AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Key topics and reactions
OpenAI Makes GPT-5.6 Sol, Terra and Luna Generally Available Across ChatGPT, Codex and API
OpenAI released the GPT-5.6 family in a three-tier structure: the flagship Sol for orchestrating long-running agents, the balanced Terra for everyday small tasks, and the fast, low-cost Luna for quick first-pass processing. The models began rolling out across ChatGPT, Codex (including GitHub Copilot integration) and the API, moving from a limited preview to general availability.
In a CNBC interview, Altman said GPT-5.6 Sol cuts output tokens by up to 54% on agentic coding tasks while matching or exceeding rival models on performance, framing cost efficiency as a differentiator for enterprises focused on return on investment. Pricing is tiered: Sol at $5/$30 per million input/output tokens, Terra at $2.50/$15, and Luna at $1/$6, with a 90% discount on cached prompt reads and a new charge for cache writes.
The series was initially limited to trusted partners at U.S. government request, and OpenAI said it completed safety reviews with officials including Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent before lifting the restrictions on July 8. Independent analysis by The Decoder found Sol nearly matches Fable 5 on aggregated benchmarks at roughly one-third the cost.
OpenAI Introduces ChatGPT Work, a Codex-Based Agent for Cross-App Automation
ChatGPT Work is designed to automate cross-app and long-duration tasks, and it combines Codex and ChatGPT into a single desktop application. The product targets workflow automation for large organizations.
OpenAI said NVIDIA's internal teams are already using the agent for workflow automation and insight extraction, indicating early enterprise deployment. Developers using Sol as an orchestrator reported passing high-level goals to the model, which then plans, executes and self-corrects.
Early testers cited reduced failure rates when Sol retains context and delegates subtasks to Terra, though some flagged high token consumption and rate limits as practical constraints, and noted that fully unattended operation remains unstable.
More Than 22 University Researchers Moved to Major AI Companies in 2026
According to The Information, at least 22 academics moved to industry AI labs in 2026, typically via leaves of absence. High compensation and access to large compute resources were cited as the main draws, while universities struggle to retain staff amid salary gaps and limited resources.
High-profile moves include Transformer co-author and Gemini co-lead Noam Shazeer going from Google to OpenAI, AlphaFold leader and 2024 Nobel chemistry laureate John Jumper moving from Google DeepMind to Anthropic, and Gemini contributors Jonas Adler and Alexander Pritzel leaving Google for Anthropic.
Some remaining faculty warned that the outflow could weaken the West's ability to compete with open-source AI models from China. Analysts note the trend extends a long-running shift of researchers from academia into industry over the past 15 years.
xAI's Grok-4.5 Rises to Top Benchmark Rankings at Lower Cost
Grok-4.5 reached #3 on Code Arena (Frontend) and claimed #1 on Terminal-Bench v2 and SWE Marathon. On CursorBench it placed third overall at 66.7%, with a cost of about $1.51 — roughly one-tenth that of comparable models.
Users described the model as balanced across coding, vision and reasoning, with several making it their default. Frontend design results were reported as competitive with GLM-5.2 and Claude Opus 4.8 (thinking).
Skeptics said the model still lags Claude on code quality and that real-world evaluation is pending, while many are waiting for a planned expansion to 1M-token context.
Category highlights
Runway Launches Developer Platform for AI Media
Runway released what it calls its first developer platform for AI media, offering zero-day access to the latest models such as Seedance and exposing a real-time inference engine via API. Meta previewed Muse Video for image-to-video generation, Stability AI updated Stable Video Diffusion for 4K output and longer clips, and Seedance 2.0 saw wide platform adoption including PixVerse and CapCut integration.
Meta Muse Image, Google Gemma 4, and Microsoft Aurora 1.5 Update
Meta announced its first image generation model, Muse Image, featuring complex prompt understanding, multi-image blending and Instagram/WhatsApp integration. Fine-tuning cases for Google Gemma 4 E2B and an update to Microsoft's Aurora 1.5 weather foundation model also drew attention.
Perplexity Computer and Google AlphaEvolve Advance
Perplexity Computer previewed a new orchestrator model based on GLM 5.2 at 0.344x the cost of Opus. Google AlphaEvolve reached general availability on GCP, and MiniMax M3 advanced multi-model routing for use-case-based model selection.
Dreamina Seedream 5.0 Pro, fal and QuiverAI Arrow 1.1 Move
Beyond Reve 2.1's Arena rise, Dreamina Seedream 5.0 Pro added pixel-level editing and multilingual prompts. fal touted 0.45-second inference and QuiverAI's Arrow 1.1 offered SVG generation.
OpenAI Build Week, fal Video Hackathon and SIGGRAPH 2026 Ahead
OpenAI Build Week begins July 13. fal is running a 72-hour video hackathon with $150K in prizes and a near deadline. NVIDIA's SIGGRAPH 2026 panel and Tencent Worlds of Play are also upcoming.
Key trends
Long-Running, Tool-Using Agents Enter Enterprise Use
Competition among long-running, cross-tool agents accelerated with ChatGPT Work, Muse Spark 1.1, Perplexity Computer and Grok Build. Real-task evaluation on venues such as Agent Arena is becoming established, and operational use has begun at large organizations such as NVIDIA.