#ai

An open bug filed against Claude Code on 2026-07-04 by an [Enterprise ZDR](https://docs.anthropic.com/en/docs/build-with-claude/zero-data-retention) user describes a working session on Sonnet 5 that suddenly starts referencing an unrelated Minecraft temple build, then doubles down on the wrong task in its recap. The reporter (GitHub: [@milesrichardson-edb](https://github.com/milesrichardson-edb), issue [anthropics/claude-code#74066](https://github.com/anthropics/claude-code/issues/74066)) is on Enterprise Zero Data Retention, the tier Anthropic specifically advertises as session-isolated. Triage on the reporter's local session JSONL at `~/.claude/projects/<encoded-cwd>/<session-id>.jsonl` finds the leaked text is not in the transcript, ruling out a local context bleed by file overlap. Four other users in the comments (with work histories going back to last year) describe near-identical behavior across Claude Code, Claude Mobile, and Claude deep research. The most plausible architectural fit is shared KV-cache state in inference ([per @yv3nne in the comments](https://github.com/anthropics/claude-code/issues/74066#issuecomment-4880448776)), but no Anthropic engineer has commented on the issue in the 22 hours since it was filed, and the issue reached the top of [Hacker News](https://news.ycombinator.com/item?id=42481789) on 2026-07-04. The tone in the thread is split: half suspecting a real platform cache reuse, half suspecting a [sonnet-5-specific hallucination triggered by a Pygments lexer](https://github.com/anthropics/claude-code/issues/74066#issuecomment-4880334711). Both readings are credible.

Anthropic released Claude Code [v2.1.199](https://github.com/anthropics/claude-code/releases/tag/v2.1.199) on 2026-07-02, one day after the v2.1.198 feature release (Chrome GA, background-agent auto-PR, /dataviz, Gateway on AWS) and two days after v2.1.197 made Sonnet 5 the default model. v2.1.199 is a bug-fix release with 24 entries: one new feature (stacked slash-skill invocations load up to 5 leading skills in one prompt), and a long list of reliability fixes for subagent error reporting, the background-agent daemon on Linux, streaming-response preservation on mid-stream API errors, SSL/TLS error handling behind corporate proxies, plan mode tool-call handling, and a per-env retry policy that no longer caps the user-tunable retry count at 15. Most of the items are in the background-agent / subagent area, which is consistent with the v2.1.197 + v2.1.198 direction of treating delegation as the product surface for Sonnet 5.

Anthropic released Claude Code [v2.1.198](https://github.com/anthropics/claude-code/releases/tag/v2.1.198) on 2026-07-01T20:45:36Z, the second consecutive day of Claude Code shipping after v2.1.197 (the Sonnet 5 default release) the day before. The headline change is that Claude in Chrome is now generally available inside Claude Code: the browser-side assistant that was a research preview moves to a stable channel and is no longer gated behind a feature flag or a waitlist. The release also lands the first end-to-end automation for background agents in `claude agents`, which now commit, push, and open a draft pull request when they finish code work in a worktree, instead of stopping to ask for permission. Other notable additions: a `/dataviz` skill for chart and dashboard design with a runnable color-palette validator, a Gateway upstream for Claude Platform on AWS (`anthropicAws`) with model-not-found responses advancing the failover chain, a built-in Explore agent that now inherits the main session's model (capped at opus) instead of running on Haiku, and subagents that inherit the parent session's extended thinking configuration. The release also removes the `/agents` wizard and replaces it with the conversational flow, fixes a long list of reliability bugs around background agents, network drops, and platform auth, and improves syntax highlighting by upgrading to highlight.js 11.

Anthropic announced Claude Sonnet 5 on 2026-06-30 ([Introducing Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5)), and Claude Code [v2.1.197](https://github.com/anthropics/claude-code/releases/tag/v2.1.197) (released the same day) made it the default model. The launch is the first Sonnet-line model that closes the agentic gap with Opus 4.8, ships a native 1M-token context window, and lands on introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026 (then $3/$15 per Mtok). Sonnet 5 is the new default for Free and Pro plans in Claude and is the default model in Claude Code, available immediately on update. The release ships with cyber safeguards enabled by default (the same safeguards as Opus 4.7 and 4.8, less strict than the Fable 5 set), a refreshed tokenizer that inflates token counts 1.0 to 1.35x, and rate-limit increases across Chat, Cowork, Claude Code, and the Claude Platform to absorb the higher token usage of the new effort-level controls. The Anthropic system card and the Claude Code release notes both ship the same day, so the Claude Code audience gets the new model the same day it is announced.

