Ethan Mollick notes Google is the only AI lab releasing full any-to-any multimodal models. OpenAI and Anthropic have notably limited or absent multimodal output capabilities.
Key facts
- Google is the only lab shipping full any-to-any multimodal models.
- Anthropic has no multimodal output capability.
- OpenAI uses selective multimodal, not unified any-to-any.
- Open-weight models offer vision input but limited output.
- Mollick posted this observation on X in 2026.
Wharton professor Ethan Mollick observed on X that full multimodal models — those accepting and generating across text, images, audio, and video — remain surprisingly niche. He points out that Google is the only major lab shipping such models, while OpenAI restricts its multimodal features to selective input/output paths (e.g., DALL·E for image generation, Whisper for audio transcription, but no unified any-to-any pipeline). Anthropic has no multimodal output at all, and open-weight models like LLaMA 3.2 Vision or Qwen2-VL offer vision input but limited or no generative output beyond text.
Why this matters
Full any-to-any capability is the architectural foundation for agents that perceive and act across modalities — a key goal for autonomous systems. Google's lead here, via Gemini 2.0's native multimodal generation, gives it a structural advantage in building end-to-end agentic workflows. The gap is not just a feature difference; it reflects deeper architectural choices about whether to train a single model on all modalities or bolt on separate specialist modules. Anthropic and OpenAI's selective approach may be a deliberate trade-off for safety or latency, but it leaves them dependent on routing and orchestration layers that add complexity and failure modes.
Open weights are catching up unevenly
Open-weight models like Llama 3.2 Vision (text+image input, text output) and Qwen2-VL (text+image input, text output) offer partial multimodal capabilities, but none match Google's any-to-any generation. The community has focused on vision-language models rather than full multimodal generation, partly due to the compute cost of training and running such models. As Mollick implies, the lack of open any-to-any models may slow progress on agentic applications that require real-time multimodal interaction.
Key Takeaways
- Mollick notes Google alone ships full any-to-any multimodal models; OpenAI and Anthropic lag.
- This gives Google a structural advantage in agentic workflows.
What to watch
Watch whether Anthropic or OpenAI announce any-to-any multimodal generation at their next major model release — likely Claude 4 or GPT-5. If neither does, Google's lead in agentic multimodal workflows will widen. Also track open-weight models like Llama 4 for any-to-any support.
