New AI Models Focus on Cost, Autonomy, and Multi-model Use

With 22 million installations and roughly 2.5% of global token usage linked to the Hermes agent, Nous Research has achieved a $1.5 billion valuation.

The company, which initially drew inspiration from 'OpenClaw,' developed its own distinct path, combining AI research with elements of performance art by building large language models through distributed, anonymous GPUs.

Significant growth and user adoption were primarily driven by the release of its Hermes agent, which now accounts for 22 million installations and approximately 2.5% of global token usage. The new funding will facilitate the launch of 'Hermes for Business,' targeting enterprise clients.

The marginal cost of software reproduction has plummeted, according to CEO Dylan Rolnick, who argues that this shift makes vendor lock-in obsolete.

Rolnick's manifesto emphasizes that users deserve observability and transparency into AI systems that integrate deeply into their lives, with 'non-negotiables' including user choice over models, data privacy, and the capability to run AI locally. The core philosophy dictates that an agent should remain entirely under the user's control and modification.

Oracle, Broadcom, and SpaceX are pursuing private debt arrangements to fund massive chip acquisitions, opting to sidestep traditional public bond markets.

Oracle is actively pursuing private funding from Apollo and Goldman Sachs for chip acquisitions, while Broadcom is reportedly negotiating a $50 billion deal with Apollo and Blackstone to finance custom chip production for OpenAI. SpaceX is also seeking a $40 billion loan, led by Apollo, for NVIDIA chip purchases, although some lenders have expressed skepticism due to insufficient documentation.

Despite these concerns, Bloomberg reported that Pimco is considering the financing deal. SpaceX's credit default swap pricing currently implies a 15% probability of default by 2031, despite proponents insisting that backing Musk’s vision takes precedence over standard investor disclosure.

This project aims to facilitate virtual experiments, reducing the necessity for physical lab work, with Biohub’s Alex Rivz emphasizing that higher-quality reasoning in lab questions will significantly increase the value of empirical results. The initiative represents a crucial infrastructure buildout for AI in biology, focusing on standardizing data at scale.

The program features three distinct access tiers: Defense, designed for incident response and malware analysis; Red Team, for penetration testing; and Specialized, which requires government approval. The Specialized tier is unique in that it mandates data retention policies, unlike the other tiers.

Under this new system, Anthropic provides authorized professionals with versions of its Opus and Sonnet models that have reduced safety guardrails, allowing for more specific and sensitive cybersecurity applications.

Researcher Nathan Lambert states that an accidental model release would amplify existing risks, while JP Morgan CEO Jamie Dimon claimed that cybersecurity threats surged tenfold following the introduction of Mythos. Despite his warnings, Dimon advocates for practical mitigation efforts over alarmist rhetoric.

This update also brings autonomous, goal-oriented workflows designed specifically for small businesses, enabling Muse to work independently towards a high-level business objective. Additionally, a Windows port of Muse was announced at a recent hardware event.

Although the AI industry lacks traditional seasonality, this concentration of new releases reflects broader trends that are currently shaping the AI landscape and dictating future development.

Jev, developed by Typesafe, leads in growth relative to its size and market share, rapidly gaining ground alongside Anthropic in foundation LLMs, while SpaceXI/Grockbot is also showing significant enterprise traction for the first time. Growth rankings show the emergence of model routers like Open Router and Featherless, marking a structural transition toward advanced model architectures.

Eric Kerazian noted that the introduction of highly performant standard and light models, along with token-efficient models designed for specific enterprise use cases like Jev, is poised to challenge the revenue growth of OpenAI and Anthropic among businesses.

Scale AI is showing renewed growth, indicating an increased demand for reinforcement learning with human feedback (RLHF) or fine-tuning needs in the AI sector.

The emergence of Runway and Higsfield demonstrates growing enterprise interest in expensive image and video generation models, which currently impact social media marketing and are expected to see future adoption in construction and manufacturing automation. Eric Kerazian pointed out that these use cases are comparatively expensive relative to the coding and language model applications that have historically dominated enterprise AI adoption, thereby pushing AI spending upwards.

To balance speed and intelligence, Grok will maintain a proprietary 'lightning-fast' model for simpler requests. This decision has sparked concerns among some users regarding potential additional API costs and philosophical opposition to integrating models from other labs.

The strategy allows xAI to capture market share through inference while simultaneously developing and maintaining its core models, aiming to deliver the best possible results for users.

Investor Anisha Sharia characterized xAI’s Grok strategy as a method to 'eat their cake and have it too,' highlighting how the company benefits from both being a lab and acting as a 'cursor-style' router.

This approach allows xAI to avoid being dictated by other labs by maintaining its own state-of-the-art model, while simultaneously capturing margins on third-party inference. Sharia noted that this creates a win-win scenario, ensuring the platform consistently benefits regardless of which model is used.

The AI industry is rapidly trending towards multi-model ecosystems as companies seek to avoid vendor lock-in, with businesses and power users increasingly demanding the flexibility to swap models without changing products.

OpenAI is also moving in this direction, selling open-model inference through its B2B marketplace, though current integrations often depend on existing company relationships, such as Anthropic utilizing xAI data centers. This trend signifies a broader desire for more choice and less restrictive model environments.

Haiku 3.5 scored 72.4% on OS World, outperforming GPT-4o Luna’s 48.9% in benchmarks covering coding, computer use, and general knowledge tasks. Artificial Analysis further rated Haiku 3.5 with an impressive 43 on its intelligence index, confirming its competitive edge.

To encourage efficient, quick queries, an 80% discount is applied to smaller prompts under 100,000 tokens. Artificial Analysis benchmarks confirm that this new Haiku is highly competitive with GPT-4o-mini and substantially cheaper than the full GPT-4o.

Investor Hibeshi stated that this release represents a 'narrative violation,' as the market for the cheapest and most efficient models is now dominated by US labs rather than previous leaders like Chinese competitors such as GLM or DeepSeek.

Enterprises must now weigh whether their primary concern is purely cost efficiency or the imperative to run models on local hardware to guarantee data privacy. The effective price parity achieved by US labs shifts the burden to China to respond to this evolving intelligence frontier.

Ethan Mollick highlighted a long-term trend towards 'just-in-time' interfaces that are built specifically for a user's immediate problem, signaling a fundamental shift in product interaction.

As raw model intelligence becomes increasingly commoditized, product experience is emerging as a primary differentiator. Foundation labs are moving away from unintuitive, text-heavy products towards design-led, context-aware interactions that generate interfaces on demand.