carl.franzen@venturebeat.com (Carl Franzen)

Meta has launched Muse Voice Transcribe, a real-time speech-to-text model priced at $0.18 per hour of audio that combines transcription, speaker diarization for 20+ speakers, and endpoint detection in a single model. The system supports over 70 languages, with 25 extensively validated for release, and handles multilingual code-switching without separate post-processing. While competitors like Speechmatics support higher speaker counts (up to 100), Meta's combination of low latency, high-capacity diarization, and aggressive pricing targets enterprise developers building meeting systems, call analytics, and ambient AI applications.

Anthropic released Claude Fable 5.1 and Mythos 5.1, its latest large language models, alongside a 75% cost reduction for cached context reads and a new Enterprise Frontier Safeguards security architecture. The release targets enterprise deployment of persistent agents capable of multi-hour problem-solving tasks. Fable 5.1 shows significant benchmark improvements across scientific research, coding, and business workflow tasks, though results are vendor-reported rather than independently verified.

OpenClaw 2.0, an open-source AI coding harness, has shifted from a personal developer tool to enterprise-focused infrastructure with multi-user collaboration, shared cloud sessions, and enhanced security controls. The update includes a redesigned browser interface that consolidates conversations, files, approvals, and agent activity into a single workspace. Creator Peter Steinberger reports the team built the update using OpenClaw itself, moving away from individual local coding environments toward shared team workspaces.

NanoCo has released a Slack integration for NanoClaw, its open source AI agent framework, allowing users to create and manage teams of persistent AI agents directly from Slack messages. Each agent gets its own identity, avatar, role, memory, and permissions, functioning more like a digital department than a single chatbot. The integration simplifies setup compared to traditional Slack bot deployment and allows agents to work together in channels and across other messaging platforms like Telegram and WhatsApp.

Serval made its Catalyst AI agent generally available Thursday, enabling it by default for customers. Catalyst acts as an administrative layer that can discover repetitive IT work from ticket history, automatically build workflows to handle it, and create background agents that proactively identify and fix problems before tickets are filed. The distinction from competitors like ServiceNow and Atlassian is Catalyst's ability to move from opportunity discovery through deployment of governed automation to continuous proactive agent creation.

Block, the company behind Square and Cash App, has open-sourced Berd, a desktop application for managing AI agents across different models and tools. Originally built for Block's internal use, Berd is now available under Apache 2.0 license for macOS, Windows and Linux, with version 0.6.2 released on August 18. The application stores conversation history locally and emphasizes visible operational state, configuration transparency, and persistent projects rather than functioning as a generic chatbot wrapper.

Alibaba released Qwen3.8-27B, a 27-billion-parameter open source model on Friday that runs locally without cloud APIs and delivers frontier-class coding and reasoning capabilities. Third-party benchmarks show it matches or exceeds proprietary models from months ago, with scores equivalent to OpenAI's GPT-5.6 Luna and outperforming Claude Opus 4.8 on agentic tasks. The model runs on consumer hardware when quantized to 4-bit, making frontier-class AI accessible without vendor dependency.

DeepSeek launched DeepSeek-V4-Pro, an updated flagship model for agentic workloads, alongside DeepSeek Harness v0.1, an open-source agent framework available under MIT license. The releases position DeepSeek as a competitor to Anthropic's Claude Code and OpenAI's Codex by offering developers an alternative agent infrastructure layer. Simultaneously, DeepSeek is shifting from flat API pricing to peak and off-peak rates starting August 16, with substantially higher prices across the board.

SpaceXAI released Grok 4.6, scoring 61 on Artificial Analysis Intelligence Index and tying OpenAI's GPT-5.6 Sol for third place globally. The model targets long-running agents, coding, and knowledge work with pricing starting at $2 per million input tokens and $6 per million output tokens, less than half the cost of GPT-5.6 Sol standard mode. The release emphasizes improvements in agent behavior and task persistence rather than isolated benchmark gains.

SpaceXAI, the rebranded xAI division, is launching Grok Bot in beta, an AI agent designed to operate continuously across enterprise software applications rather than respond to individual prompts. The product costs $120 per seat per month for teams and $200 for individuals, positioning it as a persistent digital coworker that can execute workflows autonomously. The launch arrives as Anthropic, OpenAI, and other competitors expand their own agentic capabilities for enterprise use.

LTX released LTX-2.5, an open-weights video generation model that produces 10-second clips in 6.8 seconds on Nvidia GB200 chips, with native multishot support and improved quality. The model is available free for organizations under $10 million ARR on Hugging Face, ComfyUI, and via API. LTX claims 33 million downloads across its model family and reports a 67% win rate in blind quality tests against competing models.

