Anthropic launches Claude Opus 4.7 with major coding and vision improvements, Cloudflare expands its agentic infrastructure stack, and Rust 1.95 ships language quality-of-life improvements

Simon Willison's Weblog

Qwen3.6-35B-A3B Beats Claude Opus 4.7 on the Pelican Test

Simon Willison ran his informal 'pelican riding a bicycle' SVG benchmark — a longtime gauge of model spatial reasoning — against Alibaba's Qwen 3.6-35B-A3B running locally on his laptop, and found it produced geometrically more accurate illustrations than Claude Opus 4.7. Willison is careful to note the test is deliberately absurd: a quantized local model outperforming a frontier API model on a drawing task does not imply general superiority, and the pelican benchmark has become less reliable as a proxy for overall capability. What the result illustrates is how rapidly open-weight model quality is improving — capable, locally-runnable models are closing in on frontier performance for specific tasks, with zero API cost or data exposure.

Rust Blog

Rust 1.95.0 Released with cfg_select! Macro and if-let Guards

Rust 1.95.0 ships two headline ergonomic improvements: the new cfg_select! macro, which simplifies platform-specific code branching by selecting among conditional expressions without nested cfg attributes, and if-let guards in match arms, which allow combining pattern matching with arbitrary boolean conditions in a single arm. The standard library gains a batch of stabilized APIs across atomic operations, mutable collection access, and MaybeUninit conversions, with several now available in const contexts. The release continues Rust's steady cadence of quality-of-life improvements aimed at reducing the friction of systems programming without compromising its safety guarantees.

Goose.love Blog

What Actually Uses Rust in Production

A comprehensive survey of real-world Rust adoption finds the language has crossed a meaningful threshold in production systems. Microsoft, Apple, AWS, Google, Meta, and Cloudflare all run Rust in significant parts of their infrastructure, and the language is now embedded in the Linux kernel, Windows 11, and Chromium. Government adoption has also accelerated: both the NSA and the White House have formally recommended Rust for memory-safe programming in response to systemic vulnerabilities in C and C++ codebases. Despite this momentum, the article cautions that mainstream adoption still requires sustained ecosystem maturity and education investment — Rust has passed the 'respected early adopter' phase but hasn't yet achieved the effortless reach of Go or Python.

GitHub Blog

How GitHub Uses eBPF to Block Circular Dependencies in Deployments

GitHub's deployment infrastructure faced a class of silent failures: during an outage, the deployment system would directly or transitively depend on the service it was trying to fix, creating circular loops that paralyzed recovery. Their solution uses eBPF programs attached at the kernel level to monitor and block problematic outbound connections from deployment processes, enforced through DNS redirection and domain blocklists. The approach intercepts network calls at the process group level without requiring application changes, making it transparent to existing tooling. By proactively detecting and preventing circular dependency connections, GitHub's deployment system can now self-heal during partial outages instead of becoming entangled in the infrastructure it's trying to restore.

LeadDev

JetBrains Central: Unified Platform for AI Agent Orchestration

JetBrains launched Central, a platform designed to close the gap between AI tool adoption and AI workflow integration. While 90% of developers now use AI tools, only 13% integrate them across the entire development lifecycle. Central provides governance, orchestration, and cloud infrastructure for autonomous agent workflows — managing agents like Junie for coding, Koog for cross-application tasks, and Grazie for writing assistance under one roof. The company faces credibility challenges around pricing transparency: its hybrid per-seat and pay-as-you-go model follows previous incidents where developer credits depleted unexpectedly. Whether Central simplifies JetBrains' expanding AI portfolio or adds another layer of complexity for teams evaluating the platform remains an open question.

Cloudflare Blog

Cloudflare AI Platform: 70+ Models Through a Single Inference API

Cloudflare evolved AI Gateway into a full inference platform, consolidating access to 70+ AI models across 12+ providers behind a single API endpoint. Developers can switch providers with a one-line change, use automatic failover for reliability, and route to custom fine-tuned models alongside hosted ones. The platform includes cost management tooling to control spend across provider boundaries. The design targets agent developers who want to avoid vendor lock-in: a unified abstraction over heterogeneous AI backends means agents can be retargeted to different models or providers without architectural rewrites.

Cloudflare Blog

Cloudflare Artifacts: Git-Backed Versioned Storage Built for AI Agents

Cloudflare launched Artifacts, a distributed versioned filesystem designed for agents that need to create, manage, and version repositories programmatically at scale. The service exposes standard Git operations through REST and Workers APIs, built on Durable Objects with a custom WebAssembly Git implementation backed by R2 and KV storage. Cloudflare also open-sourced ArtifactFS, which optimizes large repository cloning to 10–15 seconds by asynchronously fetching file trees rather than waiting for full pack transfers. The service is in private beta at $0.15 per 1,000 operations and $0.50/GB-month — positioning Cloudflare as infrastructure for the growing class of agents that need to read, write, and version code or data as part of their work.

Android Developers Blog

Android CLI Cuts LLM Token Usage 70% and Triples Build Speed for AI Agents

Google introduced the Android CLI, a command-line interface purpose-built for agent-driven Android development. The tool reduces LLM token usage by over 70% and delivers 3x faster task completion compared to agents working through standard tooling. It ships alongside two companion systems: Android Skills, a set of markdown-based best-practice instructions that automatically activate in supported agents, and the Android Knowledge Base, a queryable reference for current SDK guidelines and patterns. Together the three components form a structured interface that lets AI agents — in any compatible framework — operate more efficiently against the Android SDK without consuming excessive context or repeating common mistakes.