Key Signals

• The US government ordered Anthropic to disable Fable 5 and Mythos 5, disrupting Claude Code workflows worldwide. A Commerce Department export control directive citing national security concerns forced Anthropic to shut off its most capable models for all customers Friday evening—just three days after launch. Amazon researchers reportedly demonstrated a jailbreak that circumvented Fable 5's safety classifiers for cybersecurity topics, and CEO Andy Jassy escalated the findings to the White House. Any team that had already integrated Fable 5 into their Claude Code pipelines must now fall back to Opus 4.8 or Sonnet. Anthropic calls the issue "a misunderstanding" and says the vulnerabilities are minor capabilities already present in GPT-5.5. [1][2][3]

• GitHub Copilot CLI v1.0.62 shipped a sweeping update that turns the terminal agent into a plugin-extensible platform. Plugins can now ship full extensions installable via the marketplace, session-scoped extensions and canvases allow richer agent interactions, and subagent model/reasoning/context settings are now user-configurable through /subagents. The release also adds corporate proxy authentication via Kerberos/SPNEGO, a file tree sidebar with inline comment editing in /diff view, and the ability to schedule recurring slash commands with /every and /after. This is one of the largest single releases in Copilot CLI's history, consolidating features from three pre-releases in one week. [5]

• WebMCP entered Chrome 149 origin trials, giving browser-based AI agents a 90% reduction in token usage. The new standard lets websites expose typed, named tools directly to in-browser agents, eliminating the expensive screenshot-action-screenshot loop used by Playwright and Chrome DevTools MCP servers. Early implementers report dramatically improved speed and determinism alongside the token savings. The spec offers both a Declarative API and an Imperative API. [6]

• Model triage has become the critical developer skill as Fable 5's economics expose a 6x cost gap for routine work. Mitchell Hashimoto's head-to-head tests showed Fable 5 costing $9 and taking 40 minutes for tasks GPT-5.5 completed at $1.50 and Zhipu's GLM-5.1 at under $1—with equally acceptable results. The emerging pattern is to use expensive frontier models only for planning and review while routing execution to cheaper alternatives. Boris Cherny captured the trend: "I don't prompt Claude anymore. I write loops and the loops do the work." [7][10]

• Loop-driven development is replacing prompt-driven and spec-driven agentic workflows, raising new verification challenges. In this third era, developers write automated programs that prompt agents, evaluate responses, and retry—running on schedules rather than human attention. For cloud-native teams, the key bottleneck is that truthful feedback requires a real runtime, not just local tests. The four layers needed are: runtime, verification interface, feedback, and control. [8]

AI Coding News

• Fable 5 and Opus 4.8 produced near-identical results in a real-world head-to-head test, with Fable costing more than double. On a reasoning task, Fable 5 cost $2.55 vs Opus's $2.18 with slightly sharper historical analysis. On a coding task, Fable cost $12.19 vs Opus's $5.80—both finding the same bugs and implementing equivalent fixes. Fable's safety classifiers triggered mid-task and silently switched the session to Opus, meaning ~15% of Fable's "own" coding test was actually completed by Opus. The practical conclusion: for solo developers, Opus delivers most of the value at a fraction of the cost. [10]

• A practitioner shared an evolved Claude Code workflow centered on conversation-first development rather than plan mode. The approach: discuss the problem with the agent, have it distill the discussion into GitHub issues, then assign fresh parallel agents to each issue in separate worktrees. The key insight is that plan mode "jumped to solutions" without validating assumptions, whereas open-ended conversation lets both human and agent contribute unique insights—the human brings problem understanding and architectural memory, the agent brings up-to-date codebase knowledge and helpful pushback. [11]

• HashiCorp's Terraform MCP Server reached general availability, enabling AI coding agents to interact with infrastructure workflows. The open-source server lets tools like Cursor, Claude, and Gemini discover approved Terraform modules, generate compliant IaC code, analyze plan files in natural language, and query workspace state ("Which workspaces haven't been updated in 90 days?"). It enforces existing Terraform auth without exposing credentials to agents, and supports both stdio and StreamableHTTP transports. [9]

• A new tool called repo-slopscore launched for detecting AI/LLM contributions in git repositories through commit history analysis. Written in Rust and hosted on Codeberg, the tool has already scanned over 3,275 repositories including projects like Chromium, VLC, and OpenZFS. It represents the growing ecosystem of tools for auditing "vibe-coded" contributions in open-source projects—a response to concerns about undisclosed AI-generated code entering codebases without appropriate review. [17]

• A Verge writer documented the promise and pain of vibecoding a yard care Android app with Google Gemini in AI Studio. A single natural language prompt produced a functional app in minutes, but the author spent hours iterating on broken date pickers, illegible color schemes, and missing edit functionality through tedious back-and-forth. The AI plant doctor feature worked excellently out of the box, correctly diagnosing that landscape fabric was killing plants. The piece illustrates that current vibecoding tools excel at one-shot capabilities but struggle with iterative refinement. [18]

Feature Update

• GitHub Copilot CLI v1.0.62 released with plugin-shipped extensions, subagent configuration, and session-scoped canvases. Plugins can now deliver installable extensions via the marketplace. Users can configure subagent model, reasoning effort, and context tier through /subagents. Additional highlights: content search with n/N navigation in diff view, /app slash command for browser fallback, automatic Kerberos/SPNEGO authentication through corporate proxies, file tree sidebar in /diff, and scheduling with /every and /after. Shell commands now use lightweight process spawning instead of a PTY. Over 30 bug fixes cover MCP servers, permissions, plugin hooks, and accessibility. [5]

• Claude Code v2.1.177 shipped as a maintenance patch on June 13. No user-facing changelog was provided, suggesting internal fixes or infrastructure changes. The previous day's v2.1.176 had added multi-language session titles, footer link regex settings, improved Bedrock credential caching, and fixes for Fable 5 auto-mode fallback and hook path matching. [16]

• OpenCode v1.17.5 and v1.17.6 released with MCP session recovery and compatibility improvements. v1.17.5 adds external browser OAuth for Snowflake Cortex, improved project copy management in v2, recovered expired MCP sessions, and cleared stale MCP connections. v1.17.6 follows with a targeted fix declaring OpenCode's supported MCP client capabilities, improving compatibility with third-party MCP servers. Five community contributors participated across both releases. [12][13]

• Gemini CLI v0.48.0-nightly shipped with Antigravity CLI migration commands and MCP tool discovery fixes. The nightly build adds documentation and migration commands preparing users for the transition from Gemini CLI to Google's new Antigravity CLI platform (free/individual users must migrate before June 18). Technical fixes include atomic MCP tool discovery updates, zero-quota fast-fail to prevent retry loops, Vertex AI model mapping corrections, and multi-line escaped quote handling. [14]

• OpenAI Codex CLI released two rapid-fire Rust alpha builds (0.140.0-alpha.17 and .18) with no public changelogs. The daily alpha cadence indicates active development of the Rust-based CLI rewrite, continuing from 19 alpha releases in the 0.140.x series over the past week. No detailed feature notes are published for individual alphas during this development phase. [15]