AI Daily · Sep 16, 2026: StepFun Releases StepAudio 3; Perplexity Open-Sources CobbleDB with 80% Latency Cut
Today in AI, StepFun launched the StepAudio 3 audio model family with API access, showcasing strong real-time speech capabilities; Perplexity open-sourced its CobbleDB database, slashing read latency and costs. Several free models and tools also debuted.
Pick a date
-
StepFun Releases StepAudio 3 Family with Five Audio Models
StepFun launched StepAudio 3, including real-time voice, ASR, TTS, audio generation, and music models. The real-time model scores 98.9% on Artificial Analysis conversational dynamics and 99.7% on speech reasoning; ASR achieves 1.7% WER.
-
Perplexity Open-Sources CobbleDB, Cutting Read Latency by 80%
Perplexity replaced DynamoDB with its custom key-value database CobbleDB, cutting median batch-read latency from 31.4ms to 5.60ms and p99 from 123ms to 24.2ms, with at least 20% cost savings. It will open-source the database, built in two months by two engineers and hundreds of AI agents.
-
OpenCode and OpenRouter Offer Free Access to Stealth Model Union Alpha
OpenCode offers free access to its stealth multimodal model Union Alpha for one week, with 256K context and tool calling; OpenRouter also lists it as free, with no training on user prompts.
-
StepFun Releases API for New StepAudio 3 Model Family
StepFun released the API for StepAudio 3, a family of five audio models covering real-time voice, speech recognition, generation, and music.
-
Tencent AI Open-Sources BrowserSkill for Browser Control
Tencent AI released BrowserSkill, a lightweight CLI bridge that lets AI agents use an existing browser tab, preserving login state and sessions.
-
GMI Cloud Launches MCP for 150+ Multimodal Models
GMI Cloud launched GMI MCP, letting agents in Claude, ChatGPT, Cursor, and Hermes Agent use 150+ multimodal models, including Seedance 2.5, GPT Image 2.5, and Gemini.
-
Qoder Releases Cloud Agents 1.0 for Production AI Agents
Qoder launched Cloud Agents 1.0, a managed cloud harness for building, shipping, and operating AI agents in production, with end-to-end integration in as little as 10 minutes.
-
OpenCode 2 Lets Users Assign Different Models to Subagents
OpenCode 2 now supports per-subagent model selection, enabling users to route tasks like code review to more powerful models.
-
Grok Beta Lets Users Edit Text Directly on Images
Grok has released a beta feature that allows text editing on any image, including invites and posters, avoiding the need to recreate visuals.
-
OpenRouter Makes US In-Region Routing Generally Available
OpenRouter's US In-Region Routing is now GA, processing requests entirely within the US from TLS termination to model provider, complementing EU routing. Flags are added to show geofencing provider support.