Mistral AI on October 6 opened public API access to Mistral Large 4, a natively multimodal mixture-of-experts model with one trillion total parameters and 49 billion active at inference (AI News). The model accepts a one-million-token context window and was trained over two months on 4,000 Nvidia Grace Blackwell GPUs at Mistral’s European facilities (RuntimeWire). Mistral reports Large 4 reaches competitive performance against some closed frontier models on cyber, manufacturing, and finance tasks, while acknowledging it still trails on coding benchmarks. The company is staging the release through a moderated API preview before releasing weights, a pattern designed to assess real-world use before open distribution (RuntimeWire). The model is accessible now via Mistral Studio; open weights and a technical report are scheduled for October 31 under Apache 2.0, with the Hugging Face repository already staged as an upcoming release (Hugging Face).
One day earlier, Reflection AI introduced Beam, its first open-weight model: a 501-billion-parameter sparse mixture-of-experts architecture with 23 billion parameters active per token, targeting coding and agentic workloads with a 256,000-token context window and 128,000-token maximum output (Unite.ai). An OpenAI-compatible beta API is live at api.reflection.ai, currently behind a waitlist, while full weights, a technical report and a model card are expected later in October also under Apache 2.0 (OrcaRouter). Beam and Large 4 both arrive first as gated API previews, with weights to follow - a pattern that has become common for models at the largest parameter counts.
At OpenAI DevDay on September 29, the company detailed developer-facing additions now in preview: the Agents API gained support for computer use, providing applications a managed execution environment where agents operate graphical software interfaces to complete tasks (OpenAI). The same release added multi-agent coordination from Codex, tool search and context compaction, with an AWS Bedrock integration that lets teams run OpenAI agents natively within their AWS infrastructure (InfoQ). A separate announcement added support for the proposed MCP Events specification, enabling plugins to trigger agent workflows when an external application event fires - for example, a new task on a project board; the feature requires MCP protocol version 2026-07-28 and the specification remains under active development (OpenAI Developers).