AI developer tools · What shipped

Epoch AI released updated benchmark results showing GPT-6 Astra setting new records on their Epoch Capabilities Index and multiple benchmarks including mathematics, continual learning, and game-puzzle tasks. The organization also launched FrontierMath Erdős, featuring 68 unsolved mathematical problems where AI systems write solutions in Lean; GPT-6 Astra solved 2 of 68 problems (3%), with no prior models solving any. Their benchmark hub tracks 390 models across 80 benchmarks covering mathematics, coding, and other domains, using the Inspect framework for evaluations with published task definitions and accessible log viewers showing detailed model outputs and scoring methodology.

Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, voice-first AI models designed for developers building production voice agents. Gemini 3.8 Live Extended Thinking achieved top rankings on Artificial Analysis' Speech to Speech Quality Index (82.6) and τ-Voice benchmarks (68.6%), with 97.7% on Big Bench Audio, while Gemini 3.8 Live secured second place in the Speech Agent Arena. Both models are available now via the Gemini API for developers, with support for near real-time visual processing, 97 languages, and background tool execution. Developer platforms including LangChain, Pipecat, Vercel, and Vision Agents are integrating the Gemini Live API to enable developers to build high-performance voice interfaces.

Google released Gemini 3.8 Live and 3.8 Live Extended Thinking models for real-time voice applications via the Gemini API and Google AI Studio. Gemini 3.8 Live ranks #1 on Artificial Analysis' Speech-to-Speech leaderboard and supports asynchronous function calling, visual context grounding, multilingual support for 97+ languages, and configurable thinking for complex reasoning. The models are priced at $0.005/min for audio input and $0.018/min for audio output. Google also released Gemini 3.5 Transcribe, a speech-to-text model achieving 4.0% Word Error Rate (streaming) and 2.6% (non-streaming) across 85+ languages, with features including automatic code-switching and custom vocabulary biasing. Both models are available through integration partners including LangChain, Pipecat, Vercel, and Vision Agents.

AutoGPT Platform released several updates across versions 0.6.67 through 0.7.4. Recent releases (v0.7.4 in September and v0.7.3 in August) introduced features for managing AI expert agents within workflows, including scope integrations per expert, activity event logging, and user-facing memory settings. Earlier versions (v0.6.67–v0.7.2) added Slack and Telegram bot adapters, multi-workspace support, expert scheduling with attribution and triggers, and expert memory isolation. The platform also expanded model support with Claude Sonnet 5 and added provider blocks for services like Tavily (search, extract, crawl). UI/UX improvements include a new copilot chat composer, artifact panel resizing, and a redesigned sidebar with activity tracking. These updates focus on multi-agent orchestration, improved developer experience through better workflow composition, and tighter integration with third-party services via adapter blocks and API enhancements.