Meta released Muse Glimmer, a 30-billion-parameter open-weight AI model licensed under Apache 2.0, designed to run autonomous agents on consumer hardware like high-end Macs and PCs. The release marks Meta's return to fully open source after shifting to proprietary models in April, and comes with fewer restrictions than Meta's previous Llama family. Meta also announced plans to open-source Muse Spark 1.2, its frontier model powering the recently launched Muse Code terminal agent.

Liquid AI, a startup founded by former MIT computer scientists, released LFM2.5-2.6B, a 2.6 billion parameter language model designed to run on edge devices including Raspberry Pi without cloud infrastructure or GPUs. The model supports 128,000-token context windows and native tool calling, targeting agentic tasks like document management and workflow automation in regulated industries and connectivity-limited environments. Performance ranges from 30 tokens per second on smartphones to 220 tokens per second on Apple M5 Max, with the model available on Hugging Face under a custom open-weight license.

Hark, a secretive AI startup founded by serial entrepreneur Brett Adcock, launched Handoff, a computer use agent that autonomously navigates websites to complete tasks like booking flights or ordering food. The company claims top performance on the Online-Mind2Web benchmark with a 97.7 score, significantly outpacing GPT 5.4, Claude Opus 4.8, and Gemini 2.5 Pro, while pricing at less than one-tenth the token cost of competing models. Public sign-ups opened today with availability planned for later this month, though critical questions remain about how Handoff performs against the current generation of frontier models.

Alibaba's Qwen team released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model targeting autonomous software engineering and enterprise automation. The company claims the model outperforms GPT-5.6 Sol Max and Fable 5 on agentic computing benchmarks, particularly on OSWorld-Verified (86.1 vs 83.2 and 85.0 respectively). Alibaba plans to release open weights next week, though licensing terms remain undisclosed, which could reshape enterprise adoption if permissive.

OpenAI has cut prices on two models in its GPT-5.6 series: Luna by 80% to $1.40 per million tokens combined, and Terra by 20% to $14 per million tokens combined, while introducing a premium Fast mode for its flagship Sol model at double the standard price. The moves come days after Anthropic released Claude Opus 5 at competitive pricing and Google launched lower-cost Gemini models, signaling a shift in AI competition toward cost and speed rather than capability alone. Luna now competes directly with the market's low-cost inference tier, though it remains more expensive than some alternatives like DeepSeek's flash model.

Nimble, a New York City-based startup, launched Web Search Agents designed to reduce token consumption by 51% while improving retrieval accuracy by 21% compared to leading AI search alternatives. The system uses self-learning retrieval algorithms and domain-specific optimization to help AI agents perform web research more efficiently for enterprise workloads. Rather than competing as a general search engine, Nimble targets developers building autonomous agents that need continuously updated information for research, lead generation, and compliance tasks.

General Motors redesigned its autonomous vehicle engineering workflows around AI agents rather than simply adding coding assistants, resulting in a tripling of merged pull requests and faster releases with fewer defects. GM's VP of autonomous vehicles noted that engineers spend only 15% of their time writing code, with agents now handling analysis, triage, experimentation, and testing across the remaining 85% of development work. The company connected agents to internal tools and petabytes of data through customized Model Context Protocol servers and version-controlled instruction documents.

Moonshot AI released full model weights for Kimi K3, a 2.8 trillion-parameter open model with a one million-token context window and frontier benchmark performance. The release includes infrastructure for self-hosting, but comes with a custom license that imposes restrictions on larger companies and AI service providers not found in traditional open-source licenses. Enterprises with over 20 million dollars in annual revenue operating a Model as a Service business must negotiate a separate agreement with Moonshot AI before commercial deployment.

Black Forest Labs launched FLUX 3, a multimodal AI model capable of generating images and video with audio up to 20 seconds from a single prompt, along with robotic vision and action capabilities. The model is entering limited early access with pricing and full benchmarks still unannounced, and open-weight versions will arrive later this year. Early preference testing shows FLUX 3 outperforming competitors like Luma Ray 3.2 and Runway Gen-4.5, though those results are labeled preliminary.

OpenAI has launched Presence, an enterprise platform for deploying and managing real-time voice and chat AI agents in production environments. The product is available through limited general availability with support from OpenAI Forward Deployed Engineers and select systems integrators, not on a self-service basis. Presence bundles policies, system connections, guardrails, evaluation tools and escalation processes to help enterprises run reliable agents across customer-facing and internal workflows.