Cline shipped v4.0.1 on June 28, 2026 and v4.0.2 on June 29, 2026 (github.com/cline/cline), a two-step recovery cycle for the v4.0.0 SDK migration that landed on June 26. v4.0.1 ships the pre-SDK 3.89.x VS Code extension under a 4.0.1 version number, built from a dedicated `legacy-extension` branch via a new `ext-vscode-publish-legacy.yml` workflow, to resolve regressions reported in 4.0.0 (broken diff previews in the editor, run_commands errors during file edits, broken file-editing flow with GLM 5.2 and MiniMax M3 through Ollama). v4.0.2 restores the SDK-backed code path on top of the same legacy branch, adding reasoning effort support (including `xhigh`) for DeepSeek thinking models (#11938), a centralized reasoning effort control layer for ClinePass (#11954), canonical Z.ai model ids (#11951), webview env replacement fix (#11955), ClinePass and Z.ai metadata polish (#11958), and a default focus chain settings fix (#11960). CLI v3.0.32 ships the same day with SDK v0.0.54 context-compaction improvements and ClinePass onboarding polish. The release sequence shows a project recovering a major migration in 72 hours by tagging the legacy branch forward rather than reverting the SDK work.

Cline shipped v4.0.0 on June 26, 2026 (github.com/cline/cline), a major version that migrates the VS Code extension from its legacy standalone task implementation onto the shared Cline SDK, the same TypeScript engine that runs the Cline CLI, Kanban, and JetBrains plugin. The release adds ClinePass (built-in onboarding, provider selection, subscription handoff, and entitlement handling), a Customize marketplace for Skills, MCP servers, and the new Cline Plugins, queued prompts, edit-and-regenerate, a provider and model configuration rework around providers.json and a shared model catalog (Fireworks GLM 5.2, Kimi K2.6 Fast, Kimi K2.7 Code, Qwen 3.7 Plus, MiniMax M3, SAP AI Core, LiteLLM, Codex OAuth), command auto-approval disabled by default, and an extension build and package workflow moved to Bun. Subagents are temporarily disabled in the extension while the SDK-backed experience stabilizes.

The OpenAI Codex 0.142 line shipped between June 22 and June 26, 2026 (rust-v0.142.0 through rust-v0.142.3 on github.com/openai/codex). The cycle turns the agent into something a team can govern: configurable rollout token budgets that track usage across threads and abort turns when exhausted, multi-agent delegation configurable as disabled, explicit-request-only, or proactive, an indexed web-search mode that permits live search while restricting page access to server-approved URLs, MCP tools that use tool search by default, a /plugins marketplace reorganized into OpenAI Curated, Workspace, and Shared with me sections, system proxy (PAC, WPAD, bypass) support on Windows and macOS for authentication, and a long list of remote-executor reliability and security fixes. It builds directly on the 0.141 Noise-encrypted relay and cross-OS PathUri layer from June 18.

Anthropic published a blog post on 2026-06-24 (https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks) accusing Alibaba of an industrial-scale distillation campaign against Claude. Anthropic says the campaign ran from April 22 to June 5, 2026 and produced over 28.8 million exchanges with Claude through almost 25,000 fraudulent accounts. The accusation is the second in Anthropic's public distillation disclosures: the first, on 2026-02-23, named DeepSeek, Moonshot AI, and MiniMax as attackers behind 16 million exchanges across 24,000 fraudulent accounts. Reuters (Krystal Hu, Eduardo Baptista) and Bloomberg (Saritha Rai) reported the new allegation on 2026-06-24; WSJ ran Anthropic Claims Alibaba Ran 'Brazen' Campaign to Access Its Claude AI Model the same day. Alibaba has not yet issued a public statement at time of writing.

