Today's briefs
Google Releases Gemini Omni 1.2 Flash With Support for 15-Second Video Clips at $0.07/Second
Google has launched Gemini Omni 1.2 Flash, an updated multimodal model that extends video understanding to 15-second clips at a cost of $0.07 per second of video processed. The release emphasizes speed and cost-efficiency, positioning it as the go-to option for developers who need fast, affordable video analysis at scale. Key improvements include longer clip support and optimized throughput compared to prior Flash variants, making it practical for applications like video search, content moderation, and media summarization. Developers can access the model via the Gemini API today, with pricing competitive enough to enable high-volume production workloads. This release reinforces Google's strategy of offering a tiered model lineup where Flash handles cost-sensitive, latency-critical use cases.
Google DeepMind
NVIDIA Releases Cosmos 3 Open Models for Interactive World Simulation
NVIDIA has released Cosmos 3, a new set of open-weight models on Hugging Face designed for interactive world simulation, enabling developers to generate physically plausible, controllable 3D environments from prompts or partial observations. The models are built to support robotics, autonomous vehicle training, and game simulation pipelines, extending the original Cosmos family with improved interactivity and scene controllability. Because the weights are openly available, developers can fine-tune Cosmos 3 for domain-specific simulation tasks without depending on a proprietary API. This is a significant release for anyone building synthetic data pipelines, sim-to-real robotics workflows, or interactive environment generators. The open release lowers the barrier for research teams and startups to access frontier-class world simulation capabilities.
NVIDIA
Together AI Releases Together Compiler: Agentic Continual Compilation Makes Inference Up to 2.6x Faster
Together AI has launched Together Compiler, a system that uses AI agents to perform continual compilation of inference workloads, achieving up to 2.6x speedups on model serving without changing model weights or architecture. The system works by profiling running inference jobs and iteratively recompiling execution graphs to match the observed request distribution, making it particularly effective for production workloads with variable input patterns. This is a developer-facing infrastructure tool available through the Together AI platform, targeting teams running large-scale LLM inference who want to reduce latency and cost without swapping models. The agentic compilation loop means performance improves automatically over time as the system adapts, requiring minimal manual tuning. For teams cost-sensitive about inference spend, this represents a meaningful lever beyond just model quantization or batching strategies.
MarkTechPost
Thinking Machines Releases Inkling 2 Small: A 594B MoE Model With Only 17B Active Parameters
Thinking Machines has released Inkling 2 Small, a 594-billion parameter mixture-of-experts model that activates only 17 billion parameters per forward pass, targeting the efficiency-performance frontier for open-weight models. The sparse activation design allows the model to deliver strong reasoning and instruction-following quality while keeping inference compute comparable to much smaller dense models. This positions Inkling 2 Small as a compelling option for teams that want large-model capability without the serving cost of a fully dense frontier model. Developers can evaluate the model for tasks requiring strong generalization — such as coding, math, and multi-step reasoning — at a fraction of the typical compute budget. The release adds meaningful competition in the growing MoE open-weight space alongside models like Mixtral and DeepSeek.
MarkTechPost
Claude Gains Direct Bank Account Access via Plaid Integration
Anthropic has integrated Claude with Plaid, enabling Claude-powered agents to directly connect to users' bank accounts and act on personal financial data. This marks a significant expansion of Claude's agentic footprint into high-stakes personal finance workflows, including account balance queries, transaction analysis, and potentially financial planning or action execution. For developers building personal finance tools, this integration provides a sanctioned, structured path to combine Claude's reasoning with real-time financial data via Plaid's established connectivity infrastructure. The move raises the bar for what AI agents can do in production consumer contexts, but also surfaces important questions about authorization scope, auditability, and user consent that developers will need to design around carefully. This is one of the clearest signals yet that top-tier AI labs are moving agents from sandboxed demos into live, consequential real-world integrations.