Half of enterprises have deployed AI agents that passed internal evaluations but still failed in production, yet 66% are expanding autonomous deployment without human review. Only 5% trust their automated evaluation systems, creating a widening gap between the speed of agent autonomy and the assurance mechanisms to govern it. The mismatch reflects a broader pattern where companies ship agents first and retrofit control layers later.

Square has launched ChatGPT and Claude integrations that let restaurants accept orders placed directly within these AI platforms, with automatic enrollment and no marketplace commission fees. Restaurants still pay Square's standard online transaction processing fee of 2.9% plus 30 cents per transaction, significantly undercutting the 15% to 30% commissions charged by DoorDash, Uber Eats, and Grubhub. The move addresses a critical pain point for restaurant operators whose thin margins are squeezed by aggregator fees.

Google launched Nano Banana 2 Lite, a lightweight image generation model built on Gemini 3.1 Flash-Lite architecture, capable of generating 1k resolution images in 4 seconds at $0.034 per 1,000 images. The model is available immediately to enterprise developers through Google AI Studio, the Gemini API, and GEAP. It trades resolution flexibility for speed and cost efficiency, targeting high-throughput commercial workflows like programmatic advertising and e-commerce asset generation.

DeepSeek has open-sourced DSpark, an MIT-licensed framework that accelerates large language model inference by up to 85% without altering model outputs. The system uses speculative decoding, where a smaller draft model predicts likely token sequences that a larger model then validates, reducing computational overhead. DeepSeek has released technical papers, model checkpoints, and training code via GitHub and Hugging Face, making the technique available to researchers and enterprises running open-weight models.

Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter coding model that was previously operating anonymously as 'Owl Alpha' atop OpenRouter's developer rankings. The model was trained entirely on over 50,000 Chinese ASICs rather than Nvidia GPUs, demonstrating that near-frontier AI systems can scale without U.S. hardware. Released under an MIT license with aggressive pricing (promotional rates of $0.30/$1.20 per million tokens in/out), it arrives as Western governments restrict access to frontier models from OpenAI and Anthropic.

Mindstone, a London-based AI startup, launched Rebel this week, an agentic AI operating system that uses local markdown files to store agent memory and instructions. The platform automatically routes tasks to appropriate AI models, switching between local and cloud options based on data sensitivity and cost. Rebel operates under a Fair Source license, free for teams under 100 users, and has raised $5 million from investors including Pearson Ventures and Moonfire Ventures.

OpenAI and Broadcom unveiled Jalapeño, a custom AI inference chip designed specifically for large language model serving, completing development in nine months by using OpenAI's own models to accelerate chip design. The ASIC represents OpenAI's strategic move to build its own computational stack and address severe unit economics, with the company operating at a $20.92 billion loss in 2025 despite $13.07 billion in revenue. OpenAI plans to deploy Jalapeño across data centers by year-end and has already tested it with GPT-5.3-Codex-Spark in production environments.

Sakana AI launched Fugu, a multi-agent orchestration system that routes queries across a pool of specialized AI models through a single API, positioning it as an alternative to monolithic models after Anthropic restricted access to Claude Fable 5 and Claude Mythos 5 due to U.S. export controls. The system matches frontier-level performance on benchmarks while abstracting model selection and coordination from users. Sakana offers two tiers: standard Fugu for everyday tasks and Fugu Ultra for complex work, with pricing based on underlying model usage or fixed rates.

Z.ai released GLM-5.2, a 753-billion parameter open-weights LLM that outperforms OpenAI's GPT-5.5 on multiple long-horizon coding benchmarks while costing one-sixth as much. The model features a 1-million-token context window and is available under an MIT license for local deployment, positioning it as an alternative for enterprises concerned about U.S. regulatory restrictions on proprietary AI models.

Sakana AI, a Tokyo-based startup, launched Sakana Marlin, an autonomous research agent that generates 100-page strategy reports over 8 hours rather than seconds. The product targets enterprises, financial institutions, and think tanks with a pay-as-you-go pricing model. Marlin represents a shift in enterprise AI from speed-focused generation to deep, methodical reasoning using the company's Adaptive Branching Monte Carlo Tree Search technology.

NanoClaw and JFrog have launched an integration that routes autonomous AI agents through vetted software registries to block malicious code downloads. The system acts as an automated immune system, intercepting compromised packages and guiding agents to approved alternatives. The partnership offers free access for open-source users and commercial licensing for enterprises, addressing a growing security gap as AI agents autonomously install packages without human oversight.

