🧭 Claude Code v2.1.142: Opus 4.7 Is Now the Fast Mode Default

Claude Code v2.1.142, released May 14, quietly lands one of the more significant model-default changes since Fast Mode launched: Opus 4.7 is now the default model for Fast Mode sessions, replacing Opus 4.6. Opus 4.7 — released on April 16, 2026 — brings native 1 million token context (no preview flag), a new xhigh effort level, 13% improvement in coding-benchmark task resolution, and 3.75 MP image support. Since pricing is unchanged ($5/M input · $25/M output), the switch to a more capable model costs nothing extra.

What else shipped in v2.1.142

• SKILL.md plugins auto-surfaced — Any plugin that includes a root-level SKILL.md file is now automatically listed in the /skills pane, removing the need for a separate skill registration step. This brings the plugin and skill systems one step closer to a unified extension model.
• LSP servers shown in /plugin details — The plugin details panel now lists which Language Server Protocol servers a plugin provides, helping you evaluate language-tooling plugins before installing.
• MCP_TOOL_TIMEOUT fix for remote HTTP/SSE — The MCP_TOOL_TIMEOUT environment variable now correctly raises the per-request fetch timeout for remote HTTP and SSE MCP servers. Previously, the setting was parsed but not applied to the actual request layer, causing unexpected timeouts on slow external tools.
• Background session fixes — Sessions woken from idle now correctly recognise pre-existing git worktrees, and a daemon exit bug triggered by binary upgrades has been resolved.

export CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1

🧭 Claude Code v2.1.143: Plugin Marketplace Gets Cost Estimates, Dependency Guards & PowerShell Improvements

Released May 15, v2.1.143 continues the plugin ecosystem buildout started in v2.1.142, adding two features that make the marketplace practical for production use: projected context cost estimates in the browse pane, and plugin dependency enforcement that prevents you from accidentally breaking a plugin's dependencies.

Plugin marketplace: cost estimates and dependency safety

• Projected context cost in browse pane — The /plugin marketplace now shows estimated per-turn and per-invocation token costs alongside each plugin listing. This lets you assess the context overhead a plugin adds before installing it — critical for high-frequency agentic sessions where context budget matters.
• Dependency enforcement on disable — claude plugin disable <name> now refuses to proceed if another currently-enabled plugin declares a dependency on the target. You'll see a clear error message naming the dependent plugin. This prevents the silent, hard-to-diagnose breakage that occurs when shared tooling plugins are accidentally removed.

Other notable changes

• worktree.bgIsolation: "none" — A new setting lets background sessions edit the working copy directly, without needing to call EnterWorktree. Useful for background tasks that should remain on the current branch rather than an isolated worktree.
• PowerShell execution policy bypass — On Windows, the PowerShell tool now automatically passes -ExecutionPolicy Bypass, fixing the common scenario where restrictive system policies blocked Claude Code from running scripts. Opt out with CLAUDE_CODE_POWERSHELL_RESPECT_EXECUTION_POLICY=1 if your org requires policy enforcement.
• PowerShell enabled by default on Windows for Bedrock, Vertex, and Foundry users — bringing Windows deployments on managed infrastructure to parity with local CLI usage.
• Auto mode in Shift+Tab cycle — In attached agent sessions, Shift+Tab now cycles through model/effort options and includes auto mode, making it quicker to downshift to a lighter configuration mid-session.
• Background session model persistence — Sessions now preserve the model and effort level after waking from idle, preventing unintended resets to defaults.

🧭 Ramp Data: Anthropic Now Has More Business Customers Than OpenAI — 34.4% vs 32.3%

According to spend data compiled by corporate fintech platform Ramp, Anthropic has overtaken OpenAI in business customer penetration for the first time. As of May 2026, 34.4% of Ramp's business clients are paying Anthropic customers, compared to 32.3% for OpenAI. Because Ramp measures actual corporate card and invoice spend — not self-reported surveys — the data represents real purchasing decisions made by real finance and engineering teams.

Why this milestone matters

• It's spend-verified, not survey-reported — Ramp's methodology tracks expense data from participating companies, which means it captures businesses that have gone through procurement and are actively paying — a higher bar than "we tried it" usage metrics.
• The gap is narrow but the direction is clear — OpenAI still holds the mindshare advantage from GPT's consumer breakthrough, but Anthropic's enterprise product focus (longer context, stronger instruction-following, stronger safety posture for legal/compliance teams) appears to be translating into wallet share.
• Claude Code is the wedge — Multiple data points from the past two weeks — Salesforce's $300M token commitment, PwC's 30,000-strong Claude certification rollout — identify developer and engineering tooling as the primary adoption driver. Teams adopt Claude Code individually, get results, and the company-wide Claude API spend follows.
• Both are growing — The shift in relative share doesn't mean OpenAI is shrinking. Enterprise AI spending overall is rising sharply; Anthropic is simply growing faster in this cohort.