What changed, why it matters
Hand-curated tracker for the AI tools people actually use. Not every release — just the ones that change how you should work.
Tracking: Cursor · Claude · Claude Code · ChatGPT · Gemini · Perplexity · Cline · Codex · Windsurf
Opus 4.7 with 1M-token context window
Anthropic shipped a 1M-context variant for Claude Code, available via the standard CLI. For long-session refactors and large monorepos, this changes what fits in one prompt — you can now feed 30+ files plus full repo conventions without aggressive trimming. Cost per session goes up, but cache hits help. Worth testing on any codebase where you've been splitting work artificially.
Agent mode (beta) lands
Cursor 0.45 introduces an autonomous agent that plans and executes multi-file changes from a single prompt. For teams already using Claude Code, the comparison shifts: Cursor's agent is faster on navigation but weaker on long-context retention. The cmd+K shortcut moved — if you live in keyboard shortcuts, expect a half-day adjustment.
Gemini 2.5 Pro: better multimodal, same price
Google bumped Gemini 2.5 Pro to a refreshed multimodal model, sharper on document parsing and visual reasoning. Pricing held. For research workflows that involve PDFs and images, this is now arguably the most cost-effective option per pass — though still trailing Claude on long-form code reasoning.
Plus tier rate limits dropped 40%
OpenAI quietly cut Plus-tier message caps for o-series models — most heavy users will hit the cap by lunch. The official line is "balancing infrastructure load." If you rely on ChatGPT Plus for daily coding work, the practical fix is to switch to API + a UI like LibreChat, or move to Claude Pro which still has roomy limits.
Spaces with shared system prompts
Perplexity now supports team Spaces with a shared system prompt and source allowlist. Useful for company-internal research where you want every team member to query the same trusted set of docs. Trial period is 30 days; not free indefinitely.
Sonnet 4.6 — faster, cheaper, equal-or-better on coding
Anthropic's mid-tier got a real upgrade: same Sonnet pricing, ~30% faster output, and benchmark parity (or better) with the previous Opus on most coding evals. If your team uses Sonnet for chat and Opus for harder tasks, this is the moment to test whether you can drop Opus for most workflows.
Cline 3.0: native MCP support
Cline now supports the Model Context Protocol natively, meaning any MCP server (databases, browsers, APIs) plugs in without custom adapters. For solo developers building agents, this dramatically reduces glue code. Setup is still less polished than Claude Code's, but the ecosystem is broader.
OpenAI Codex (re-release) ships
OpenAI brought Codex back as a Cursor-class CLI agent, tightly integrated with their newer reasoning models. Strongest at long-form refactors and test generation. Weakest at conversational debugging where context is implicit. Worth a look if you're already on the OpenAI stack — but switching from Claude Code is non-trivial.
Want this in your inbox? Tool changes are bundled into the weekly digest.