Xiaomi has open-sourced MiMo Code V0.1.0, a terminal-native AI coding assistant that claims to outperform Anthropic's Claude Code on long-horizon, multi-step coding tasks (200+ steps) according to internal benchmarks. The tool uses a cross-session memory system with SQLite FTS5 to retain context across extended work sessions, addressing a core limitation of existing AI coding agents. Xiaomi is also offering limited free access to MiMo-V2.5, its flagship model with a million-token context window.

Anthropic CEO Dario Amodei published a policy essay calling for FAA-style government regulation of powerful AI models, arguing that mandatory third-party testing and potential deployment blocks are necessary as AI capabilities grow. The company released two policy frameworks alongside its most powerful model yet, Claude Fable 5, and a gated version with advanced cyber capabilities. The proposal would require models trained with over 10^25 FLOPs or developed by companies with $500 million in AI revenue to undergo mandatory testing, with regulators able to block or delay deployment if severe risks are identified.

Researchers from UIUC, UC Berkeley, and Chroma released Harness-1, a 20-billion parameter open-source search agent that scores 73% on information recall benchmarks, outperforming GPT-5.4 (70.9%) and other proprietary models. The model is available under Apache 2.0 license on Hugging Face. Harness-1 achieves its performance by offloading search session management to a structured software environment rather than relying on expanded context windows, suggesting that model efficiency matters more than raw parameter size for autonomous retrieval tasks.

Anthropic reports that over 80% of its production code merged in May 2026 was authored by Claude rather than humans, representing an 8x increase in code shipped per engineer per quarter since 2021-2025. The company attributes this shift to advances in autonomous agents capable of independent debugging and multi-hour task delegation. Anthropic outlines a three-step transition model for enterprises seeking to replicate this automation, moving from developer-centric coding to architectural oversight and autonomous code factories.

Google released Gemma 4 12B, an 11.95-billion-parameter open-source model that runs entirely on a standard 16GB enterprise laptop without requiring cloud connectivity. The model uses an encoder-free architecture that processes audio and video directly without secondary processing modules, reducing latency and memory overhead. It includes a 256K token context window, native tool-use capabilities, and step-by-step reasoning mode, making it suitable for enterprises with strict data privacy requirements.

Alibaba released Qwen3.7-Plus, a multimodal AI model supporting text, video, and image inputs at $0.40/$1.60 per 1M tokens, 60% cheaper than its text-only predecessor Qwen3.7-Max. The model marks a strategic shift away from Alibaba's open-source focus, available only through proprietary APIs and closed commercial licensing. It includes a 1-million token context window and a 'preserve_thinking' parameter to maintain reasoning state across multi-step tasks, positioning it competitively on cost among major AI models.

Figma has upgraded its Make AI design assistant to connect directly to production GitHub repositories, allowing designers to visually edit code and submit pull requests through standard engineering workflows. The two-way integration maintains existing governance, security checks, and code review processes while eliminating the previous one-way export limitation. This positions Figma Make as a design-first alternative to full-stack code generation platforms like Lovable and Claude Design.

MiniMax released a technical report on its M2 language model series while teasing an upcoming M3 model that uses a new sparse attention mechanism to achieve 15.6x faster response speeds on long-context tasks (up to 1 million tokens). The M2 report details the company's engineering approach to building competitive open-source models, while M3 aims to make ultra-long-context AI agent deployment economically viable through a custom sub-quadratic framework that balances speed with accuracy.

Alibaba released Qwen3.7-Max, a proprietary AI model capable of 35 hours of continuous autonomous execution, marking a shift toward closed, paid models rather than open-source releases. The model demonstrated superior performance on complex engineering tasks compared to open-source competitors, executing over 1,100 tool calls to optimize code without human intervention. However, its availability only through Chinese endpoints may limit adoption among Western enterprises concerned with data sovereignty and compliance.

Cohere released Command A+, a 218-billion-parameter sparse mixture-of-experts language model under an Apache 2.0 open-source license, marking the company's first fully open-weight model release. The model achieves near-lossless compression through 4-bit quantization while maintaining reasoning performance, enabling deployment on a single NVIDIA Blackwell B200 GPU or two H100s. The release reflects Cohere's strategic bet on sovereign AI, allowing enterprises and governments to run frontier-grade models within their own secure environments without relying on proprietary cloud services.

