Generative AI Digest: OpenClaw and Moltbook make a splash

Several new models were released this month, many of them featuring improved complex problem-solving, orchestration and task decomposition in a way that makes them better suited to agentic workflows. February was also a big month for Chinese startups MiniMax Group Inc., Zhipu AI (Knowledge Atlas Technology Joint Stock Company Ltd.) and Moonshot AI, all of which launched models that are rapidly closing the gap with Western rivals in specialized benchmarks, particularly in coding and agentic reasoning.

The loudest AI story of the month centered on OpenClaw, an open-source, locally running agent framework, and Moltbook, a Reddit-style social network built entirely for AI agents. Together, they ignited a debate about autonomy and safety and forced a shift in how we perceive AI systems. OpenClaw vaulted from a solo side project to one of the most widely deployed autonomous personal agents, characterized by agents with persistent identities that iteratively refine their behavior, enabled by local execution with real system-level permissions. This marks a transition from thinking of AI as a chat-based assistant to treating it as an operating-system-like execution environment that can act on users' behalf, without a human in the loop. That same shift echoed in enterprise launches like OpenAI's Frontier platform, where deeply integrated agentic stacks heighten the stakes. Persistent state and powerful tool access expand capabilities, but also collapse traditional safety boundaries, making threat modeling, sandboxing, permissioned scopes and continuous behavioral monitoring nonnegotiable rather than optional add-ons.

Product releases and updates

In February, OpenAI substantially reshaped its model lineup by launching GPT-5.3-Codex-Spark, a specialized coding model designed in partnership with Cerebras Systems Inc. to deliver faster generation speeds. This release coincided with the debut of OpenAI Frontier, an enterprise-grade platform that transitions AI from a conversational tool into a digital co-worker capable of autonomous execution across siloed business data. The product is positioned as a single control plane that connects agents to enterprise systems of record (e.g., data warehouses, CRMs, internal apps) via a shared business context and then runs them through an agent-execution layer with built‑in evaluation loops. To streamline this new agent-first ecosystem, the company also retired several legacy models, including the once-flagship GPT-4o, signaling that the industry has fully entered the era of reasoning-heavy and agentic AI.

In late January 2026, Moonshot AI introduced Kimi K2.5, positioning it as a major step beyond earlier Moonshot AI models by combining native multimodality with stronger agentic execution. It suggests this was achieved by Kimi K2.5 building on Kimi K2 with continued pretraining over approximately 15 trillion mixed visual and text tokens. Perhaps the most notable aspect of the announcement was an "Agent Swarm" research preview, which the Chinese startup suggested allows for the model to self-direct up to 100 sub-agents and coordinate up to 1,500 tool calls. The company also launched Kimi Code, an open-source coding tool.

Bytedance Ltd. unveiled Seedance 2.0, describing it as a unified multimodal, audio‑video joint generation system that can take text, image, audio and video inputs. The quality of motion depiction, prompt adherence and controllability featured heavily in the commentary around the release, and the model has proven popular in both China and the West. Concerns about AI-driven identity forgery emerged around the release, prompting ByteDance to announce that it would no longer allow realistic photos or videos of humans to be used as reference subjects. The company also announced the Doubao 2.0 large language model, with the announcement focusing on the model's agentic enablement capabilities.

An AI agent platform, ai.com, emerged as an advert during the Super Bowl. Little is known about the platform other than its claim that it will provide private, personalized agents and that the website domain was purchased by Kris Marszalek, the CEO of Crypto.Com, Inc. The stated differentiation by the platform is that its agents will be able to autonomously build out missing features and capabilities, as well as share these advances with other agents on the network.

Chinese model-maker Zhipu AI launched its fifth-generation flagship, GLM-5, a 744-billion-parameter Mixture-of-Experts (MoE) model. The model also establishes a new milestone for hardware independence; the company confirmed that GLM-5 was trained and serves inference entirely on domestic Huawei Ascend and other Chinese-made chips, proving that frontier-level intelligence can be achieved without access to restricted NVIDIA Corp. hardware. The model marks a significant leap in agentic engineering, and the release narrows the gap with Western leaders like GPT-5.2 and Claude Opus 4.5, while offering open-weight accessibility under an MIT License.

FourKites Inc. has launched Loft, an AI-native orchestration platform designed to transition supply chain management from data visualization to autonomous task execution. The system utilizes an AI developer agent named Sophie to convert natural language instructions into functional automations by integrating internal enterprise data with the FourKites Intelligent Network. To maintain transparency and recordkeeping, the platform utilizes agent operating procedures to document the logic behind AI-driven actions, such as purchase order reconciliations or supplier escalations.

Alibaba Group Holding Ltd. introduced Qwen3‑Max‑Thinking as its latest flagship reasoning model. The model is proprietary, not open-weight, unlike those of some prominent Chinese AI competitors, and is listed with a context window of approximately 256,000/262,000 tokens. Available in Qwen Chat and via API, the release messaging emphasizes stronger general reasoning alongside adaptive tool use.

Over the past few weeks, Google LLC began rolling out Project Genie, a Google Labs prototype that exposes DeepMind's Genie 3 world model to Google AI Ultra subscribers in the US. Users can develop worlds from text or images, then explore and remix interactive scenes, with sessions capped at 60 seconds. Its release impacted several video game stocks. Google DeepMind disclosed Aletheia, an internal math research agent built on Gemini Deep Think, designed to iteratively generate, verify and revise research-grade solutions.

