The Tool Diff #2

Anthropic's Code with Claude event dominated the week. A SpaceX data center deal doubled rate limits overnight, while Cursor shipped self-improving code review and Windsurf put Devin in every terminal. Meanwhile, GitHub's June 1 token-billing deadline looms as Copilot still can't accept new individual signups, and a Microsoft Research paper found that even frontier models silently corrupt a quarter of your documents during long editing sessions.

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

GPT-5.5 Instant (May 6)

• Replaced GPT-5.3 Instant Mini as the default ChatGPT model.
• OpenAI claims 52.5% fewer hallucinations compared to the previous default.¹
• Available to all ChatGPT tiers including free.

Gemini 3.1 Flash-Lite (May 7)

• General availability. Google's smallest Gemini 3.x model.²
• Targets high-throughput, cost-sensitive workloads.

Mojo 1.0 Beta (May 8)

• Modular's Python-superset language hit its first beta milestone.³
• 354 points on Hacker News. The community is watching.

FEATURE UPDATES

Claude Code — Code with Claude Event (May 7)

• Rate limits doubled. Anthropic announced a deal with SpaceX's Colossus data center, adding enough compute to double rate limits across all Claude Code plans immediately.⁴
• Managed Agents. Claude Code can now spawn and coordinate sub-agents for parallel tasks. Agents run in isolated environments with their own tool permissions.⁴
• Dreaming. Background indexing mode where Claude Code analyzes your codebase overnight and pre-computes suggestions, diffs, and refactoring plans for the morning.⁴
• Auto mode GA. Previously in beta, Auto mode (Claude decides when to ask permission vs. act autonomously) is now generally available to all plans.⁴

Cursor 3.3 (May 6)

• Bugbot. Self-improving code review agent that runs on every PR. Cursor claims an 80% issue resolution rate across internal testing. Bugbot learns from accepted and rejected suggestions to improve its own accuracy over time.⁵
• Canvases. Visual workspace for system design, architecture diagrams, and multi-file planning. Think whiteboard inside your editor.⁵
• PR review agent integration with GitHub and GitLab.⁵

Devin for Terminal (May 5)

• Cognition released Devin as a standalone CLI agent, separate from the cloud workspace product.⁶
• Windsurf is bundling it as the default terminal agent, calling it 30% more token-efficient than Cascade.⁷

VS Code 1.119 (May 8)

• The git.addAICoAuthor bug from Tool Diff #1 is fixed. Copilot co-author lines are no longer injected into commits where Copilot wasn't used.⁸

Amp CLI (May 7)

• Rebuilt from the ground up. New architecture, faster startup.⁹
• GPT-5.5 available in deep reasoning mode.

ServiceNow Build Agent (May 6)

• Expanded from VS Code-only to all major IDEs: Cursor, Windsurf, Claude Code, and GitHub Copilot.¹⁰
• Enterprise-focused agent for building on the ServiceNow platform.

Coder Agents Beta (May 8)

• Self-hosted, model-agnostic coding agents. Bring your own LLM, run on your own infrastructure.¹¹
• No usage limits through September 2026.

PRICING & LIMITS

GitHub Copilot — June 1 Token Billing

• Token-based billing goes live June 1. Each credit costs $0.01. Monthly credit pools replace flat-rate unlimited usage.¹²
• Individual signups remain paused. No timeline for reopening. Students and indie devs who missed the window are still locked out.¹³
• Existing subscribers keep their plans but transition to the new billing automatically.

DeepSeek V4 Pro — Promo Extended

• The 75% promotional discount on V4 Pro, originally set to expire May 5, has been extended through May 31.¹⁴
• Effective price during promo: $0.44 / $0.87 per 1M input/output tokens.

RESEARCH

Microsoft Research: LLMs Corrupt Documents

A Microsoft Research paper found that frontier LLMs silently corrupt approximately 25% of document content during long editing workflows. The corruption is cumulative: models introduce small factual errors, formatting changes, and semantic drift with each pass. Over a multi-step editing session, the compounding errors degrade the source material significantly. The paper tested GPT-5.5, Claude Opus 4.7, and Gemini 3.0 Ultra. All showed similar degradation patterns.¹⁵

CONTROVERSY

Opus 4.7 Tokenizer Backlash

The backlash over Anthropic's Opus 4.7 tokenizer continues into its second week. Developers measuring real-world token counts report 12–27% cost increases compared to the same prompts on the previous tokenizer. The new tokenizer splits common programming tokens differently, inflating counts on code-heavy workloads. Anthropic has not addressed the discrepancy publicly. The original Hacker News thread hit 837 points with 660 comments, and frustration has not cooled.¹⁶

USER REACTION

• Mojo 1.0 Beta: 354 pts on HN. Cautious optimism. Most comments focused on whether the Python compatibility story holds up in practice.³
• Opus 4.7 tokenizer: 837 pts / 660 comments. The thread became a clearinghouse for developers comparing before-and-after token counts across different codebases.¹⁶
• Code with Claude event: Mixed reception. Rate limit doubling was well received. Dreaming drew skepticism about privacy implications of overnight codebase analysis.
• Bugbot (Cursor): Interest in the self-improving loop. Developers asking whether it learns from their codebase specifically or from aggregate data.