OpenAI Pushes Prices Down: GPT-5.6 Luna Gets 80% Cheaper
On July 30, OpenAI published a blog post titled "Advancing the price-performance frontier with GPT-5.6". The headline is straightforward: Luna, the cheapest of the three GPT-5.6 tiers (Luna, Terra, Sol), is getting an 80% price cut.
This isn't a temporary promotion. OpenAI explained the technical drivers: kernel-level optimizations reduced end-to-end serving costs by 20%, while experimental improvements boosted token-generation efficiency by over 15%. Together, these changes justify the price drop.
GPT-5.6 launched mid-2026 with three tiers. Luna is the fastest and most affordable, meant for high-volume tasks. Sol handles the hardest reasoning work. Luna was already priced below previous-gen models, but an additional 80% cut fundamentally changes the cost math for teams running large-scale agent workflows.
Some HN commenters compared Luna to Claude Opus 5. The general sense is that Luna's Extra High reasoning mode already covers most scenarios where you'd reach for Opus 5. If that holds, OpenAI's strategy is clear: Luna for volume, Sol for peak capability.
Notably, the blog didn't mention Sol Ultra — the version that reportedly proved the Cycle Double Cover Conjecture (a 30-year open problem in graph theory). Ultra likely remains under restricted release, currently available only through Codex.
Google's Robotics Models Arrive: Gemini Robotics 2
Google DeepMind released Gemini Robotics 2 on the same day. It's actually three models, not one.
Gemini Robotics 2 (VLA model) converts vision and language input into physical robot actions. It can control full humanoids — from toes to fingertips — and dual robotic arms. Demos showed Apptronik's Apollo 2 humanoid walking, crouching, reaching, and grasping in one continuous sequence. Previous Gemini Robotics models only controlled the upper body; this release adds full lower-body coordination.
Gemini Robotics ER 2 (reasoning model) acts as the robot's brain. It processes camera video feeds for real-time spatial reasoning, plans multi-step tasks, and coordinates collaboration between different robots. DeepMind showed a wheeled rover and a robotic arm working together on a task neither could handle alone. Key metrics: progress classification accuracy of 57.4% (tracking how much of a task is done), moment-finding accuracy of 91.3% with a 0.96-second time error.
Gemini Robotics On-Device 2 is optimized for local deployment with no internet needed. It can adapt to entirely new robot hardware in just a few hours, using fewer than 200 demonstration examples.
What Stands Out
The pricing war continues. OpenAI's cuts come against sustained pressure from Chinese model makers. GLM 5.2 and Kimi K3 pushed pricing to aggressive levels, and Luna wasn't winning on value before. Now the math changes.
The robotics deployment path is getting clearer. Gemini Robotics ER 2 is available through Google AI Studio and the Gemini Enterprise Agent Platform — meaning developers can start building and testing without a physical robot. This software-first approach lowers the barrier compared to building a full robotics stack from scratch.
Multi-robot collaboration is a real step forward. The industry has long struggled with coordinating multiple robots. ER 2 uses shared semantic understanding to let different robot types communicate — solving a practical deployment problem, not just a lab demo.
Both announcements hit the same day — one pushing prices down, one expanding capabilities. Different directions, but the underlying logic is the same: AI is moving from conversation models to systems that actually get work done.