Engineering notes on billing you can audit, caching, and shipping with the OpenAI SDK — real numbers, no fluff.
A capability listed on a model page is a claim, not evidence. So we sent all seven models in our catalog the same two images and asked what was in them. Three answered correctly. Four did not — and only one of those four said so.
August 18, 2026·5 min read
Two new models, one giant window: DeepSeek V4 Flash and Pro with 1,048,576 tokens of context — 4× our previous ceiling — measured, priced 40% off list, and OpenAI-compatible from the first request.
August 16, 2026·4 min read
No waitlist, no sales call: sign up, top up, swap base_url. Every model is 40% off list price for the launch period — applied automatically, live on the pricing page.
August 12, 2026·3 min read
Prepaid credits, integer nano-USD arithmetic, and a formula you can recompute from your own usage log. The entire billing model, with a real request worked through by hand.
August 11, 2026·6 min read
No new SDK to learn: swap base_url and the key, keep everything else. Working setups for Python, Node, LangChain, and the Vercel AI SDK — plus the two errors you will actually meet.
August 11, 2026·4 min read