The Most Important Trends in New AI Products

As access to capable AI models spreads, model selection and product experience are emerging as key differentiators. The new products discussed range from interfaces that turn model responses into interactive displays to services that route requests among models and agents designed to give users control over their software and data. The trade-off is between convenience and control: choosing a single bundled product can simplify use, while routing and open systems offer more choice but raise questions about cost, complexity, and reliability.

OpenAI’s Intelligent UI can present answers as interactive visualizations, charts, and images alongside text. Issa Fulford said the company built streamable components and a compiler so interfaces can render while the model responds, and trained the model to choose between interactive elements and plain text. Andrew Chen said the challenge was making the interfaces feel native and fast without allowing them to take over the conversation. Early users disagreed on whether the feature clarified answers or added clutter, leaving its value dependent on the task and the user.

Grokbot’s announced approach makes model routing a visible part of the product: Elon Musk said it could use third-party services including Claude Opus 5.5, Midjourney, and Suno, while sending most simple requests to a fast version of Grok 4.8 when available. It also added tools to search, read, and monitor X. Users raised concerns about potential third-party API costs and the ability to opt out, and the selection remains limited because OpenAI models are excluded. Anisha Sharia described the combination of model lab, inference provider, and router as a possible business structure; Thomas Tunguz argued that no company can provide the best model for every use case.

Anthropic positioned Haiku 5.5 as its cheapest, fastest, and most capable small model, with results that suggest lower-cost models may be useful for more than simple tasks. It scored 72.4% on OSWorld, compared with 48.9% for GPT-6 Luna, and beat Luna on GDPVal and AA Briefcase; Artificial Analysis gave it an Intelligence Index score of 43. Haiku 5.5 costs one-quarter as much per token as Sonnet 5.5 and half as much as Haiku 4.5, with an 80% token discount for prompts below 100,000 tokens. It is unusually token-hungry, according to Artificial Analysis, though estimated task costs were 21 cents at max inference and 12 cents at extra high.

Ramp’s October rankings showed JEV gaining one percentage point of market share in a month, but the rankings reflect early adopters rather than the entire enterprise market. JEV’s creator, Type Safe AI, ranked first in growth relative to size and second in market-share growth; JEV was gaining share faster than every frontier lab except Anthropic. Eric Karazian said efficient, lower-cost models could pressure OpenAI and Anthropic’s revenue growth, while increased use of expensive image and video models could lift spending. Broader use of those models in construction and manufacturing remained a possibility, not an observed trend. OpenRouter and Featherless also gained share, and OpenAI’s enterprise marketplace lets customers direct part of a spending commitment toward inference from open models.

Hermes Agent makes user control and openness central: users can choose models, control data and memory, inspect prompts and tools, choose where the agent runs, and read or change its code. The Wall Street Journal reported 22 million installations and about 2.5% of global token usage. News Research reached a $1.5 billion valuation after a funding round and plans to use the money to launch Hermes for Business; it reported $36 million in annualized revenue so far this year and a $100 million year-end target. Hermes is free for users who bring their own models, with paid options that bundle frontier-model usage and hosted instances. Dylan Roelnik said the open ecosystem aims to provide trust, observability, and freedom from lock-in, and argued that reproducing software has become much cheaper.

Anthropic expanded access to Mythos through a cyber verification program after initially limiting it to selected Project Glasswing companies. The program has defense, red-team, and specialized tiers; the specialized tier requires government approval and is intended for tests of high-risk systems including flight control, power grids, and financial infrastructure. Nathan Lambert said an accidental open-weight release would accelerate cybersecurity risk rather than create a step change in societal risk, while Jamie Dimon said cyber risks had increased “tenfold” after Mythos and that AI had created previously unknown vulnerabilities. Separately, Meta made Muse available on iPad, added GitHub, Asana, QuickBooks, and Canva connectors, and introduced a small-business feature for pursuing a user’s high-level goal autonomously; a Windows version was announced as the next step.