Anthropic has released Claude Opus 4.6, introducing dynamic reasoning control and context compression for long-running agents, along with a 1 million token context window (beta). The model achieves 76% accuracy on MRCR v2 multi-needle retrieval. It is now available via AWS Bedrock, Google Vertex AI, Microsoft Foundry, and the Claude API, supporting up to 128,000 output tokens and new Agent Teams for parallel collaboration. Pricing remains at $5 per million input tokens and $25 per million output tokens, with requests exceeding 200,000 tokens entering higher billing tiers.

The Council of the European Union has reached a common position on the 'Omnibus VII' proposal to streamline AI regulation: timelines for high-risk AI systems have been adjusted, with new deadlines set for standalone systems on December 2, 2027, and for systems embedded in regulated products on August 2, 2028. The deadline for establishing national AI regulatory sandboxes is also postponed to December 2, 2027. The updated text introduces bans on generating non-consensual intimate content and child sexual abuse material, strengthens oversight powers of the AI Office over general-purpose AI models, and reinstates certain registration obligations. Negotiations with the European Parliament will follow.

AWS has partnered with AI chipmaker Cerebras to deploy its chips in AWS data centers and offer inference acceleration services, expected to launch in the second half of 2026. The solution uses a split workflow: Amazon's custom Trainium3 handles the prefill phase of requests, while Cerebras manages the decoding generation phase, aiming to improve throughput and cost efficiency compared to solutions like Nvidia. Financial terms were not disclosed. The report notes Cerebras' valuation at $23.1 billion and a prior $10 billion chip supply agreement with OpenAI, positioning it as a growing alternative in cloud inference supply chains.

OpenAI has launched a new version of the Responses API, offering a callable 'full computer environment' for models like GPT-5.2. Developers can enable models to execute shell commands, run scripts, and reuse 'skill' packages within isolated containers, enabling more reproducible and auditable production-grade agent workflows. Concurrent infrastructure updates improve error handling and structured output processing to reduce token waste from invalid retries and enhance stability. Details on pricing, quotas, and regional availability have not been disclosed.

Verifiable AI startup Axiom Quant has secured $200 million in Series A funding at a $1.6 billion valuation, led by Menlo Ventures. The company leverages formal verification to mitigate hallucination and security risks in LLM-generated code: using the theorem-proving language Lean as a core, it enables models to generate code and reasoning steps that can be checked by deterministic verifiers, providing machine-verifiable correctness guarantees at the logical level and reducing uncontrolled risks for enterprise AI adoption. Funds will support scaling infrastructure and teams to advance training and productization of its 'verification data flywheel'.

Enterprise AI agent platform Wonderful AI has announced a $150 million Series B round at a $2 billion valuation, led by Insight Partners with participation from Index Ventures. The platform enables low-code development and deployment of agents, supporting external service integration, task constraint configuration, simulation testing, and dashboards for monitoring response latency and business accuracy. Case studies claim up to 60% reduction in task completion time. Wonderful plans to grow its team from 350 to around 900 by end of 2026 and strengthen on-site engineering delivery to shorten implementation cycles from months to weeks or less.

Google AI introduced Groundsource, a method using Gemini to extract structured disaster records from multilingual, unstructured news, releasing an open dataset of 2.6 million historical urban flooding events across over 150 countries. By combining semantic parsing with geospatial mapping, the approach addresses gaps in traditional remote sensing and existing databases caused by cloud cover, satellite revisit cycles, and data scarcity. The dataset has already trained new predictive models and supports up to 24-hour early warnings for urban flash floods via Google Flood Hub, serving as a foundation for rapid-disaster early warning systems.

Security firm Permiso has disclosed a cross-prompt injection (XPIA) vulnerability in Microsoft Copilot's email and Teams message summarization features: attackers can embed hidden instructions in HTML/CSS of emails to manipulate Copilot into generating summaries containing fake alerts and phishing links, exploiting user trust in AI assistants to prompt clicks and information disclosure, potentially leading to data leaks or account compromise. The report indicates Copilot in Teams is more vulnerable, with inconsistent protection across interfaces. Mitigations include employee training, strict DLP policies, email filtering, and secure link protection.

Random Labs, supported by Y Combinator, has launched Slate V1, described as a 'swarm-native' coding agent. Using dynamic pruning and a Thread Weaving architecture, it separates strategic decision-making from execution tasks, employs recursive language models and 'episodic memory' mechanisms to reduce context loss in long-cycle projects, and supports multi-model collaboration (e.g., coordination, coding, retrieval) with routing between cost and quality. The company reports a 2/3 success rate on the make-mips-interpreter task in Terminal Bench 2.0 internal tests, and offers usage-based billing, organizational monitoring, and billing tools for professional R&D teams.

An evaluation of 525 real-world attacks focused on the 'lethal triad' scenario—privileged data access, untrusted content injection, and outbound exfiltration paths—revealed high susceptibility in agent systems. Testing showed attack success rates of 90.3% for GPT-4o-mini, 82.4% for Gemini 2.5 Flash, and 6.7% for Claude Sonnet; control groups showed no leakage and results were statistically significant. The research team open-sourced Cerberus, a runtime protection tool achieving 28.5% overall detection, with better performance in data source monitoring and provenance tracking layers. Performance overhead is minimal, with p99 latency at 0.23ms, making it suitable for live monitoring.

Multiple sources indicate Meta's foundational AI model codenamed 'Avocado' has been delayed from its planned March 2026 release to at least May, due to underperformance in internal evaluations on critical capabilities such as reasoning and coding relative to key competitors. Reports also highlight internal disagreements over open-sourcing and roadmap direction, and that planning for the next-generation model 'Watermelon' has already begun, reflecting pressure on Meta's foundational model iteration pace and organizational alignment. The delay may affect its competitive window against Google, OpenAI, and Anthropic in enterprise and developer ecosystems, though Meta has not disclosed specific evaluation metrics or revised timelines.