OpenAI cuts GPT-5.6 Luna 80%, Anthropic halves Opus 5
A DeepSeek-driven price war reaches the frontier labs — plus River's API to train your own open model cheaply and OpenAI's gated cyber LLM.
Copy markdown
OpenAI's Luna drops to $0.20 per million
GPT-5.6 Luna's input price falls 80%, from $1.00 to $0.20 per million tokens; mid-tier Terra is down ~20%, and premium Sol holds its rate but runs faster. High-volume Luna work — extraction, classification, agent loops — just got 5x cheaper to run.
Anthropic halves Opus 5, freezes Sonnet 5
Claude Opus 5 lands at $5 / $25 per million input/output tokens — about half of Fable 5 — and the September price hike planned for Sonnet 5 is cancelled, holding launch rates indefinitely. Frontier Claude is now materially cheaper to keep in production.
What lit the fuse: Chinese rivals
The cuts follow DeepSeek V4 Pro at $0.87 per million output tokens, Moonshot's Kimi K3, and Z.ai — which has passed both Claude and ChatGPT in token volume on OpenRouter. If you'd been eyeing a cheaper base, the leverage just moved to your side.
River API: train and own a custom model for under $1K
River's API runs LoRA fine-tuning and RL on open weights from 35B to 1T params — Qwen3.6, Kimi K2.6, GLM 5.2 — billed per token, not per GPU-hour, and serves any checkpoint from an OpenAI-compatible endpoint. A full RL run on Qwen3.6-35B reportedly lands under $1,000; the v0.1 preview is open to sign-ups.
OpenAI's GPT-5.6-Cyber ships — but not to you
OpenAI launched a security-tuned GPT-5.6 with relaxed safeguards for exploit development, gated behind 'Daybreak' approval for vetted firms like CrowdStrike and NCC Group. It's credited with surfacing real browser vulnerabilities — a marker of where autonomous vuln-hunting is heading, even if you can't run it yet.