AI API

Jev is TypeSafe AI's new decision model: state in, typed Choice/Score/Boolean answers out, no text generation. What it actually is, how it compares to GPT, Claude and Gemini, and when it's the right tool.

A token is roughly four characters or three-quarters of a word in English. Here is the quick reference table, how to estimate from your own text in seconds, and the five cases where the ratio breaks badly.

A plain-English breakdown of GPT-6 Astra: the real specifications, what actually changed versus the GPT-5.6 line, what the launch benchmarks do and do not prove, and whether you need it at all.

Same context window, same maximum output, 2.5x the price. A specification-by-specification and benchmark-by-benchmark comparison of GPT-6 Astra and GPT-5.6 Sol, and a straight answer on when the upgrade pays for itself.

A practical guide to building on GPT-6 Astra: what reasoning effort actually costs you, what strict-mode structured outputs guarantee and what they do not, how the agentic tool surface fits together, and how to keep the bill sane.

Input tokens, output tokens, cached tokens and cache writes — what each one costs on GPT-6 Astra, what actually counts toward your input bill, and three worked examples showing where the money really goes.

Markdown fences, truncated objects, refusals, trailing prose and schema-valid nonsense — the failure modes specific to model-generated JSON, and a checklist for diagnosing each one quickly.

Cached input tokens bill at a tenth of the standard rate. Here is exactly what gets cached, the 1,024-token minimum, why writes cost 1.25x, and the prompt-ordering mistakes that silently destroy your hit rate.