The week's main developments, grouped by original publication date. Every item links to our full story and the original source.
Google announced Gemini 4 Argon for complex tasks. The model is rolling out to trusted cyber defenders through the Fairwind Program.
The 1 million output token limit, up from 64,000, is intended to support deep reasoning across complex, long-horizon workflows.
Cohere released Embed 5, a new retrieval model series featuring a 128K-token context window and shared embedding space for enterprise use.
Shared embedding spaces allow organizations to balance retrieval performance and cost by using Pro for indexing and Fast for queries.
Quine is a Microsoft Research effort linking a biology world model with scientific tools; initial access is limited to its Fellows program and select research collaborations.
Quine is designed to help researchers use computational predictions to prioritize hypotheses and experiments, not to replace laboratory testing.
OpenAI has introduced GPT-6.1 Sol, a cost-efficient model providing performance comparable to GPT-6 Astra for coding and professional tasks.
The release provides a more economically viable option for developers building complex AI agents and software engineering tools that require high reasoning capabilities.
Anthropic reports that GLM-5.3 can generate end-to-end exploits, highlighting significant limitations in the model's safety protocols.
The release of a powerful, open-weight model capable of autonomous exploitation shifts the threat landscape, forcing both defenders and software developers to account for automated, high-speed vulnerability research.
Anthropic positions Claude Sonnet 5.5 for everyday work. Here is what changed and why it may matter for AI teams.
For teams running a high volume of routine tasks, speed and cost can change the economics of automation. Those figures come from the model provider, so teams should compare results on their own workloads before switching.
GitHub has launched computer use capabilities for Copilot in public preview, enabling interaction with desktop applications on macOS and Windows.
Beyond coding, Copilot can automate workflows in desktop applications that lack an API, command-line interface or MCP integration.
NVIDIA has released Metropolis Blueprint VSS 3.3, featuring a new vss-build-vision-ai skill for creating video analytics agents from text prompts.
This update streamlines the development of vision-based AI workflows by allowing engineers to generate functional agents through simple prompting rather than manual configuration.
OpenAI is rolling out Dots, persistent AI agents that manage third-party applications and perform proactive research for Pro and Enterprise users.
The launch of Dots shifts AI utility from interactive chat to persistent automation. By offloading multi-step tasks to agents that operate 24/7, users can minimize manual coordination between different software platforms.
Salesforce reports customer examples of AI agents in service, including Canada Goose’s reported resolution rates and AT&T’s use of customer context to brief retail experts.
The examples show how service automation can involve staff preparation and controls on model behavior. The results are company-reported; the material provides no methodology or baselines and does not independently verify them.
Barclays is expanding Claude use across the bank. Anthropic describes an employee assistant and email processing; Barclays expects most of its software engineers to use Claude Code in 2027.
The announcement describes Claude use in software development and internal banking operations. Adoption and email-processing figures are reported by Anthropic; the Claude Code figures are Barclays’ expectations, not achieved results.
A machine-learning system estimates space-weather risks at 66,935 substations in the continental US, producing location-specific forecasts 30–60 minutes before potential impacts.
After further validation with utilities and operational data, the estimates could help operators prioritize engineering reviews and consider targeted protective actions.
Whitbread is expanding Microsoft Copilot to 1,650 employees following a pilot project that reported daily time savings for support staff.
The deployment aims to improve operational efficiency, allowing staff to delegate routine tasks to AI and focus on high-value decision-making.
Salesforce introduced the Koa CRM model built on NVIDIA Nemotron 3 architecture and a new Enterprise AI Harness for secure AI governance.
The platform enables organizations to select and govern specific AI models for business tasks, ensuring that AI agents remain connected to enterprise data and security policies.
Meta has introduced new skills and connectors to its Muse AI agent to assist small business owners with operational automation.
By connecting scattered administrative tasks and communications into a single AI-driven interface, Meta aims to help small business owners save time on manual operations.
Accounting AI firm Basis published an internal test comparing the efficiency of GPT-6 Astra against GPT-5.6 Sol in completing 50-tab tax workbooks.
The case shows why reasoning effort can be adjusted during a long accounting task and why output quality needs a separate check. These performance figures come from Basis's own tests.
