OpenAI Launches GPT-5.6 Series Amid US Government Restrictions; ByteDance Unveils Seedance 2.5 Video Model

Key Takeaways

1. OpenAI released GPT-5.6 series with Sol, Terra, and Luna models; US government limits availability to a controlled preview, with Sam Altman criticizing the process and announcing a custom AI chip.
2. ByteDance's Seedance 2.5 video model supports 30-second 4K video with audio and local editing, achieving $2B ARR with 3.3M hours of video generated, reinforcing China's lead in video AI.
3. Open-source momentum continues: Meta's PRX Pixel 7B text-to-image model and NVIDIA's Nemotron 3 Ultra top an agentic benchmark, while Hugging Face joins the OpenEnv agent protocol effort.
4. Anthropic restores access to its strongest cybersecurity model Mythos 5 for US critical infrastructure organizations after government approval, highlighting ongoing regulatory pressure on frontier AI.

1. Major Model Releases and Regulatory Constraints

1. OpenAI unveiled the GPT-5.6 series, comprising three models: Sol (frontier), Terra (balanced), and Luna (efficient/fast). Sol offers significant performance improvements over GPT-5.5 at the same price, while Terra delivers 5.5-level performance at half the cost. However, due to US government requirements, the models are only available in a limited preview, with OpenAI working to expand access. — via 1 2
2. Sam Altman criticized the current deployment process as suboptimal and called for a transparent, reliable early-access pipeline. He also announced OpenAI's custom AI chip, codenamed Jalapeño, developed with Broadcom, designed for LLM workloads like ChatGPT and Codex, aiming to expand the full-stack platform. — via 1
3. Anthropic announced that, after close collaboration with the US government since June 12, its strongest cybersecurity model Mythos 5 can now be redeployed to US organizations operating and defending critical infrastructure. The company is working to restore access rapidly and continues efforts to make Fable 5 generally available. — via 1
4. Ethan Mollick noted that new licensing regimes will end vague posts from labs when releasing models, implying increased regulatory transparency. — via 1

2. Open Source and Open AI Momentum

1. Yann LeCun released PRX Pixel, an open-source 7B text-to-image model that generates images directly in pixel space. Weights are available, and integration into Diffusers is underway. He also argued that slowing rollouts and regulatory capture damage US competitiveness, and that open AI will gain increasing funding and pressure. — via 1 2
2. NVIDIA's Nemotron 3 Ultra ranked top among open models on the new AA-Briefcase benchmark, which evaluates complex agentic tasks, including tasks unseen during training. — via 1 2
3. Hugging Face announced that RadixArk joined the OpenEnv community, an agent environment protocol layer standardizing environment publishing and consumption. Hugging Face will join the committee alongside PyTorch, NVIDIA, and others, first integrating Miles with end-to-end examples. — via 1

3. Video Generation Advances: ByteDance Leads with Seedance 2.5

1. ByteDance is set to release Seedance 2.5 video model in early July, supporting 30-second videos with audio and 4K resolution, up to 50 reference sources, and local editing. Seedance 2 has already achieved $2 billion ARR and generated over 3.3 million hours of video, equivalent to 30 billion movies. — via 1
2. Deedy observed that video generation length doubles every 6 months, predicting that within 18 months AI can generate a full music video in one click, and that China continues to lead the US in video models. — via 1