Astro 7.0.0, published on 2026-06-22, lands the 7.0 stable branch after ten weeks of beta. The headline changes: a Vite 8 upgrade, the Rust-based Astro compiler as the default (the Go compiler is removed), the Sätteri Markdown pipeline as the default (remark/rehype moves out of the default install), advanced routing promoted from experimental (with src/fetch.ts as the new default entrypoint), route caching promoted from experimental (top-level cache and routeRules, with cacheNetlify() and cacheVercel() providers landing in the same release), a background dev server mode designed for AI coding agents (astro dev --background, .astro/dev.json lockfile, astro dev stop|status|logs), compressHTML: 'jsx' as the new default, the @astrojs/db package removed, and every official integration bumped a major version (vue 7, react 6, svelte 9, preact 6, solid-js 7, vercel 11, netlify 8, node 11). create-astro@5.1.0, shipped in the same release wave, now writes an AGENTS.md file into every new project with a CLAUDE.md symlink pointing at it.

Sakana AI launched Fugu on June 22, 2026, a multi-agent orchestration system delivered as a single OpenAI-compatible API. Two models, Fugu and Fugu Ultra, coordinate a pool of closed-source LLMs using learned orchestration from two ICLR 2026 papers (TRINITY and Conductor). Fugu Ultra posts benchmark scores competitive with Anthropic's Fable 5 and Mythos Preview, but the release draws sharp criticism over cost transparency, closed-source dependencies, and benchmark methodology.

Codex 0.141.0 (June 18, 2026) makes Noise IK the default transport between orchestrator and exec-server, ships a PathUri / NativePathString layer that round-trips POSIX, Windows-drive, and UNC paths without leaking the URI encoding, opens a `created-by-me-remote` plugin marketplace, raises the MCP tool timeout to 300 seconds, and pins the bundled SQLite to 3.51.3 to keep the WAL-reset corruption fix in place after dependency refreshes.

On June 12, 2026, Google Cloud published the Open Knowledge Format (OKF), an open specification that formalizes the LLM-wiki pattern into a portable, interoperable format: a directory of markdown files with YAML frontmatter, one required field (type), five recommended ones, and zero required tooling. The tweet from Google Cloud Tech on June 16 drove 117,000 views in 24 hours and made the spec the most-discussed knowledge-format launch of the year. This long read walks through the v0.1 spec section by section, the design choices that make it deliberately minimal, what Google is shipping alongside it (an enrichment agent for BigQuery, a static HTML visualizer, three sample bundles, and a native BigQuery Knowledge Catalog integration), and the open question every AI agent builder and data platform team should be tracking over the next six months.

On June 16, 2026, Z.ai released GLM-5.2 under MIT, with a solid 1M-token context, an IndexShare sparse-attention trick that cuts per-token FLOPs by 2.9×, an anti-hack module for coding RL, and benchmark numbers that put it within a few points of Claude Opus 4.8 on Terminal-Bench 2.1, ahead of GPT-5.5 on three long-horizon suites, and at the top of the open-weights leaderboard on every coding benchmark the company chose to publish. This long read walks through the architecture, the IndexShare + MTP story, the slime agentic RL infrastructure, the anti-hack module, the full benchmark table with all the footnotes, and what this means for the rest of the open-weights long-context race.

On June 16, 2026, four trading days after SpaceX's record $85.7bn IPO made Elon Musk the world's first trillionaire, the company confirmed it will acquire Anysphere, the parent of the Cursor AI coding editor, in an all-stock deal valued at $60 billion. The price is roughly 16x Cursor's late-2025 private valuation, twice the round it was about to close, and the deal closes the loop on a curious April arrangement in which SpaceX had the right to buy Cursor for $60bn or pay $10bn for the partnership instead. This long read walks through the deal mechanics, the IPO-as-acquisition-currency story, the technology bet on Composer + Colossus, what xAI's collapse had to do with the timing, and the questions every developer who uses Cursor, Claude Code, Copilot, or any other AI coding tool should be asking this week.

On June 15, 2026, at VivaTech, the Staan search API opened self-service to any developer. It is the public face of European Search Perspective, the 50/50 Qwant and Ecosia joint venture that runs the European search index, and the launch lands in the gap left by Microsoft's August 2025 retirement of the Bing Search API and Google's constrained programmatic access. This long read walks through the pipeline, the three product tiers, the pricing, the sovereignty story and its limits, the 'American backup' fallback, and what it means for the AI agents and coding tools that depend on a fresh web index.

On the night of June 14, 2026, a fake Mistral product page leaked to X. By the morning of June 15 it had acquired 30 to 100 trillion parameters, an EU suspension, a FrontierMath 4 score of well beyond 100, recursive self-improvement, and an Andrej Karpathy hire notice. None of it was real. All of it landed. Here is the timeline, the community notes, and what it tells us about how the AI ecosystem absorbs a launch.

