OpenAI launches GPT-5.6 Sol, Terra and Luna

#1 📝 OpenAI News

How GPT-5.6 fuses frontier intelligence with frontier efficiency - OpenAI's GPT‑5.6 family balances capability and cost: flagship GPT‑5.6 Sol outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index at less than half the cost, Terra matches GPT‑5.5 on intelligence benchmarks at half the price, and Luna is priced 80% lower than Sol. By using Sol to autonomously rewrite Triton/Gluon kernels, tune load‑balancing and KV‑cache configurations, and improve speculative decoding, OpenAI reports a 20% reduction in end‑to‑end serving costs and more than a 15% gain in token‑generation efficiency.

#2 📝 OpenAI News

Accelerating scientific discovery with ChatGPT for Academic Researchers - OpenAI is launching ChatGPT for Academic Researchers to give 100,000 scientists, mathematicians, and engineers free access to frontier models—starting with 10,000 this summer at institutions including the Institute for Advanced Study and École normale supérieure—with participants able to invite up to four collaborators, access ChatGPT/ChatGPT Work/Codex (including GPT‑5.6 Sol Pro at launch), and opt out of using their data to train models by default. The program offers expanded research features (higher usage limits, larger context windows), over 75 life-science skills and connectors, training and hands-on support, and is part of a $250M+ commitment through 2027 (including $50M for NextGenAI); OpenAI cites ~1.3M weekly advanced-science ChatGPT users generating 8.4M messages and benchmark claims such as GPT‑5.6 Sol scoring 83% on FrontierMath Tier 4 (vs 72.5% for GPT‑5.5) and GPT‑5.6 Sol Pro solving 31.5% of GeneBench Pro tasks.

#3 𝕏

NVIDIA AI unveiled Cosmos-Dreams, a neural closed-loop simulator demoing world models as the data engine of physical AI at SIGGRAPH 2026. They highlighted that “data stops being something you collect; it becomes something you compute.”

#4 𝕏

NVIDIA AI thanks @perplexity_ai for open-sourcing Numbat, an agent-detection and response layer, to the Open Secure AI Alliance.

#5 𝕏

Santiago introduced Kimi K3, an open‐weight 2.8T‐parameter coding model with 1 M‐token context, multimodal inputs and agentic long‐run sessions. He highlighted the Verdent + Moonshot AI partnership behind its optimized agentic coding workflows.

#6 𝕏

Peter Yang walks through six steps using Claude Design and Claude Code to build Tastemaker—an app for sharing movies, TV shows, and games. He stresses that fleshing out requirements and up-front AI-driven design are crucial to avoid sloppy products.

#7 𝕏

Guillermo Rauch uses React/Next.js’s `` to one-shot, parallel-stream multiple models/providers from the Vercel AI Gateway API, yielding ultra-clean code.

#8 𝕏

Santiago launched BAND, an interaction layer enabling personal agents to communicate across identities, channels, and routing, showcased with a calendar collaboration demo.

#9 𝕏

Hugging Face launched a “Sign-In with Hugging Face” OAuth button for websites/apps, letting users share their email, create model/dataset repos, store data in Buckets, or kick off GPU-backed Jobs seamlessly.

#10 𝕏

Harrison Chase demos openwiki, a “dreaming” memory–powered wiki that runs scheduled background jobs to parse LangSmith traces of coding agents and automatically update your codebase documentation.

#11 𝕏

Guillermo Rauch built a harness that adds vision capabilities to every AI model, enabling them to process and interpret images seamlessly.

#12 𝕏

Cursor launched its iOS app, letting you create, review, and merge code from anywhere—download it now on the App Store.

#13 𝕏

v0 launched support for open-weight models, starting with Kimi K3—try it at https://v0.app/?kimi3

#14 𝕏

Cursor launched native iOS support with an organized inbox and a full PR review experience on iPhone and iPad—covering comments, checks, and approvals.

#15 𝕏

Peter Yang highlights Claude Design’s simple clarifying-questions feature that prompts you to refine design requirements before prototyping. He’s surprised other AI design tools haven’t adopted this effective approach.

#16 𝕏

OpenAI found that GPT-5.6 Sol was underperforming on the ARC-AGI-3 2D puzzle benchmark because its evaluation harness wasn’t retaining learned context; enabling two API settings tripled its scores with 6x fewer output tokens.

#17 𝕏

OpenAI launched ChatGPT for Academic Researchers, offering free access to its frontier models for 10,000 scientists, mathematicians, and engineers now—and expanding to 100,000 by 2027 to accelerate cross-disciplinary discovery.