OpenAI gives GPT-6.1 Sol an Ultrafast tier: up to 8x faster via API
A faster, pricier tier for latency-bound agents, as Reflection preps a 501B open-weight coder, Mistral Large 4 opens preview, and Decisions API goes public.
Copy markdown
Sol Ultrafast: 8x the speed, 6x the price
GPT-6.1 Sol's new Ultrafast tier runs up to 8x faster at $12 input / $60 output per 1M tokens — six times standard Sol — with rate limits from 1M to 40M tokens per minute. If your coding or agent loop is bottlenecked on latency rather than token cost, this is the dial you've been missing; price it before you flip it on.
Reflection Beam: a 501B open-weight coder lands this month
Nvidia-backed Reflection unveiled Beam, a 501B-parameter sparse MoE (23B active) that scores 80.9 on SWE-bench Verified and 80.1 on Terminal-Bench v2.1, shipping under Apache 2.0 later in October. It's the first US open-weight model built to rival the GLM/DeepSeek/Qwen tier — self-hostable frontier-grade coding is nearly here. The catch: weights aren't out yet, and access is waitlist-only for now.
Mistral Large 4 opens its API preview
Mistral's new flagship — a ~1T-param MoE with 49B active, a 1M-token context, and 160+ languages — is live in public preview at $1.36 input / $4.18 output per 1M tokens, with open weights promised by month's end. A cheap, long-context, multilingual frontier option you can wire in today and swap to a local copy later.
OpenAI's Decisions API hits public beta
The Decisions API — the small gpt-6-luna model picking one answer from a fixed set at $0.10 per 1M input tokens — is now open to all developers. It's built for classification, request routing, and choosing an agent's next action, and Simon Willison already shipped an llm-openai-decisions plugin. A dirt-cheap replacement for brittle prompt-and-parse routing.