Releases

OpenAI says billing for the trusted-access GPT-Rosalind research model begins on October 5, 2026, at $5 per million input tokens, $0.50 per million cached input tokens, and $25 per million output tokens.

CoreWeave launched reactive Python and SQL notebooks powered by marimo and connected to Forge experiment, artifact, and evaluation data.

ARIA became generally available in CoreWeave Forge with experiment analysis, code-aware hypothesis generation, and approval-gated run launches.

CoreWeave opened ARENA Pass to qualified enterprise teams seeking workload-specific production evidence and deployment recommendations.

CoreWeave said Cognition was running production workloads on its NVIDIA Vera Rubin NVL72 system shortly after deployment.

CoreWeave announced it would offer NVIDIA Vera, a CPU designed for agentic and data-intensive AI workloads.

CoreWeave introduced a curated network of production-validated partner integrations across the AI stack.

CoreWeave launched Forge as a connected development environment spanning model and agent production, observation, curation, training, evaluation, and continuous improvement.

Perplexity launched scheduled and event-triggered automations that build on prior Computer work.

SpaceXAI launched shareable Grok Bots that work from common team files, apps and expertise and learn as teams use them.

Microsoft makes Fabric IQ grounding generally available in Copilot Chat and Cowork, connecting governed Power BI business context to AI workflows.

NVIDIA launched an open full-stack agent safety platform combining OpenShell secure runtime and the Sentry reference design.

Cohere opened a managed Compass Cloud private beta for retrieval-heavy and agentic applications.

Microsoft launches Copilot Managed Runtime in preview as governed hosting infrastructure for Copilot-created apps and automations.

Microsoft introduces Cowork for delegating complex end-to-end work and returning completed deliverables for iteration.

Microsoft introduces Autopilot, formerly Scout, as a persistent agent that proactively performs recurring and long-running work.

Microsoft introduces Home as Copilot's unified starting experience combining Chat, Cowork and editable Office documents.

Microsoft introduces Copilot Code for creating sandboxed applications, dashboards, automations and workflows through natural language.

The Mission Control Agent entered console preview with deployment diagnostics and recommendations based on platform telemetry.

Transformers added support for running llama.cpp GGUF quantized models within the library’s model-definition and inference workflows.

Perplexity launched a marketplace for reusable Computer skills with enterprise governance.

Perplexity launched local-first Portable Computer for supported NVIDIA Windows and Linux systems.

Perplexity introduced configurable Light, Standard, High and Ultra effort levels for Computer tasks.

Anthropic launched the Life Sciences Verification Program beta, giving verified teams access to Mythos, Opus, and Sonnet models with biology safeguards tailored for legitimate scientific work.

CoreWeave added cross-region write acceleration and an Archive storage class, while LOTA served cached reads at up to 7 GB/s per GPU.

CoreWeave announced a multi-rack Vera Rubin NVL72 cluster and supporting compute, networking, storage, and orchestration capabilities.

SpaceXAI added persistent memory capabilities to Grok Build.

Modal added sidecar processes to sandbox environments.

DeepSeek reduced API prices following the V4.1 Flash release.

CoreWeave launched an embedded field-engineering offering that helps customers turn proprietary physical-world data into production AI systems.

AWS made OpenAI's GPT-6 Astra generally available through Amazon Bedrock.

SpaceXAI launched enterprise deployment and organization-wide access for Grok Bot.

Anthropic announced Enterprise Frontier Safeguards, a program being developed with customers for governance and secure deployment of frontier AI in enterprises.

AWS expanded GovCloud Bedrock access to Amazon Nova, Anthropic Claude, Meta Llama, NVIDIA Nemotron, OpenAI and xAI Grok model families.

SpaceXAI added tighter X integration to Grok Bot.

Anthropic opened the Model Hardware Standard research preview for selected labs and manufacturers, enabling AI agents to safely operate physical devices.

SpaceXAI expanded Grok Bot access across SuperGrok, Cursor Pro and Cursor Teams plans.

NVIDIA announced an entry-level robotics computer with higher local model inference performance and lower power use.

NVIDIA put its Groq 3 LPX interactive inference accelerator into full production for ultrafast token generation.

OpenAI reduced GPT-5.6 Sol API and credit pricing by more than 20% for three months, following earlier reductions for Terra and Luna.

Mistral AI launched Agentic Search for autonomous research, retrieval and evidence-grounded agent workflows.

SpaceXAI made Grok Build available across plans on the web and mobile.

Mistral AI expanded in-region inference for open models using new European AI infrastructure.

SpaceXAI launched persistent AI teammates with their own cloud computers, connectors, memory, approvals and routines.

