AI News: April 18, 2026

1. Alibaba’s Qwen3.6-35B-A3B Beats Gemma 4-31B on Agentic Coding Benchmarks

Alibaba. Qwen3.6-35B-A3B, the first open-weight variant of the Qwen3.6 generation, is a sparse MoE with 35B total and 3B active parameters released under Apache 2.0. It scores 73.4 on SWE-bench Verified versus Gemma 4-31B’s 52.0 and 51.5 on Terminal-Bench 2.0 versus Gemma 4-31B’s 42.9, while also edging out Gemma on GPQA (86.0 vs. 84.3) and AIME26 (92.7 vs. 89.2). Quantized, the model fits in 21GB and runs on consumer hardware, making it one of the most capable open-weight coding models you can self-host today. Source

2. Similarweb: ChatGPT Slips to 56.7% as Gemini Surges to 25.5% and Claude Triples to 6%

Similarweb. New Similarweb traffic data shows ChatGPT’s share of AI chatbot traffic falling from 77.4% a year ago to 56.7%, while Gemini jumped from 6% to 25.5% on aggressive Android integration and Claude went from 2.2% to 6.0% in a single month. Altman’s recent “ChatGPT has as many users in Texas as Claude has in the whole US” quip is aging quickly: Claude’s monthly growth is the fastest in the field, and Gemini’s inclusion in Android and Search notifications is reshaping the distribution game. Source

3. “Tokenmaxxing” Data Suggests AI-Heavy Dev Teams Ship More Code, Less Product

TechCrunch. A new piece synthesizes data from GitClear, Faros AI, and Jellyfish to argue that maximizing AI token budgets has become a status metric that obscures declining output quality. GitClear finds AI users have 9.4x higher code churn than non-AI peers, Faros reports an 861% increase in code churn under high AI adoption, and Jellyfish shows engineers with the largest token budgets achieving 2x throughput at 10x cost. Engineers initially accept 80-90% of AI suggestions, but real acceptance falls to 10-30% once revisions and reviews run their course. Source

4. Sam Altman’s World Pushes Orb Verification into Tinder, Zoom, DocuSign, and Ticketmaster

Tools for Humanity. Sam Altman’s World (formerly Worldcoin) announced a wave of new integrations for its Orb-based human verification, starting with a global Tinder rollout that lets users prove a profile belongs to a real person. Additional partners include Zoom for deepfake detection, DocuSign for signature verification, and Ticketmaster, Eventbrite, and Concert Kit for ticket sales. The scale-out is a direct response to the rising agent-versus-human identity problem facing consumer platforms. Source

5. Cal.com Closes Source Code, Citing Security — Alpha Signal Pushes Back on the “5-10x Easier to Exploit” Claim

Cal.com. Cal.com moved its 30,000-star codebase behind a closed license, citing vulnerabilities identified by Gecko Security and claiming open source makes the app “5-10x easier to exploit.” Alpha Signal’s analysis argues the actual security evidence is real but that the 5-10x statistic is unsubstantiated, and that the shift is better explained as a commercial decision framed as a safety one. The episode is a preview of how AI-enabled exploit discovery will put pressure on open-source maintainers to justify openness. Source