Users on Copilot Pro, Pro+, Max, Business, and Enterprise can request code reviews through supported REST and GraphQL APIs and optionally set the effort level for each request.
Teams on the listed plans can request reviews through the supported APIs from their own scripts, workflows, and internal tools. They can set the effort level per request, while the Balanced default leaves an explicitly selected Lite setting intact.
The change covers every Copilot experience. GitHub lists three alternatives; Copilot Enterprise administrators may need to enable access through model policies.
The change affects model choices across Copilot experiences. Users should review workflows and integrations, while Enterprise administrators may need to check access to alternatives in their settings.
GitHub Copilot adds support for multi-step automated workflows across its CLI, app, and SDK platforms.
This update allows developers to define and execute reusable, code-based orchestration processes directly within the Copilot environment.
Microsoft Foundry offers streaming transcription across 60 languages and two speech-generation models. Microsoft has published prices and named platforms it plans to add.
Developers can pair streaming transcription, with first partials arriving within the low hundreds of milliseconds, with either of two text-to-speech models. Published per-hour and per-1M-character prices support cost planning. Microsoft’s performance claims have not been independently verified here.
Google Cloud has released its Data Agent Kit, providing Model Context Protocol tools for integrating AI coding agents with cloud data services.
This toolkit enables developers to connect their existing coding agents directly to analytics and operational databases for more efficient data workflows.
CoreWeave provides early-access to NVIDIA Vera Rubin NVL72 systems with Spectrum-X networking in its cloud, alongside new tools for agentic AI workflows.
These updates are significant for AI engineering teams as they provide quantified performance improvements in inference throughput and deployment speeds for agentic workflows.
Hugging Face’s Open TTS Leaderboard compares open TTS models using WER/CER, RTFx, TTFA and speaker similarity. Its scores complement rather than replace human preference rankings.
For developers, the leaderboard offers a faster and more reproducible way to compare models than collecting votes for weeks. Its metrics are still proxies: WER and speaker similarity do not measure naturalness, expressiveness or listener preference, and the leaderboard does not replace human rankings.
Developers can now build AI agents that navigate websites and interact with browser interfaces within an OpenAI-hosted environment.
The update provides a framework for automating web-based tasks like testing or data collection, while placing the responsibility for security limitations and action-scope control on the developers implementing the tool.
OpenAI describes Ultrafast as its fastest API service tier. It is available to all API users for GPT-6 Astra at low rate limits, while GPT-5.6 Sol has preview access.
Developers can select OpenAI’s fastest API service tier for GPT-6 Astra while weighing its higher cost, low rate limits, and regional processing restrictions.
The GPT-6.1 Sol API documentation specifies context and output limits, reasoning settings, token prices, and restrictions for EU data residency.
These specifications help developers budget API use and assess deployment constraints. The model announcement is covered separately.
The GPT-6.1 Sol model has been added to Amazon Bedrock, providing developers with enhanced reasoning capabilities and lower task costs.
This release allows developers to execute complex multi-step business tasks with greater efficiency and reduced operational costs compared to previous versions.
Amazon Bedrock now supports Grok 4.7, featuring a 500K token context window and adjustable reasoning effort for developers and enterprise users.
This update provides professional users with granular control over AI reasoning intensity and extended context handling, which are essential for building robust agent-based applications.
Amazon has integrated Claude Sonnet 5.5 into its Bedrock service, offering lower costs and faster processing for coding and documentation tasks.
This update provides enterprises with a more cost-effective way to automate complex technical workflows. By lowering operational overhead for coding and analysis, developers can integrate high-performance AI capabilities into their AWS workloads more sustainably.
Anthropic's Claude Sonnet 5.5 is now available in GitHub Copilot for eligible professional and enterprise users via usage-based billing.
This integration allows developers to leverage Claude Sonnet 5.5 directly within their coding environments, offering greater flexibility in model selection.
NVIDIA announced an open software platform and reference system design to govern AI agents from testing through deployment across software and infrastructure.
The platform adds governance and control beyond the model and agent harness, including at the infrastructure level. This gives organizations tools to define and enforce boundaries as autonomous agents work.