OpenRouter's Fusion API runs multiple LLMs in parallel, has a judge model extract consensus and contradictions, then synthesizes a single answer. On the DRACO deep research benchmark, a panel of budget models matched the best frontier model at half the cost, and a frontier panel surpassed every individual model tested. This analysis breaks down the architecture, the benchmark methodology, and why the synthesis step, not model diversity, is where most of the performance gain lives.

On June 12, 2026 at 5:21pm ET, Anthropic received an export control directive from the US government requiring it to disable Fable 5 and Mythos 5 for every user, including foreign nationals inside the United States. Anthropic is complying while publicly disputing the technical basis of the order.

On June 12, 2026, the US government ordered Anthropic to suspend Fable 5 and Mythos 5 for every foreign national worldwide. The directive did not target a country, and it did not target a capability. It targeted a class of person. This long read unpacks the legal mechanism, the 'US person' concept that decides who keeps access, why Anthropic chose to disable for everyone, what Project Glasswing loses, and what precedent this sets for the rest of the frontier model industry.

Fable 5 is the first widely-available Mythos-class model, with state-of-the-art results on software engineering, knowledge work, and vision, priced at $10/$50 per million tokens. Two days later, Anthropic reversed course on its stealth distillation throttling.

Moonshot AI released Kimi K2.7-Code, a 1T-parameter MoE coding model with 32B activated and a 30% drop in thinking tokens versus K2.6. Xiaomi followed with MiMo Code, another open-weights coding model. Here is what is actually on the model card.

The 2026 State of Web Dev AI survey of 7,258 developers reveals the AI-generated code share hit 54%, AI coding agents are rising, and 60% of respondents believe we are in an AI bubble.

The JavaScript runtime, bundler, and toolkit built by a 14-person team and relied on by millions of developers has been acquired by Anthropic. Bun will remain open source and MIT-licensed, but the roadmap now bends toward AI coding infrastructure.

OpenCode, the open source AI coding agent with 145K GitHub stars, has rebuilt its Desktop app on Electron after concluding that WebKit performance and bundled CLI startup issues made Tauri a poor fit for its all-TypeScript architecture.

Vercel engineers used AI coding agents to optimize Turborepo's Rust codebase, achieving 81–96% faster task graph computation. Here's the process, the wins, and the sharp limits they ran into.

JetBrains' March release turns on a service-based TypeScript engine by default, puts Junie, Claude Agent, Codex, and Cursor in the AI chat sidebar, sunsets Code With Me, and adds native Wayland support on Linux.

In one month, Claude Mythos Preview found thousands of zero-day vulnerabilities that survived decades of human review, in OpenBSD, the Linux kernel, FFmpeg, and every major browser. We dug into the technical details, the industry coalition, and what it means for every security team on the planet.

Starting April 24, 2026, GitHub will use interaction data from Free, Pro, and Pro+ Copilot users to train AI models, unless they manually opt out. Business and Enterprise tiers are unaffected. Here's what changes and how to protect your code.

YC and a16z-backed Rivet built an agent runtime on V8 isolates and WebAssembly that cold-starts in 4.8ms, 92x faster than E2B, at 1/17th the cost. We deeply researched the architecture, the benchmarks, and what it means for every agent framework.

On March 30–31 2026, developers discovered that the npm package @anthropic-ai/claude-code@v2.1.88 included a production source map file that exposed the full TypeScript source code, revealing undocumented multi-agent orchestration, a hidden Chrome MCP server, an internal query engine, a tool permission system, and a three-tier telemetry system.

Hermes Agent just added native OpenClaw migration. We deeply researched both platforms, here's everything that matters.

Bun v1.3.11 drops a 4MB smaller binary, ships Bun.cron for OS-level scheduled jobs, and marks a pivotal moment as the runtime joins Anthropic to power Claude Code and future AI coding tools.

LogRocket's March 2026 AI development tool rankings show major shifts as Antigravity climbs to second place and OpenAI's Codex re-enters the top five.

By early 2026, Claude Code held a 46% 'most loved' rating among developers, leaving Cursor at 19% and GitHub Copilot at 9%. But love ratings and usage rankings don't tell the whole story. Here's what each tool actually does well, and when to use which.

How a developer found a hidden model ID, sparked a global debate about attribution, and revealed how dependent the AI industry has become on Chinese open-source models.