The flat-rate AI coding era just ended (and Opus 4.7 is the reason)

Hey — Three things broke in April that change how we all code: Anthropic shipped Opus 4.7, GitHub froze new Copilot Pro signups, and Cursor 3 turned multi-agent coding from "demo trick" into "default workflow." If you've been ignoring AI tooling news because it all sounds the same, this is the month to look up. The flat-rate subscription era is over — coding tools are now metered compute, and the model under the hood matters more than ever. ## This Week in AI Coding - **Claude Opus 4.7 is GA on everything.** Anthropic shipped it April 16 across Claude, the API, Bedrock, Vertex, and Foundry — same $5/$25 per-million pricing as 4.6, but a 12-point jump on CursorBench and the first Claude with high-res vision (2576px input). Notable change: it follows instructions more literally and stops "silently generalizing" what you asked for. Worth the migration if you do agentic work. - **The flat-rate party is over.** April 20: GitHub paused new signups on Copilot Pro / Pro+ / Student. April 21: Anthropic pulled Claude Code from the $20 Pro plan. Anthropic also chopped the server-side prompt cache from 1 hour to 5 minutes, and Cursor moved frontier models behind Max Mode for legacy plans. Translation: agentic tools are compute products, and compute products always meter. - **Cursor 3 ships an Agents Window.** Run multiple AI agents in parallel — local, worktree, SSH, or cloud — from one window. This is the version of "agentic coding" that actually matches the marketing. - **OpenAI's Agents SDK got serious.** April 15 update: model-native harness, configurable memory, sandbox-aware orchestration, Codex-like filesystem tools. Built-in support for E2B, Modal, Vercel, Daytona, Cloudflare. This is the moment "agent SDK" stopped being a wrapper and became infrastructure. - **MCP crossed 97M installs in March.** It's now foundational infrastructure for agentic coding, not a curiosity. If you haven't built an MCP server yet, you're missing the easiest leverage point in the stack. - **Salesforce shipped Headless 360.** Every Salesforce capability as an API, MCP tool, or CLI command — 100+ new tools immediately. The enterprise vendors are catching the MCP wave faster than expected. - **Anthropic's Mythos exists, and you can't have it.** Anthropic confirmed it built a model called Mythos that's significantly better than anything public, then locked it behind a 50-company "Project Glasswing" firewall. OpenAI countered with GPT-5.5-Cyber for cyber defenders. Frontier capability is consolidating into curated access — worth watching if you build in regulated spaces. ## What We Shipped - **Opus 4.7 migration** across the whole VeriduxLabs stack — agents, Claude Code workflows, Kynvo render pipeline. The literal-instruction-following change cleaned up a few flaky agents that were inferring requests we didn't make. - **YouTube pipeline cleared its backlog.** [@endofcoding](https://www.youtube.com/@endofcoding) is back on a daily Sun-Thu cadence, 110+ videos live. The latest covers AI hiring trends, vibe coding for beginners, and the "$0/month AI coding stack" that keeps surprising people. Sub if you haven't. - **JaaS (Johnny-as-a-Service) shipped a 5-agent overhaul.** New CAIO, AI/ML Architect, enhanced QA + CFO, plus a heartbeat system and procedural memory. 1,867 lines. Kanban view next. - **LemonSqueezy went live across CyberOS, the Vibe Coding Ebook, and LLMHire.** Paddle is officially deprecated. Building takes weeks, switching payment processors takes a single afternoon — funny how that works. ## Tool of the Week: Claude Opus 4.7's File-System Memory The understated win in 4.7 is *better file-system memory for agents*. If your agent maintains a scratchpad, notes file, or structured memory store across turns, 4.7 is materially better at writing useful notes to itself and using them later. I've been moving long-running agents over this week and the difference shows up in 5+ turn tasks — fewer context-window stalls, fewer "what was I doing?" moments. If you build agents, the upgrade pays for itself within a day. ## Quick Links - [Anthropic — Introducing Claude Opus 4.7](https://www.anthropic.com/news/claude-opus-4-7) - [The New Stack — Cursor, Claude Code, and Codex are merging into one AI coding stack](https://thenewstack.io/ai-coding-tool-stack/) - [TechCrunch — Google to invest up to $40B in Anthropic](https://techcrunch.com/2026/04/24/google-to-invest-up-to-40b-in-anthropic-in-cash-and-compute/) - [Activated Thinker — The flat-rate AI coding subscription era is ending](https://medium.com/activated-thinker/the-flat-rate-ai-coding-subscription-era-is-ending-what-github-copilot-claude-code-and-cursor-9763e043a63a) --- Building in public from Tel Aviv — **Guy** [endofcoding.com](https://endofcoding.com) · [vibecodingebook.com](https://vibecodingebook.com) · [llmhire.com](https://llmhire.com) · [cyberos.dev](https://cyberos.dev) · [YouTube @endofcoding](https://www.youtube.com/@endofcoding) *Reply hit. I read everything.*