Anthropic
Perplexity Launches Teammates: AI Agents Designed for Workplace Task Execution
Perplexity has launched Teammates, a suite of AI agents built for workplace productivity tasks including research, drafting, scheduling coordination, and information retrieval across connected work tools. Unlike Perplexity's core search product, Teammates is designed to take autonomous action across integrated SaaS applications, positioning it as a direct competitor to agent platforms from OpenAI, Microsoft, and Google in the enterprise productivity space. The product targets teams that want AI agents embedded in their existing workflows without building custom tooling from scratch. For developers evaluating agent platforms, Teammates represents a new entrant with Perplexity's search-grounded retrieval as a differentiator, potentially giving agents more accurate real-world context than pure LLM-based competitors. Early access details and integration support are available through Perplexity's enterprise channel.
Perplexity
OpenAI Launches GPT-Image-2.6 Flare and Sunburst: Adds GIF Output, SVG Export, and 6-Second API Generation
OpenAI has released two new image generation model variants — GPT-Image-2.6 Flare and Sunburst — adding meaningful new output capabilities including animated GIF generation, SVG export, and a 6-second end-to-end API generation target. These features meaningfully expand what developers can build on top of OpenAI's image API, enabling animated asset creation and scalable vector outputs suitable for UI design, marketing automation, and content pipelines. SVG export in particular is notable for developers building design tooling, as it produces resolution-independent graphics that can be edited programmatically downstream. The 6-second generation target makes the API viable for more interactive, near-real-time applications than prior generations allowed. Developers already using GPT-Image APIs can access the new variants through the existing endpoint with updated model selectors.
OpenAI Blog
Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE Model With 14B Active Parameters
Baidu has open-released ERNIE 5.5 Flash Base, a 280-billion parameter mixture-of-experts language model that activates 14 billion parameters per forward pass, targeting efficient large-scale inference for Chinese and multilingual applications. The model follows a similar sparse-activation design philosophy to other large MoE releases, allowing high-capacity reasoning at comparatively low serving cost. ERNIE 5.5 Flash Base is released as a base model, meaning developers will need to apply instruction tuning or RLHF fine-tuning for specific task alignment, making it most relevant for research teams and enterprises with custom fine-tuning pipelines. Its release adds to the rapidly growing pool of open large MoE models from non-Western labs, increasing competitive pressure on the open-weight frontier. Developers building multilingual or East Asian language applications now have an additional high-capacity open option to benchmark against.
MarkTechPost
NVIDIA's Vera CPU Architecture Delivers Highest Throughput and Lowest Latency for AI Agent Workloads
NVIDIA has detailed its Vera CPU architecture, designed specifically to maximize throughput for AI inference and agent execution workloads in data center deployments. Vera is positioned as the CPU layer of NVIDIA's full-stack AI computing platform, optimized to keep up with GPU-side throughput by reducing CPU-side bottlenecks that commonly throttle agentic and multi-step inference pipelines. The architecture targets large-scale deployments where agents make rapid sequential decisions, making CPU-GPU bandwidth and latency a first-order concern rather than an afterthought. For infrastructure engineers designing AI serving stacks, Vera represents a purpose-built option to pair with NVIDIA's GPU lineup rather than relying on general-purpose server CPUs. This signals NVIDIA's intent to own the full compute stack for AI workloads, not just the accelerator layer.
NVIDIA
Agility Robotics Makes Nasdaq Debut as Humanoid Robot Market Faces Investor Scrutiny
Agility Robotics, maker of the Digit bipedal humanoid robot, has begun trading on the Nasdaq, marking one of the first major public market entries for a humanoid robotics company. The IPO invites fresh scrutiny of the broader humanoid robot market, with analysts examining whether commercial deployment timelines and unit economics justify current valuations across the sector. Agility's Digit robots have been piloted in warehouse and logistics environments, with Amazon among its notable partners, giving the company a real deployment track record — though scale remains limited. For developers and engineers working on physical AI and embodied agents, the IPO raises the question of whether the capital markets will fund the long development cycles required to bring humanoids to production viability. The debut adds public market data points to a space that has been largely private, potentially informing how robotics AI startups are valued going forward.
Ars Technica