The Most Important Trends Showing Up in New AI Products

New AI products are competing on more than model capability: price, speed, choice of model, product experience and access to high-quality data are becoming central. Ramp’s October rankings provide an early signal from customers it describes as bleeding-edge adopters, not a representative picture of the whole enterprise market. Jev and Type Safe AI ranked first in growth relative to company size and second in market-share growth; Jev gained one share point in a month and was gaining faster than every frontier lab except Anthropic. Anthropic, SpaceX AI and OpenAI were also among the top five for share growth, while DeepSeek and model routers OpenRouter and Featherless appeared in the rankings.

Anthropic’s Claude Haiku 5.5 illustrates the push toward cheaper, faster models. Anthropic calls it its cheapest, fastest and most capable small model to date, and it substantially outperformed Haiku 4.5 on the benchmarks discussed. Haiku scored more than twice GPT-6 Luna on Terminal-Bench 4.0, reached 72.4% on OSWorld versus Luna’s 48.9%, and beat it on GDPVal and AA Briefcase. Its price per token is one quarter of Sonnet 5.5’s and half of Haiku 4.5’s, with an 80% discount for prompts under 100,000 tokens. Artificial Analysis estimated task costs of 21 cents at max inference and 12 cents at extra high, while cautioning that Haiku uses unusually many tokens. Its Intelligence Index score was 43, between Kimi K3 and GLM 5.3 Flash.

More capable, token-efficient low-cost models could reduce enterprise AI spending and pressure OpenAI’s and Anthropic’s revenue growth, according to Eric Arazzian. Expensive image and video models could increase spending, but their use is currently concentrated in social-media marketing and advertising; broader automation in construction and manufacturing remains a possibility rather than an established trend. Hasib Koreshi called the prospect of buyers choosing American labs for the cheapest models a “narrative violation.” The episode argues that US labs’ price cuts and efficiency gains have shifted the cost-and-intelligence frontier, though they do not resolve sovereignty concerns about Chinese open models.

Portability and avoiding vendor lock-in are also becoming product priorities. Hermes offers open, user-controlled agent software, while Noose’s stated principles include choosing models, controlling data and memory, deciding where an agent runs, inspecting prompts and tools, and reading or changing code. Hermes supports free bring-your-own-model use as well as subscriptions for frontier-model usage and hosted instances. It was reportedly valued at $1.5 billion after a funding round; the Wall Street Journal reported 22 million installations and about 2.5% of global token usage. Noose reported $36 million in annualized revenue so far this year and a $100 million year-end target, and plans to pursue enterprise customers with Hermes for Business.

SpaceX plans for Grokbot to route tasks to what Elon Musk calls the best backend, including Claude Opus 5.5, Midjourney and Suno, while using a fast version of Grok 4.8 for most simple requests when available. Anisha Sharia said SpaceX can combine its models and inference infrastructure with routing; Thomas Tungus argued that no company can provide the best model for every use case. The available choices are limited to providers with relationships to Musk or Grok, exclude OpenAI, and have prompted objections from users concerned about third-party API costs being passed to subscribers. OpenAI’s enterprise marketplace similarly lets customers direct part of their spending commitment to open-model inference. Together, these products point toward model choice and portability through an agent’s harness and context as priorities for businesses and power users.

The user experience is changing alongside model access. OpenAI says it is bringing a new GPT-6 model to free ChatGPT, which it says serves more than 1.2 billion people weekly, and adding an Intelligent UI that can generate interactive visualizations, charts and images alongside text. Its team built streamable components and a compiler to render interfaces as the model generates them, while training the model to choose between interactive elements and plain text. Early reactions were mixed: some users found the feature unclear or over-elaborate, while others welcomed an alternative to walls of text. Meta’s Muse expanded to iPad, added GitHub, Asana, QuickBooks and Canva connectors, and gained a feature for pursuing a high-level business goal autonomously; a Windows version was announced.

Access to capable systems and the data behind them raises distinct security and research questions. Anthropic opened applications for Mythos and reduced-guardrail versions of Opus and Sonnet through a cyber verification program, with tiers for defensive work, limited red-team testing and government-approved testing of high-risk systems. Specialized Access replaces and expands Project Glasswing and is the only tier with mandatory data retention. Separately, Biohub launched a $1.8 billion initiative with Google and the US government to create an open biological dataset and train models on complex systems, including entire cells, with the aim of helping scientists prioritize virtual experiments before physical lab work. Meanwhile, major AI infrastructure financing is drawing scrutiny: a reported $50 billion private-credit deal would finance Broadcom’s custom-chip production for OpenAI, and SpaceX is reportedly seeking $40 billion for Nvidia chips. Some lenders passed on SpaceX’s deal after receiving a short memo describing data centers to be built “somewhere in the universe”; reported credit-default-swap pricing implied a 15% chance of default by the end of 2031.