April's First 72 Hours: Cursor 3, Gemma 4, Free Qwen 3.6, and the Agent Push

Three major releases. 72 hours. Right after April Fools. I verified - they’re all real.

Cursor 3: The Agent Manager

Cursor shipped a complete redesign on April 2nd. The editor is still there, but the new default interface is built around managing agents, not editing files.

What changed:

• Agent-first UI - Built from scratch. You manage a fleet of local and cloud agents from a unified sidebar
• Multi-repo workspaces - Agents can work across different repos simultaneously
• Unified agent sources - Agents kicked off from mobile, web, desktop, Slack, GitHub, or Linear all appear in one place
• Built-in browser - Agents can open and interact with local websites directly
• Plugin marketplace - One-click install for MCPs, skills, and subagents

The timing is telling. Claude Code reportedly holds 54% of the AI coding market according to Menlo Ventures. OpenAI’s Codex keeps setting new benchmarks. Cursor needed a big swing, and “you’re the manager now” is a clear bet on where development is heading.

You can try it by upgrading Cursor and hitting Cmd+Shift+P → Agents Window. The classic IDE view is still available.

Software development will continue to evolve as we enter the third era, where fleets of agents work autonomously to ship improvements.

— Cursor Blog

Gemma 4: Open Models, Actually Open

Google released Gemma 4 on April 2nd with four model sizes and, for the first time, an Apache 2.0 license.

The lineup:

• 2B & 4B - Edge models for phones and embedded devices
• 26B MoE - Mixture of Experts, good balance of capability and efficiency
• 31B Dense - Currently ranked 3rd among open models on Arena AI Text

The Apache 2.0 shift matters more than the benchmarks. Previous Gemma releases had restrictions that blocked some enterprise and commercial use. That’s gone now. The 31B model is genuinely competitive and you can ship it in a product without lawyering up.

Other highlights: 140+ language support, multimodal (text, image, audio), and the edge models are up to 4x faster and 60% less battery than previous generations. Available on Hugging Face, Kaggle, Ollama, and Google AI Studio.

Qwen 3.6 Plus: Free and Surprisingly Good

Alibaba dropped Qwen 3.6 Plus on OpenRouter around March 30th. No press release, no launch event. Just a free model that quietly showed up.

The specs: 1M context window, up to 65K output tokens, always-on chain-of-thought reasoning, native function calling. It beats Claude 4.5 Opus on some benchmarks (Terminal-Bench 2.0: 61.6 vs 59.3) and trails on others (SWE-bench Verified: 78.8 vs 80.9). Community reports suggest roughly 3x the speed of Opus 4.6.

I Tried It on My Own Codebase

I pointed Qwen 3.6 at this blog’s codebase with a generic “find usability issues” prompt. It came back with 19 issues.

After review:

• 5 false positives - Flagged things that weren’t actually problems
• 7 nits - Real but not worth fixing (style preferences, minor inconsistencies)
• 7 actionable issues - Genuine usability improvements worth making

I verified the actionable ones with Opus 4.6 afterwards. They held up. That’s a free model doing work that’s numerically comparable to Codex.

You can use it right now via OpenRouter with the model ID qwen/qwen3.6-plus-preview:free. If you use Claude Code, my claude-launcher package lets you point it at OpenRouter backends including Qwen.

The Connecting Thread

Every release is pushing the same direction: agents, not chat.

• Cursor rebuilt its entire UI around managing agent fleets
• Gemma wants agents running locally on your phone with Apache 2.0 freedom
• Qwen ships with native function calling and agentic coding as a headline feature

A year ago, “AI coding” meant autocomplete and chat sidebars. Now it means deploying autonomous agents across repos while a free Chinese model finds real bugs in your code. The gap between frontier and free is narrowing fast.

What Else Is Coming

April is stacking up. DeepSeek V4 is confirmed for this month (multimodal, text/image/video generation). The Claude Mythos leak from March 26th suggests Anthropic has something significant in the pipeline. Grok 4.20 is already out with a novel multi-agent architecture.

It’s going to be a busy month.

Update: One Week Later (April 11)

Eight days in and half of “What Else Is Coming” is already stale. Resolutions and corrections:

• Mythos landed as Project Glasswing on April 8. The tease at the end of this post resolved. Apple, Microsoft, Google, AWS and 40+ others are now deploying Claude Mythos Preview to find zero-days in critical infrastructure. $100M in credits, curated partner list, $25/$125 per Mtok, not publicly available. Mythos also escaped its sandbox during red-teaming and posted exploit details to public websites on its own initiative. This is the biggest AI story of the month.
• Claude Code leaked via npm the same day this post went up. Six internal packages accidentally published to the public registry, exposing the harness architecture. Full breakdown.
• Anthropic is under siege (April 8). Pentagon blacklist as a supply chain risk, source code leak, user revolt over Claude quality regressions, critical CVEs in their own tooling, CMS misconfiguration exposing 3,000 internal assets. Five-front pressure event in a single week.
• DeepSeek V4 is late. “Confirmed for this month” now looks optimistic. As of April 11 it’s still nowhere in sight, with the release window reportedly slipping to late April. The interesting angle isn’t the delay: V4 is targeting Huawei Ascend chips, which turns the launch into a test of whether Chinese silicon can power a real Nvidia competitor. Hong Kong Free Press and Japan Times are both running “waiting for DeepSeek” pieces.
• Grok 4.20 enterprise distribution expanded to Microsoft Foundry (April 8) and Palantir AIP (April 7). Pricing landed at $2/$6 per Mtok. The 4-agents-on-a-shared-backbone architecture from the Feb 17 beta is quietly becoming a front-line enterprise option.
• Qwen 3.6 Plus is still free. The expiry warning up the page was wrong, which is why I flipped it to an info note.

The gap between what’s “coming” and what’s already shipped keeps collapsing. This is the last time I pretend I know what next week looks like.