Claude Code 2.1.116–2.1.117: Native Binary, Opus 4.7 Context Fix, Performance
Claude Code versions 2.1.116 (April 20) and 2.1.117 (April 22) bring a structural change to how the CLI operates: the main process now spawns a native platform binary instead of the bundled JavaScript runtime. On macOS and Linux native builds, the Glob and Grep tools are replaced by embedded bfs (breadth-first search) and ugrep (ultra-fast grep) binaries, delivering significantly faster file and content searches compared to the JavaScript implementations.
Version 2.1.117 also fixes a long-standing issue affecting Opus 4.7 sessions: the context window calculation was incorrectly using 200K tokens instead of the actual 1M token limit, causing artificially inflated context usage warnings. The OAuth token refresh flow was fixed to handle mid-session 401 errors gracefully — previously these resulted in "Please run /login" errors that interrupted long-running sessions. WebFetch no longer hangs on large HTML pages.
For performance, /resume on large sessions (40 MB and above) is now up to 67% faster in 2.1.116, and MCP server startup is faster due to concurrent connection establishment replacing sequential startup. The default effort level for Pro and Max subscribers on Opus 4.6 and Sonnet 4.6 was raised from medium to high, meaning more thorough analysis by default. Plugin install now auto-installs missing dependencies, and plugin marketplace restrictions such as blockedMarketplaces and strictKnownMarkets are enforced at install and update time.
Read more — Claude Code Docs
GitHub Copilot Plans Restructured: Pro and Pro+ New Signups Paused
GitHub announced on April 20 that new sign-ups for Copilot Pro, Copilot Pro+, and student plans are temporarily paused. The change was framed as part of "ongoing efforts to ensure service reliability and a sustainable" service. Existing subscribers are unaffected and can still upgrade between tiers. Copilot Free remains available to new users.
The restructuring also introduced formal usage limits: Copilot Pro+ now delivers more than five times the usage allocation of Pro. Both VS Code and Copilot CLI now display warnings as users approach their limits. Copilot Pro no longer includes access to Opus models — Claude Opus 4.7, which began rolling out on Copilot on April 16, remains a Pro+ exclusive. Earlier Opus versions are being phased out from all tiers.
Two additional GA milestones arrived in the same window: auto model selection became generally available on April 17 in Copilot CLI across all plans, allowing the service to choose the most efficient model automatically. A new gh skill command launched on April 16, providing terminal-native discovery, installation, and management of AI agent capabilities without leaving the shell.
Read more — GitHub Changelog
Stanford AI Index 2026: Coding Benchmarks Saturate, Entry-Level Jobs Decline
The 2026 Stanford AI Index report from HAI documents the fastest single-year jump in AI coding capability on record. On SWE-bench Verified — a benchmark requiring models to resolve real GitHub issues — scores climbed from 60% to nearly 100% in a single year. AI agents on the OSWorld benchmark improved from 12% to 66% task success, demonstrating rapid progress at desktop-level autonomous operation.
The employment picture for software developers is mixed. Employment among developers aged 22–25 has declined nearly 20% since 2024, with the report noting that jobs with higher AI exposure show similar patterns. Mid-career and senior developers show stable or growing headcount, suggesting that experience and specialized expertise buffer against displacement. The report is careful to note that broader economic factors complicate attributing the trend solely to AI.
On the ecosystem side, global AI investment reached $581.7 billion in 2025, up 130% from the prior year. GitHub now hosts 5.58 million AI-related projects, a 23.7% increase. Model transparency is declining: the Foundation Model Transparency Index saw average scores drop from 58 to 40 points year over year, even as the most capable models are the least transparent. The report was published in April 2026 by Stanford's Human-Centered AI institute.
Read more — Stanford HAI
Links & Sources
You May Also Enjoy
