Anthropic Releases Claude Opus 4.7 Outperforming 4.6 in All MetricsNew Model Trained on Radiologist Gaze Improves X-ray DiagnosisZo Computer Achieves 20x AI Reliability Improvement on VercelGoogle Launches 7 AI-Powered Travel Tools for Summer PlanningGitHub Builds Emoji List Generator Using Copilot CLITool Scans Website AI Agent CompatibilityVercel Preserves Active Branch Deployments Under Retention PoliciesClaude Code: Native Binary Support and Domain ControlsGitHub Status Page Now Provides Detailed Health MetricsPyCon US 2016 to Host AI and Security Tracks in Long BeachAI Companies Buying Slack Data from Failed StartupsCredo: Controlling LLM Pipelines via Beliefs and PoliciesCounterfactual Routing Activates Dormant Experts to Fix MoE HallucinationsCompressed-Sensing-Guided Structured Pruning for LLMsGeometric Routing Enables Causal Control in MoE ExpertsMixAtlas: Uncertainty-aware Data Mixture Optimization for Multimodal LLMsOpenAI Codex Releases Version 0.122.0-alpha.9Datasette 1.0a28 Fixes Compatibility Bug from 1.0a27Anthropic Releases Claude Opus 4.7 Outperforming 4.6 in All Metrics

Anthropic has released its flagship model Claude Opus 4.7, demonstrating improvements over the 4.6 version across all dimensions. The new SOTA model shows enhanced performance in code generation, reasoning, and task handling. This release strengthens Anthropic’s competitive position in the enterprise AI market, with the model now available through official channels.

An arXiv paper introduces a new vision-language model trained on radiologists’ gaze patterns and reasoning processes, significantly improving chest X-ray interpretation accuracy. While existing models can automate readings, a gap remains between their outputs and radiologist expertise. By incorporating physician decision pathways, the new model bridges this clinical utility gap, enhancing AI diagnosis reliability and interpretability.

Zo Computer achieved a 20x improvement in AI service reliability on Vercel, reducing retry rate from 7.5% to 0.34%, achieving 99.93% chat success rate (up 1%), and cutting P99 latency by 38% (131s to 81s). New model deployment now takes under 1 minute. These improvements were achieved through optimized model management and load balancing, providing a scalable architecture for high-concurrency AI services.

Google has released 7 new AI-powered tools to help users plan trips, find deals, and explore destinations. These tools integrate AI search, price comparison, and personalized recommendation features to provide customized suggestions based on user history and preferences. They simplify decision-making for everything from flight deals to hotel bookings and attraction recommendations.

GitHub demonstrated how to build an emoji list generator using the GitHub Copilot CLI during a Rubber Duck Thursday stream. The showcase highlighted the tool’s code completion and generation capabilities, allowing developers to create utilities through natural language descriptions. This practical demonstration proves the value of AI-assisted programming in improving development efficiency.

Developers can now check their website’s compatibility with AI agents using isitagentready.com. The tool evaluates whether websites can be properly crawled and understood by AI agents, helping developers optimize site structure for better AI interaction. As AI agents become more prevalent, this tool helps identify and fix compatibility issues early, ensuring sites work effectively with next-gen AI services.

Vercel updated its retention policies to preserve the latest preview deployments for branches with open or unmerged pull requests. Previously, deployments could be removed if they exceeded the configured retention window, even for active branches. This change allows developers to safely use shorter retention periods without risking the loss of in-use preview environments, offering more flexibility in managing deployment resources.

Claude Code released v2.1.113 with key updates: the CLI now spawns a native Claude Code binary via platform-specific optional dependencies, added sandbox.network.deniedDomains setting to block specific domains even when broader allowed wildcards exist, and improved fullscreen mode. These changes enhance performance, security controls, and flexibility for developers configuring their coding environment.

GitHub updated its status page with more specific health metrics. Users can now view detailed platform health data including API response times and database connection status. The update enhances transparency, helping developers identify service issues faster.

PyCon US 2026 will be held in Long Beach, California from May 13-19, marking the first return to the West Coast since Portland in 2017. The conference will feature new AI and security tracks, with core talks scheduled for May 15-17 and tutorial/sprint days surrounding the main event. As the Python community’s largest annual gathering, it expects global developer attendance to explore latest technology trends.

Multiple AI firms are purchasing Slack chat data from defunct startups, containing internal discussions and roadmaps. This raises data privacy concerns as it may violate user agreements and regulations. Developers should be aware of legal risks when using such training data.

The arXiv paper ‘Credo’ proposes belief-and-policy-based control for LLM pipelines. It models intermediate reasoning beliefs in long-decision AI systems, enabling pipeline-level policy control. Experiments show improved accuracy in complex decision-making tasks for evolving AI applications.

An arXiv paper proposes counterfactual routing to fix MoE hallucination on long-tail knowledge. The method activates dormant experts by simulating counterfactual queries, reducing hallucinations by 37%. It outperforms traditional routing on WikiText and MMLU benchmarks.

An arXiv paper introduces compressed-sensing-guided structured pruning for LLMs. It achieves 40% parameter reduction and 30% lower decoding latency while maintaining accuracy. By preserving critical neuron connections, it balances speed and performance for resource-constrained deployments.

The arXiv paper ‘MixAtlas’ proposes uncertainty-aware data mixture optimization for multimodal LLM training. It dynamically adjusts multimodal data ratios based on sample uncertainty, improving training efficiency. It achieves 15% higher downstream accuracy on Image-Text tasks compared to traditional methods.

OpenAI Codex has released version 0.122.0-alpha.9, continuing the series of incremental updates. While specific improvements aren’t detailed, this latest test version likely contains performance optimizations or new feature support for OpenAI’s code generation model. Developers can access the test version through official channels to provide feedback.

Datasette released version 1.0a28, fixing a compatibility bug in executewritefn() callbacks introduced in 1.0a27. This update addresses accidental breakages discovered during Datasette Cloud’s upgrade to 1.0a27. Developers should update to avoid callback-related errors.