NanoCo AI, founded by former Wix engineer Gavriel Cohen and his brother Lazer Cohen, has raised a $12 million oversubscribed seed round to commercialize NanoClaw, an open source AI agent framework designed for enterprise deployment. The startup plans to offer each employee a personalized, secure AI assistant that builds persistent memory of their work through emails, documents, and call notes, functioning as a productivity multiplier rather than a replacement. NanoClaw maintains a minimalist 500-line TypeScript core that can be audited in eight minutes and runs agents in isolated Docker containers with policy-enforced API gateways to prevent unauthorized actions.

fal, a generative media platform serving 2.5 million developers, has selected AWS as its preferred cloud provider following a $300 million Series D funding round that valued the startup at $4.5 billion. The partnership aims to combine fal's optimized inference engine with AWS's global infrastructure to deliver 99.99% uptime for millions of daily API calls across image, video, audio, and 3D generation workloads. The deal signals a shift in the generative AI market from model development toward infrastructure and scaling for commercial consumption.

Corti, a Copenhagen-based healthcare AI company, launched Symphony for Speech-to-Text, a clinical-grade speech recognition model that achieves 1.4% word error rate on medical terminology, significantly outperforming OpenAI's Whisper (17.4%), ElevenLabs (18.1%), and other general-purpose models. The model also reaches 98.3% recall on formatted clinical entities like dosages and measurements, compared to 44.3% for the strongest general-purpose baseline. The launch underscores a broader shift in enterprise AI where specialized, domain-specific models can outperform foundation model providers in regulated industries, particularly as healthcare moves into an agentic era where accurate transcription becomes foundational data for downstream AI systems.

Raindrop AI launched Workshop, an open source MIT-licensed debugging and evaluation tool for AI agents that runs locally on developers' machines. The tool captures real-time traces of agent behavior, tool calls, and decisions in a lightweight SQL database file, displayed on a local dashboard, eliminating the need to send telemetry to external servers. A standout feature is the self-healing eval loop, which allows coding agents like Claude Code to autonomously read traces, write evaluations, identify logic errors, and fix broken code until assertions pass. The tool supports TypeScript, Python, Rust, and Go, and integrates with major frameworks including Vercel AI SDK, OpenAI, Anthropic, LangChain, LlamaIndex, and CrewAI.

New VB Pulse survey data shows Microsoft and OpenAI dominating enterprise agent orchestration platforms, but Anthropic's first measurable appearance in the tracker signals a strategic shift in how the AI market will compete. Rather than fighting over which model answers prompts best, the next battleground is the control plane where agents plan, access tools, run workflows, and prove compliance to security teams. Anthropic moved from 0% to 5.7% adoption in orchestration between January and February, a small number that gains significance because it coincides with Claude's dramatic rise in enterprise model adoption, from 23.9% to 56.2% in the same quarter.

Anthropic has reinstated support for third-party AI agents like OpenClaw on its paid Claude subscriptions, reversing an April 2026 ban that prohibited such usage. However, the reinstatement comes with a structural change: programmatic usage now draws from a separate, fixed monthly credit pool (worth $20 to $200 depending on plan tier) billed at API rates, rather than from the general subscription allowance. Credits do not roll over and cannot be supplemented by regular subscription usage, forcing users to purchase additional credits if they exceed the monthly Agent SDK allocation. This shift addresses Anthropic's infrastructure constraints by shifting the cost of inefficient third-party code back to users rather than subsidizing it across the broader subscriber base.

Perceptron Inc., a two-year-old startup led by former Meta and Microsoft researchers, released Mk1, a video analysis AI model priced at $0.15 per million input tokens and $1.50 per million output tokens, undercutting competitors like OpenAI's GPT-5, Anthropic's Claude Sonnet 4.5, and Google's Gemini 3.1 Pro by 80-90 percent. The model demonstrates strong performance on spatial and video reasoning benchmarks, including a score of 85.1 on EmbSpatialBench and 88.5 on VSI-Bench, while maintaining native video processing at up to 2 frames per second across a 32K token context window. Perceptron positions Mk1 as a practical tool for enterprise use cases including security monitoring, video content analysis, and behavioral assessment, moving video understanding from research-grade capability to mainstream accessibility.

Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati and researcher John Schulman, has unveiled a research preview of 'interaction models' designed to enable near-real-time, simultaneous voice and video conversation with AI. Rather than the current turn-based model where users input and wait for output, these systems use a full-duplex architecture that processes 200ms chunks of input and output concurrently, allowing the AI to listen, speak, and see in real time. The company demonstrated the approach with TML-Interaction-Small, a 276-billion parameter Mixture-of-Experts model, paired with a background reasoning system for complex tasks. A limited research preview will launch in coming months, with broader availability expected later in 2026.