No results found
We research, test, and compare products so you don't have to. Find the best tools for every category.
Updated April 18, 2026 (5:00 AM PST)
xAI quietly drops Grok 4.3 with native video input, document generation, and a 2M token context window — but only for SuperGrok Heavy subscribers at $300/month. No model card, no benchmarks, no announcement.
Anthropic's Opus 4.7 jumps to 87.6% on SWE-bench Verified, adds 3.3x vision resolution and a new xhigh effort level — but an updated tokenizer means the same input now costs up to 35% more.
Huawei's new AI chip delivers 2.8x the FP4 performance of Nvidia's H20 with 112GB of in-house HBM. ByteDance has already committed $5.6 billion in orders.
An anonymous 15B-parameter video model appeared on Artificial Analysis, took #1 in text-to-video with 1389 Elo, and then turned out to be Alibaba's. HappyHorse-1.0 generates video and audio in a single pass.
Meta Superintelligence Labs ships its first model — a small, fast, closed-source reasoner that's 10x more compute-efficient than Llama 4 Maverick. Three days after releasing open-weight Llama 4, Meta went proprietary.
Llama 4 Scout runs 109B total parameters across 16 experts with a 10 million token context window. Maverick scales to 400B across 128 experts. Both are open-weight, natively multimodal, and Meta's first MoE release.
Microsoft's first in-house speech models: MAI-Transcribe-1 beats Whisper on all 25 languages at half the GPU cost, and MAI-Voice-1 generates 60 seconds of expressive audio in under one second.
Four model sizes, Apache 2.0, natively multimodal from 2B to 31B — and the 26B MoE variant scores 88.3% on AIME 2026 with only 3.8B active parameters.
IBM watsonx Orchestrate integrates ElevenLabs TTS and STT, bringing 10,000+ voices across 70 languages to enterprise AI agents — with PCI compliance and HIPAA-ready data handling built in.