi…
People are trying @typesafeai's Jev as an LLM judge. We run AI Judges, so we tested it on roughly 15,000 requests.
It can't replace one. But as a filter picking what the judge reads, it helped catch 3.3x the failures of random sampling at 1/47 the cost.
three.dev/blog/jev-vs-the-ll…
DeepSeek and Qwen just agreed on where a model's facts should live: in RAM or on an SSD, off the GPU.
In Qwen3.8-Flash-Next, 51B of 176B parameters no longer need GPU memory.
Our CTO on how this works and what it means for inference cost: three.dev/blog/knowledge-jus…
