The Day in AI
In short
- Bland released Tofu-1 on 15 November, a speech-to-speech model for phone calls that listens and talks at once, in a closed enterprise preview.
- Bland says over 92% of listeners in its screenings could not tell Tofu-1 was a model, and it has published no benchmarks.
- Artificial Analysis scores Google's Gemini 4 Hydrogen at 51 on its Intelligence Index, one point above the larger Gemini 4 Argon.
- Meituan released LongCat-2.5 on 2 November, a 1.6 trillion parameter multimodal model that uses about 48 billion parameters per token.
- Anthropic released Claude for Manufacturing on 12 November, a plugin of 17 skills for quoting, quality management and maintenance.
- The US and its Pax Silica allies reportedly agreed to a joint list of AI technologies to protect from China, according to the Financial Times.
The day in brief
Bland released Tofu-1 on 15 November, a full-duplex speech-to-speech model for phone calls, and says over 92% of listeners in its screenings could not tell it was a model.
Most voice agents chain speech recognition, a language model and speech synthesis, passing text between them. Tofu-1 takes audio in and gives audio out with no transcript in between, so the model can still use the caller's tone, hesitation and background sound. It is full-duplex, which means it listens while it speaks. A caller can interrupt it mid-sentence, and it keeps the conversation going while it works on a request.
The 92% figure comes from Bland's own screenings, as relayed by MarkTechPost. Bland has not published the architecture, model size, latency or any benchmark results, so there is no independent measure yet of how Tofu-1 compares with other voice systems. Access is a closed research preview for enterprise customers.
Independent scores for Gemini 4 Hydrogen, which Google released on 12 November, put the cheap model close to the top of its own family. Artificial Analysis rates it 51 on its Intelligence Index at the default high thinking setting and 54 at xhigh, against 50 for Gemini 4 Argon. Vals ranks Hydrogen first on its Tax Agent Bench at 52.3% and tied first on its Legal Research Bench at 55.8%, and arena.ai's Agent Arena places it fifth. It costs $1.50 per million input tokens and $7.50 per million output tokens.
Meituan's LongCat-2.5, released on 2 November, is a 1.6 trillion parameter model that uses about 48 billion parameters for each token and is built for long agentic coding sessions. It represents images as multi-level discrete tokens in the same stream as text, and Meituan turned failed runs from real agents into verifiable training tasks for large-scale reinforcement learning. Meituan reports 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0, and says it optimized 4-bit weight and activation (W4A4) inference on domestic Chinese chips. A limited-time price is roughly €0.27 per million input tokens and €1.07 per million output tokens, and no independent evaluation has been published.
Also in the news
- Anthropic's Claude for Manufacturing reportedly connects to Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite, and it drafts quotes and candidate root causes for staff to review.
- The US and its Pax Silica allies, including Japan, the UK and India, agreed to build a joint list of AI technologies and supply chains to protect from China, the Financial Times reports.
- US Congress members want to tighten export rules on AI chips after an Nvidia spy scandal, according to The Verge.
- Nvidia built a multibillion-dollar business letting Chinese companies use its chips in Southeast Asian data centers, the Wall Street Journal reports.
- UniPat AI published the SPU benchmark, which asks agents to predict real wet-lab results, and reports that no agent scores above 0.5.
- OpenAI reportedly had one model read another model's internals and raised its math score from 7.9% to 48.1%, though details are thin so far.
- Nidus, which builds AI for manufacturing robots used by US defense contractors, raised $30 million across pre-seed and seed rounds, Upstarts Media reports.