Chinese startup MiniMax released its M2.5 and M2.5 Lightning models, utilizing a sparse MoE architecture. While the total parameter count is 230 billion, the models only activate 10 billion parameters per token, allowing them to achieve performance levels on the SWE-Bench Verified benchmark comparable to flagship models like Claude 4.6. This efficiency enables the standard version to operate at lower costs, with output pricing roughly 1/60 that of rival frontier models. The Lightning variant further optimizes for speed, delivering a stable throughput of 100 tokens per second to facilitate real-time agentic workflows.

Waymo LLC introduced its Waymo World Model, a simulation platform built on Google DeepMind's Genie 3 that generates hyper-realistic 3D driving environments. Unlike traditional simulators limited by real-world data, this system leverages Genie 3's vast world knowledge to invent long-tail scenarios — such as flooded streets or rogue wildlife — that a vehicle fleet might never encounter in person. By converting 2D video intelligence into 3D lidar and camera outputs, the model allows the Waymo driver to practice what-if maneuvers in a safe, virtual space at accelerated speeds.

Robbyant, an embodied AI unit within Ant Group Co. Ltd., announced the open-source release of LingBot‑World, positioning it as a real‑time "world model" designed for embodied intelligence (intelligence that can interact physically within its environment), autonomous driving simulation and game development. The company frames LingBot‑World as a controllable "digital sandbox," emphasizing long-horizon stability to address video-generation drift and claiming up to nearly 10 minutes of continuous, coherent generation.

Yahoo Inc. has unveiled Scout, its answer to search players like Perplexity AI Inc., OpenAI and Google, which are increasingly looking to disrupt the search game. The new platform will enable conversational searching and summarization. While it is not the first answer engine to come to market, Yahoo still boasts a sizable presence in search and email that gives the pivot some legs. Yahoo plans to drill specifically into personalization opportunities in AI, given its existing trove of data from its users and consumers. The company is simultaneously introducing the Scout Intelligence Platform, which will bring generative capabilities to other pieces of its product suite (e.g., Yahoo Mail, News, Finance and Sports), enabling features like email summaries and game analysis.

Anthropic released two major models in February. The release of Claude Opus 4.6 marks a leap in its utility with agentic AI, in that it allows the model to split complex tasks into parallel sub-processes via new Agent Teams and now includes a 1-million-token window and conversation compaction to prevent memory drift. It has been optimized for stronger coding, planning and long-task performance. This release was shortly followed by Claude 4.6 Sonnet, the company's mid-tier model, intended to balance capability, latency and cost for broader deployment. This update included Real-time Recursive Reasoning, a feature that allows the mid-tier model to refine its logic internally before responding.

Funding and M&A

Anthropic closed a $30 billion series G, valuing the company at $380 billion post-money, led by GIC and Coatue. This figure included a portion of previously announced Microsoft Corp. and NVIDIA commitments. As part of the announcement, Anthropic stated that its run-rate revenue is $14 billion, a figure that has grown over tenfold annually in each of the past three years. This represents the second-largest venture funding deal of all time, behind only OpenAI's $40 billion round in 2025.

AI audio startup Eleven Labs Inc. — best known for its text-to-speech and voice-generation capabilities — announced a $500 million series D led by Sequoia Capital at an $11 billion valuation. The round comes about a year after ElevenLabs' $180 million series C.

xAI has been folded into SpaceX via a share‑exchange/all‑stock acquisition, pitching a tighter, vertically integrated structure. It appears that the brands and operations will remain at least partially separate in the near term. In the wake of the SpaceX tie‑up and a subsequent reorganization, a number of senior leaders departed xAI, including co‑founders Tony Wu and Jimmy Ba.

World Labs Technologies Inc., the spatial intelligence startup founded by Fei-Fei Li, secured $1 billion in new funding at an undisclosed valuation, though industry sources peg it at around the $5 billion mark. This significant investment round included backing from major tech companies such as NVIDIA, Advanced Micro Devices Inc., and Autodesk Inc. The capital will accelerate the development of large world models designed to understand and interact with the 3D physical world.

Mistral AI SAS inked its first-ever acquisition, buying French cloud startup Koyeb. By integrating Koyeb's serverless infrastructure, Mistral will augment its Mistral Compute platform, allowing developers to deploy and scale AI models without managing underlying servers. This supports emerging use cases, such as agents, which involve less predictable infrastructure needs. This acquisition follows Mistral's massive €1.2 billion investment in Swedish data centers, signaling a concerted effort to build a sovereign European AI cloud that can compete directly with US hyperscalers. Mistral will bring on Koyeb's 13-person engineering team, including its three co-founders, as the company targets $1 billion in revenue for 2026.

Politics and regulations

ByteDance responded to media inquiries surrounding threats of legal action from Walt Disney Co. following the release of its latest video model, with a brief statement acknowledging IP concerns and pledging to strengthen safeguards. Clips generated by Seedance 2.0, re-creating popular film scenes and characters, were widely distributed online, prompting Disney to send a cease-and-desist letter.

The December 2025 US Executive Order establishing a "minimally burdensome national AI policy framework" has gained teeth with the AI Litigation Task Force, operational since Jan. 9. As states including California, New York, Texas and Illinois have not paused enforcement of AI laws, tensions have continued to escalate. Another safety tension point has arisen around Anthropic's relationship with the US Department of Defense, as the model company has resisted requests to extend the areas in which permits the technology to be applied.

The UK government brought AI chatbots under the Online Safety Act's illegal-content duties. This move appears to have been accelerated by xAI's Grok chatbot producing sexualized images, and has been presented as "closing a loophole" in the government's strategy to "keep children safe online."

451 Research from S&P Global Energy Horizons provides technology industry research, data, and advisory solutions. For more information or to contact us, please visit 451 Research.
This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global.