AI Dev Weekly is a Thursday series where I cover the week's most important AI developer news, with my take as someone who actually uses these tools daily.
The flat-rate AI subscription era ended this week. Anthropic pulled Claude Code from the $20 Pro plan. GitHub froze all new Copilot signups. And while Western companies were busy raising prices, two Chinese labs dropped frontier models within 48 hours of each other. Let's get into it.
Claude Code removed from Pro plan
Anthropic quietly removed Claude Code from the $20/month Pro plan on April 21. The pricing page now shows an "X" next to Claude Code for Pro subscribers. Access starts at Max ($100/month).
Anthropic's head of growth called it "a small test on ~2% of new prosumer signups." But the public pricing page already reflects the change for everyone. Sam Altman's response on X: "ok boomer."
The real reason: engagement per subscriber surged after Opus 4, Cowork, and long-running agents. Pro subscribers at $20/month are consuming 10x or more in token value. The math doesn't work.
My take: This was inevitable. Unlimited AI coding for $20/month was never sustainable. If you're on Pro, you still have access for now. But start planning for either Max ($100/month) or cheaper alternatives like Kimi K2.6 ($0.60/M tokens) or MiMo V2.5 Pro ($1/M tokens).
GitHub Copilot freezes all new signups
GitHub paused new registrations for Copilot Pro, Pro+, and Student plans on April 20. Only the Free tier accepts new users. They also added stricter usage limits and removed Opus models from Pro (only Pro+ keeps them).
The reason: "unsustainable compute demands from AI-powered coding agents." Same story as Anthropic. Agentic AI usage broke the pricing model.
My take: Two of the three biggest AI coding platforms raised prices or froze signups in the same week. The third (Cursor) is probably next. The era of $10-20/month unlimited AI coding is over. Open-source and Chinese models are the hedge.
Kimi K2.6 launches with 300-agent swarm
Moonshot AI released Kimi K2.6 on April 20. The highlights:
- 80.2% SWE-Bench Verified (matching Claude Opus 4.6)
- 300 sub-agent swarm (up from 100 in K2.5)
- 54.0% on HLE-Full with tools (beating GPT-5.4's 52.1%)
- $0.60/M input tokens (25x cheaper than Opus)
- Modified MIT license (open weights)
- Available on OpenRouter and Cloudflare Workers AI
The agent swarm is the standout feature. K2.6 scored 86.3% on BrowseComp (Agent Swarm) vs GPT-5.4's 78.4%. For coding agent workloads, K2.6 is the strongest open-source option available.
My take: K2.6 is the first open-source model to genuinely match Opus 4.6 on coding benchmarks. At 25x cheaper. The timing with Anthropic's price hike is not a coincidence. See our K2.6 vs Opus 4.6 comparison.
MiMo V2.5 Pro: 40-60% fewer tokens than Opus
Xiaomi dropped MiMo V2.5 Pro on April 22, just 48 hours after K2.6. The headline number: 40-60% fewer tokens than Opus 4.6 at comparable capability.
- 57.2% SWE-bench Pro
- 64% Pass^3 on ClawEval with only ~70K tokens per trajectory
- 1,000+ tool calls in single sessions
- Built a complete SysY compiler in Rust in 4.3 hours (672 tool calls, 233/233 tests)
- Works with Claude Code as a harness
- Coming open-source soon
The token efficiency is the real story. Same capability, half the tokens, fraction of the price. The V2.5 Standard model adds native multimodal (image, audio, video) and actually outperforms V2-Pro on some agent benchmarks.
My take: V2.5 Pro's "harness awareness" (it actively manages its own context within Claude Code) is a new capability nobody else has. Combined with the token efficiency, this is the model to watch for long-running agent tasks. See our full V2.5 series guide.
The flat-rate subscription is dead
Three data points in one week:
- Anthropic removes Claude Code from $20 Pro
- GitHub freezes all Copilot signups
- Both cite "unsustainable compute demands"
The pattern is clear. Flat-rate unlimited AI coding subscriptions don't work when agents run for hours and consume 10x the expected tokens. Expect token-based billing everywhere within 6 months.
The winners: Chinese models (Kimi K2.6, MiMo V2.5 Pro, Qwen 3.6 Plus) that were already priced per-token at 10-25x less than Western alternatives. If you haven't explored them yet, now is the time. See our Chinese AI models ranking.
Quick hits
- OpenAI Workspace Agents: ChatGPT now has workspace agents for enterprise teams. Not relevant for individual developers yet.
- OpenAI Privacy Filter: New privacy filter for enterprise data. Good for compliance, not a developer tool.
- Vercel data breach: Vercel disclosed a breach that exposed some user data. Check your account if you use Vercel.
What I'm watching next week
- Whether Claude's serverless function limit forces architectural decisions (it broke one of our race agents)
- How MiMo V2.5 Pro performs in real-world agent tasks (we just upgraded our Xiaomi race agent to V2.5 Pro)
- Whether any race agent gets its first paying customer
See you next Thursday. If you found this useful, subscribe to AI Dev Weekly for the full archive.
Originally published at https://www.aimadetools.com