Daily AI CatchupNew wave of fast models: Mercury 3, MiniMax, Exa 3
Inception Labs released Mercury 3, a diffusion-based language model that reportedly beats GPT-6 Luna and Claude Haiku 5.5 on some benchmarks while generating over 1,000 tokens per second. MiniMax released M3.1 Flash, aimed at agentic tasks and coding, claiming nearly 3x faster inference and 70% lower cost than its predecessor; weights are promised later. Exa shipped version 3.0 of its AI search engine, including a cheaper 'Exa Lite' tier, built on an index of over 2 trillion URLs.
Why it matters
💻 Developer · Mercury 3's diffusion architecture and MiniMax's cost cuts matter if you're picking models for latency-sensitive agent pipelines. Exa 3 gives you a cheaper search API option for RAG systems.
📦 Product · Cheaper, faster inference options mean more headroom to add AI features without blowing your cost budget per request.
🎨 Design · Faster token generation means snappier AI-assisted interfaces, less spinner time for users waiting on model output.
📈 Business · Falling per-token costs across multiple vendors keep squeezing margins for anyone reselling AI capability, good news if you're a buyer.
🤔 Just Curious · AI models are getting faster and cheaper at a steady clip, this week it's three different companies racing on speed instead of just raw intelligence.
Sources: Mercury 3, MiniMax M3.1 Flash: Redefining Agent Performance and Cost, Introducing Exa 3.0