AI Coding Tools Weekly: March 21, 2026

#This Week in AI Coding Tools

A busy stretch since our last weekly. Pricing models shifted at Cursor and GitHub Copilot, Anthropic shipped Claude Code features at an impressive clip, the Windsurf/Cognition acquisition settled into a new normal, and a free open-source tool called Roo Code has quietly become one of the most talked-about tools among power users.

Here’s what changed and what it means for your stack.

---

#Major Updates

#Claude Code: Voice, /loop, Channels, and 1M Context

Anthropic has been moving fast on Claude Code in 2026, and the March update is the most substantial yet.

Voice Mode (/voice): Push-to-talk voice coding is now live. Hold spacebar to speak, release to send. It sounds gimmicky but it’s genuinely useful for reviewing code while moving around, or for developers who think better out loud. 20 languages supported.

/loop Recurring Tasks: A lightweight cron job inside your session. Run loop 5m check the deploy and Claude Code checks every 5 minutes. Useful for monitoring builds, watching for test failures, or polling external status without writing a custom script.

Channels (Discord/Telegram): Connect Claude Code to your Discord server or Telegram account. Send it tasks from your phone, get results in DMs. This is a meaningful workflow change for developers who want to trigger background coding work while away from their desk.

1 Million Token Context Window: With Opus 4.6 (now the default model), Claude Code can ingest entire large codebases in a single session. This is the largest context window of any mainstream AI coding tool. For monorepo work or cross-service refactoring, it’s a real advantage.

Opus 4.6 as Default: Every new Claude Code session now starts with the most capable Claude model. Previously you had to explicitly select Opus; now it’s the baseline.

Our updated Claude Code tool page reflects these changes. The features listing and ratings have been updated accordingly.

---

#Cursor: New Pro+ Tier and Credit-Based Pricing

If you subscribed to Cursor last year and noticed your monthly requests felt lower, here’s why: Cursor switched from a request-count model to a credit-based system in June 2025.

How credits work:

• Each paid plan’s dollar amount equals your monthly credit pool ($20 Pro, $60 Pro+, $400 Ultra)
• Auto mode (Cursor-routed model selection) is completely free and unlimited
• Manually selecting a premium model (Claude Sonnet, GPT-4o, o1) draws from your credit pool
• The shift from “500 fast requests” to “$20 of credits” was controversial — many users perceived it as a cut (Cursor’s CEO issued a public apology)

New Pro+ tier ($60/month): Sits between Pro and Ultra, giving 3× the credit pool of Pro. Aimed at heavy Composer/Agent users who hit their Pro limits mid-month but don’t need the full Ultra allocation.

The free tier is now called Hobby (functionally the same, just renamed).

Our pricing file has been updated with all tiers: Hobby, Pro, Pro+, Ultra, Business.

---

#GitHub Copilot: Premium Request Limits

GitHub quietly introduced a significant change: premium request quotas are now part of the model.

What changed:

• GPT-4.1 and GPT-5 mini are “free” base models — unlimited use on any paid plan
• Advanced models (Claude Sonnet, Claude Opus, o3, GPT-4.5, Gemini Pro) now consume premium requests
• Pro plan: 300 premium requests/month
• Pro+ plan: 1,500 premium requests/month
• Extra requests: $0.04 each
• Heavier models have multipliers — Claude Opus 4.5 counts as 3× per request

What this means in practice: If you use Copilot heavily with Claude or o3, you’ll feel this. 300 requests sounds like a lot until you realize a Claude Opus session on a complex refactor can burn through 20-30 in an afternoon.

For developers who stick to the base GPT-4.1 model, nothing changes. For premium model users, the $39/month Pro+ tier is now worth considering — 1,500 requests gives genuine headroom.

We’ve updated the GitHub Copilot tool page and pricing with the new model.

---

#Windsurf: Post-Acquisition Update

The dust has settled on the Windsurf/Cognition story. Quick recap for anyone who missed it:

In July 2025, Windsurf’s CEO and co-founders departed for Google in a $2.4 billion licensing deal. Days later, Cognition AI (the team behind Devin) acquired the remaining Windsurf team and technology. About 30 staff were subsequently laid off.

Where things stand now:

• Windsurf continues as a product under Cognition ownership
• Combined ARR more than doubled post-acquisition
• Cognition is building toward an integrated Devin (autonomous agent) + Windsurf (IDE) experience
• Pricing simplified: Free (25 credits/month), Pro ($15/500 credits), Teams ($30), Enterprise ($60 with ZDR)
• The $200/month Max tier is no longer in the lineup

The Windsurf and Devin tool pages have both been updated to reflect the acquisition and new ownership context.

Should existing Windsurf users worry? The product is actively maintained and the Cognition backing is substantial ($10B+ valuation post-acquisition). The uncertainty is around roadmap direction, not product survival. If the Devin+Windsurf integration vision lands, it could become uniquely powerful.

---

#New Tool: Roo Code

We’re adding Roo Code to our tools directory this week. It’s been on our radar for months — GitHub star growth has been exceptional — and it’s now mature enough to recommend.

What it is: A free, open-source VS Code extension that gives you role-based AI agents. Architect, Code, Debug, Ask, and Custom modes, each with its own focus and behavior. BYOK — you bring your own API keys and pay your LLM provider directly.

Why it matters: Roo Code answers a real problem — hitting monthly usage walls mid-sprint. With BYOK, you’re never rate-limited by a subscription tier. You pay exactly for what you use, at your LLM provider’s published rates.

The standout feature: Boomerang Orchestration. Spawn sub-agents for specific tasks and collect results back. This is proper multi-agent coordination — an Architect agent designs the system, Code agents implement components, Debug agents verify them. Sophisticated stuff, available for free.

Who it’s for: Developers who want maximum control and BYOK flexibility. Power users who’ve outgrown Cline. Anyone doing local/offline AI development via Ollama. Budget-conscious developers who don’t want another $20/month subscription.

See the full Roo Code tool page for features, pricing breakdown, and our take.

---

#Pricing Updates Summary

Tool	What Changed
Cursor	Added Pro+ tier ($60/month); free tier renamed Hobby; credit-based model explained
GitHub Copilot	Premium request limits: 300/month Pro, 1,500/month Pro+; advanced models consume requests
Windsurf	Max tier removed; Enterprise now $60/user with ZDR; credits clarified (25 free, 500 Pro)
Tabnine	No free tier (dropped April 2025); enterprise-focused at $39–59/user

---

#What We’re Watching Next

• Cognition’s integrated Devin+Windsurf roadmap: The combined company has the pieces for something interesting. A fully autonomous agent that can also be an IDE experience is a different kind of product than anything on the market now.
• Claude Code adoption curves: Voice mode and Channels are differentiators that could pull in a new segment of developers. We’ll be tracking whether usage grows among non-CLI developers.
• Cursor credit model pushback: The June 2025 change created real community friction. Whether Cursor adjusts the model or developers migrate to alternatives (Windsurf, Roo Code) is worth watching.
• Roo Code vs Cline: The two main BYOK open-source options are diverging in focus — Cline leaning into simplicity, Roo Code into orchestration. Both are worth following.

---

Have a tool or stack to suggest? The site tracks what’s actually being used. Drop suggestions in our community channel.