The Tool Diff #1

This is The Tool Diff, a weekly roundup of what actually changed in the AI developer tools you use. No hype, no hot takes. Just the diff: what shipped, what costs more, what broke. Every Monday.

NEW MODELS

DeepSeek V4 Pro + Flash (April 24)

• 1.6T total params, 49B active (Pro) / 284B total, 13B active (Flash).
• MIT licensed, open-weight, 1M token context.
• Matches GPT-5.5 on most coding benchmarks at a fraction of the cost.²

Kimi K2.6 (April 20)

• Moonshot AI. 1T params, 32B active. Open-weight, Modified MIT.⁴
• 80.2% SWE-Bench Verified. 256K context window.
• Agent swarm: scales to 300 sub-agents across 4,000 coordinated steps.

GPT-5.3 Instant Mini

• New ChatGPT fallback model, replaces GPT-5 Instant Mini.⁵

FEATURE UPDATES

Claude Code (April 28)

• alwaysLoad option for MCP servers — skip tool-search deferral.
• claude ultrareview CLI subcommand for non-interactive use in CI.
• PostToolUse hooks can now replace output for all tools (was MCP-only).
• Windows: PowerShell works natively, Git Bash no longer required.
• OAuth login fixes for IPv6 devcontainers, slow connections, proxy setups.⁶

OpenAI Codex

• Quick reasoning controls in TUI: Alt+, lowers reasoning, Alt+. raises it.
• Multi-environment support: choose environment and working directory per turn.
• First-class Amazon Bedrock auth (AWS SigV4 signing).
• Updated bundled docs for GPT-5.5 and gpt-image-2.¹¹

Pi (pi.dev) v0.71–0.72 (April 30 – May 2)

• Removed built-in Google Gemini CLI and Google Antigravity support.
• Added Cloudflare AI Gateway, Moonshot AI, and Xiaomi MiMo as built-in providers.
• WebSocket cached transport for OpenAI Codex — keeps connection open, sends only new items after first request.
• Model thinking level metadata — models declare which thinking levels they support.
• Mistral Medium 3.5 and DeepSeek V4 Flash thinking support added.¹²

VS Code 1.117 (April 23) + 1.118 (April 27)

• BYOK for Copilot Business/Enterprise — bring your own API keys.
• Remote control of Copilot CLI sessions from GitHub.com or mobile.
• Semantic codebase search across any workspace.
• Chat session insights — generates standup reports from your history.
• Enterprise lockdown for AI features.⁷

Windsurf

• Devin Local agent — same harness as terminal, 30% more token-efficient than Cascade.
• GPT-5.5 and GPT-5.2 added to model roster.
• New terminal CLI workflow with cloud handoff.⁸

ChatGPT

• Model picker moved into the composer.
• Thinking effort controls moved into model picker.
• Advanced Account Security — phishing-resistant sign-in, login alerts, automatic training exclusion.⁵

PRICING & LIMITS

GitHub Copilot

• Pausing new signups for individual plans (students, indie devs).¹
• Opus models removed from Pro plans. Opus 4.7 stays in Pro+.
• Model availability tightened; usage limits adjusted.

DeepSeek V4

• V4 Pro: $1.74 / $3.48 per 1M input/output tokens. 97% below GPT-5.5 pricing.²
• V4 Flash: $0.14 / $0.28 per 1M tokens.
• 75% promo discount on V4 Pro through May 5.
• For comparison: Claude Opus 4.7 is $5/$25. GPT-5.5 is $5/$30.

Cursor

• Now fully on credit-based system. Monthly credit pool = plan price in dollars.³
• Six tiers: Hobby, Pro, Pro+, Ultra, Teams, Enterprise.
• Security Review in beta (Teams/Enterprise only).

CONTROVERSY

VS Code git.addAICoAuthor

VS Code has been injecting "Co-Authored-by: Copilot" into git commits even when Copilot wasn't used. The setting shipped enabled by default in v1.110 (March). One developer wrote their own commit message from scratch — the co-author line still appeared in git history. Microsoft tagged it as an important bug, targeting v1.119.⁹

USER REACTION

• VS Code co-author bug: 1,458 pts on HN. Developers called it a "trust violation" and questioned whether it inflates Microsoft's "30% of code is AI-written" claim.¹⁰
• DeepSeek V4: 566 pts / 343 comments. Split between "commoditization is here" and "benchmark gaming."
• Kimi K2.6: 293 pts / 156 comments. Interest in agent swarm capabilities, skepticism about benchmark reproducibility.