1. Claude Opus 4.7: Anthropic's new frontier model for agentic coding and complex tasks
Anthropic has released Claude Opus 4.7, its most capable generally available model to date, featuring a 1M token context window. The model demonstrates improved performance in advanced software engineering, long-running agentic tasks, and high-resolution vision compared to Opus 4.6. It is available immediately via the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing remains unchanged at $5 per million input tokens and $25 per million output tokens. Notably, the model includes an adaptive thinking requirement without a manual override, which may affect routing for simpler tasks.
2. OpenAI Codex update: Background computer use and native web browsing for Mac
OpenAI has released a major update to its Codex desktop app, introducing background computer use capabilities for macOS users. The agent can now operate desktop applications in parallel with the user by seeing, clicking, and typing with its own cursor. The update also adds an in-app browser for frontend iteration, image generation via gpt-image-1.5, and over 90 new plugins for tools like GitLab and Jira. These features aim to automate longer-running development tasks and testing workflows directly within the local environment.
3. Cloudflare Artifacts (Private Beta): A Git-compatible versioned file system for AI agents
Cloudflare has introduced Artifacts, a distributed, versioned file system designed specifically to handle the high volume of code generated by AI agents. The service exposes a REST API and a native Workers API, allowing developers to programmatically create repositories, generate credentials, and commit code from serverless functions. It functions as a standard Git remote, enabling seamless cloning, forking, and pushing by both agents and standard Git clients. Artifacts is currently in private beta for paid Workers users, with a public beta planned for early May 2026.
4. Qwen3.6-35B-A3B: A new 35B MoE vision-language model optimized for local agentic coding
Alibaba has released Qwen3.6-35B-A3B, a 35B parameter Mixture-of-Experts vision-language model with 3B active parameters. The model features advanced agentic coding capabilities and thinking preservation, making it highly capable for complex development tasks. It is available on Hugging Face and can be run locally on consumer hardware, such as a MacBook Pro, using quantized formats. Early community testing indicates strong performance in visual generation and coding tasks compared to larger proprietary models.
5. Mozilla Thunderbolt: A sovereign AI client for self-hosted infrastructure
Mozilla has introduced Thunderbolt, a front-end client designed for developers and businesses running self-hosted AI infrastructure. Built on the open-source Haystack framework, the client connects to any ACP-compatible agent or OpenAI-compatible API. It features integration with locally stored enterprise data using an offline SQLite database as a local source of truth. The tool also includes optional end-to-end encryption and device-level access controls to ensure data privacy when interacting with local models.
6. Android CLI: A new terminal interface optimized for AI agents
Google has launched a new Android CLI designed specifically to facilitate agentic workflows outside of Android Studio. The tool provides a lightweight, programmatic interface for agents like Claude Code and Codex to interact with the Android SDK and development environment. It includes commands for environment setup, project creation, and device management, which Google claims reduces LLM token usage by over 70% for project setup tasks. The release is accompanied by an Android Knowledge Base to provide agents with up-to-date context on modern Android development practices.
7. Cloudflare Email Service: Public beta brings native email sending to Workers and AI agents
Cloudflare has moved its Email Sending service into public beta, allowing developers to send transactional emails directly from Workers using a native binding without managing API keys. The release includes a new Email MCP server, Wrangler CLI email commands, and an open-source agentic inbox reference application. These tools are designed to make email a first-class, bidirectional interface for AI agents, enabling workflows like customer support and invoice processing. The service also supports REST API access for integration with any platform or language.
8. Cloudflare AI Platform: Unified inference API for third-party models
Cloudflare has updated its AI Platform to serve as a unified inference layer, allowing developers to call third-party models from providers like OpenAI and Anthropic using the existing AI.run() Workers binding. This enables seamless switching between Cloudflare-hosted models and external APIs with a single line of code. The platform also integrates AI Gateway to handle automatic retries, connection drops, and response buffering, which is particularly useful for long-running agentic tasks. REST API support for non-Workers environments is planned for the coming weeks.
9. Gemini API billing risk: Unrestricted Firebase keys lead to large unexpected charges
A developer reported incurring over €54,000 in unexpected Gemini API charges within 13 hours after enabling Firebase AI Logic with an unrestricted browser key. The automated, non-user-driven traffic bypassed standard budget alerts due to delayed cost reporting. In response, Google highlighted newly rolled-out billing account caps, project-level spend limits, and a transition to prepaid billing for new accounts to prevent similar overages. Developers using Firebase AI Logic or client-side Gemini API calls should immediately implement strict quotas, App Check, and spend caps.
10. Humwork A2P: A human-in-the-loop fallback marketplace for AI agents
Humwork has launched an Agent-to-Person (A2P) marketplace that connects AI agents with verified human experts when they encounter errors or get stuck. The platform integrates with AI coding tools like Claude Code, Cursor, and Replit via the Model Context Protocol (MCP). When an agent triggers a handoff, the system securely shares the full session context—including code and error logs—with a human expert who resolves the issue and pushes the solution back to the agent. This provides developers with a structured human-in-the-loop fallback mechanism for autonomous workflows.
11. Salesforce Headless 360: Entire CRM platform exposed as APIs and MCP tools for AI agents
Salesforce has announced Headless 360, a major architectural shift that exposes its entire platform's capabilities as APIs, CLI commands, and Model Context Protocol (MCP) tools. The initiative ships with over 100 new tools and skills immediately available to developers. This allows AI agents to interact with, reason over, and execute tasks within the Salesforce ecosystem without requiring a graphical user interface. The move signals a significant enterprise shift toward agent-first infrastructure and programmatic CRM management.
12. Kampala: A MITM proxy for agentic reverse-engineering of legacy workflows
Kampala has launched a man-in-the-middle (MITM) proxy designed to help AI agents reverse-engineer existing web, mobile, and desktop workflows into APIs. Unlike brittle browser automation, the tool leverages existing session tokens and anti-bot cookies to interact deterministically at the network request layer. Developers can use Kampala's MCP integration to manually perform a workflow once, allowing an AI coding agent to automatically generate a script or API to replicate it. This provides a more stable alternative to computer-use agents for automating legacy dashboards and on-premise solutions.