OpenAI's Decisions API is GA: typed answers 10x faster than Responses
The cheap, fast endpoint for classification and routing ships to all devs — alongside 12-hour prompt caching and NVIDIA's open RAG retrieval stack.
Copy markdown
The Decisions API leaves beta with gpt-6-luna
OpenAI's gpt-6-luna Decisions API is now generally available — it turns text and images into schema-typed answers roughly 10x faster than the Responses API, now with up to 400 questions per request (up from 200 in beta), per-question image inputs, and structured inputs via input_json. At $0.10/$0.50 per 1M input/output tokens ($0.01 cached), it's the cheap, fast layer for routing, moderation, extraction, and bulk labeling at scale.
12-hour prompt caching lands on the Responses API
The Responses API can now hold a prompt cache for 12 hours, up from the usual few minutes. Cache writes cost 2.5x the input price, but a long-running agent or a pipeline that reuses one big system prompt or document can now stay warm across a full workday instead of repaying full input on every call.
NVIDIA open-sources a full RAG retrieval stack
At GTC Berlin, NVIDIA released open-weight Nemotron Retriever embedding and reranking models (1B-8B, text and image, 34 languages) plus Nemotron Parse for pulling text and charts out of PDFs at ~10.5 pages/sec on a single H100. The 1B embedder fits on a DGX Spark, and weights ship with training data and recipes — a drop-in upgrade for multilingual, multimodal RAG pipelines.
OpenAI lets orgs lock projects to specific service tiers
Platform settings can now restrict which service tiers a project — or a whole organization — is allowed to use. It's a quiet one, but it's the switch that stops a side project from silently burning spend on priority traffic, or pins a team to the tier you actually budgeted for.