The Best New AI Products and the Key Trends They Represent
New AI products this week show three trends: model-agnostic agents, cheap capable models from US labs, and labs finally competing on product experience, not just raw intelligence.
Watch on YouTube ↗Charlie's summary
- ▶ 0:28
Open source has a viable business: Hermes agent reached a $1.5 billion valuation by giving the agent away free.
Nous Research offers Hermes free with bring-your-own-model, and sells subscriptions bundling frontier model usage and hosted instances. The Wall Street Journal reports 22 million installs and roughly 2.5% of global token usage. Annualized revenue is $36 million, with a $100 million target by year end. CEO Dylan Rolnick argues that with software reproduction costs plummeting, openness gives users a “foundation of trust”.
- ▶ 10:37
The Ramp data shows a structural realignment: cheaper, token-efficient models are pulling enterprise spend away from the big two labs.
Ramp's list of fastest-growing vendors is a view of the bleeding edge, which tends to set patterns others follow. Typesafe's Jev ranked first in growth relative to size and gained a full point of market share in a month. SpaceX, Anthropic and OpenAI all landed in the top five. Ramp economist Eric Kerazian says cheaper models push spend down, while expensive image and video models (Runway, Higgsfield) push it up.
- ▶ 14:12
Multimodelness: Grokbot will use whichever back-end model is best, including Claude Opus, Midjourney and Suno.
Elon Musk said SpaceX will use “the best back-end model for any given task”, while simple requests go to a fast Grok version. The host reads this as a move to compete on both harness and model. Businesses and power users want to avoid lock-in to one ecosystem, so the harness and context become the lock-in and models become portable. OpenAI's DevDay move to sell open-model inference through its B2B marketplace points the same way. The host's view: no lab can be best at everything.
- ▶ 17:48
Claude Haiku 5.5 is a narrative violation: the cheapest capable models now come from US labs, not China.
Anthropic is competing at the low end. Haiku 5.5 costs a quarter of Sonnet 5.5 and half of Haiku 4.5 per token, with an 80% token discount on prompts under 100,000 tokens. It scored 72.4% on OS World against 48.9% for GPT-6 Luna, at about 12 to 21 cents per task. Investor Hibeshi says the whole Pareto frontier is now US-owned. The host notes that sovereignty and data control, which US labs haven't addressed, may still drive interest in Chinese open models.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.