AWS added restricted access to OpenAI Daybreak Red and Daybreak Blue cybersecurity models.

OpenAI expanded Daybreak as a cyber-defense program and access platform for defensive security work.

Perplexity launched builder-focused Computer workflows spanning coding, deployment, monitoring and business operations.

Perplexity released Numbat, an open-source monitoring and response layer for AI coding agents.

Perplexity released Personal Computer for Windows with local files, background sessions and desktop workflows.

Mistral AI launched Shieldstral, a security and guardrail layer for AI applications and agents.

NVIDIA released Alpamayo 2 Super under an open commercial license for autonomous-vehicle development.

Perplexity evolved Spaces into Projects with persistent files, memory, connectors and shared agent workspaces.

Perplexity brought Model Council into Computer for multi-model analysis and artifact generation.

SpaceXAI launched an early beta mode for creating websites, apps, games and interactive dashboards inside Grok.

Cohere launched governed workflow orchestration in North with branching, approvals, schedules and model routing.

SpaceXAI launched a Grok add-on for Google Sheets, Slides and Docs.

Grok Build gained workflow orchestration for fanning tasks across many agents, verifying results and reporting in the background.

SpaceXAI launched a Microsoft Outlook add-in for thread summaries and drafting replies.

SpaceXAI launched a Microsoft Excel add-in for natural-language analysis, formulas and scenarios.

SpaceXAI added scheduled and trigger-based automations that run jobs and report results.

Mistral AI Studio added centralized prompt and skills management for enterprise AI systems.

AWS made OpenAI GPT-5.6 Sol, Terra and Luna generally available through Amazon Bedrock.

OpenAI launched ChatGPT Work, a workspace for longer, multi-step professional tasks powered by GPT-5.6.

SpaceXAI added 21 multilingual voices and improved the naturalness of its original voice set.

SpaceXAI launched a no-code builder for creating personalized voice agents.

Apollo Go launched fully driverless commercial operations in Dubai, including availability through Uber.

Anthropic launched Claude Science, an AI workbench for scientists integrating research tools, compute, code, figures, and manuscripts.

CoreWeave made W&B Weave’s agent-development capabilities generally available alongside the ARIA preview.

CoreWeave launched ARIA in preview as an AI research and iteration agent inside Weights & Biases that analyzes experiments and drives follow-on runs.

HF Jobs added exposed-port workflows for running private services such as OpenAI-compatible vLLM servers directly on Hugging Face compute.

Baidu moved ERNIE web services to an upgraded site with chat, creation and web-search capabilities.

Modal documented state-of-the-art speculative decoding performance in Auto Endpoints.

Perplexity launched legal research, analysis and drafting workflows for counsel.

Mistral AI added centralized enterprise controls for connectors, access and governance.

Modal introduced dedicated single-tenant endpoints for production model serving.

Modal introduced shared managed endpoints as part of its inference product line.

Modal introduced long-running high-performance server deployments.

Modal introduced automatically optimized and managed model endpoints.

Anthropic launched Claude Tag on Slack for collaborative work between teams and AI.

SpaceXAI added /goal for long-running autonomous task execution in Grok Build.

Perplexity introduced self-improving memory for agents and persistent Computer workspaces.

SpaceXAI launched a Microsoft Word add-in for drafting, formatting and research inside documents.

AWS introduced an AI-native security service for discovering, validating, prioritizing and remediating code vulnerabilities.

AWS introduced a governed knowledge graph for giving agents shared context across enterprise data.

AWS added continuously running autonomous agents, an activity feed and cross-source enterprise insight retrieval to Amazon Quick.

AWS launched a managed web-search tool that grounds agents in current cited information without customer-data egress.

AWS introduced fully managed enterprise retrieval capabilities that abstract RAG ingestion, indexing and retrieval infrastructure.

Databricks announced Unity AI Gateway as a unified governance and control plane across hosted and external AI assets.

Databricks expanded Agent Bricks with model choice, memory, document intelligence, sandboxes, registry support and multi-agent development capabilities.

SpaceXAI launched a Microsoft PowerPoint add-in for creating and refining presentations with Grok.

SpaceXAI added a dashboard for managing many concurrent coding sessions and dispatching work.

Perplexity integrated Deep Research into Computer so research could flow directly into documents, spreadsheets, presentations and workflows.

SpaceXAI added a built-in plugin marketplace to Grok Build.

IBM released modular adapter functions for RAG, safety and core model capabilities.

IBM positioned its open Mellea generative-programming library as the runtime for modular Granite adapters.

Perplexity launched a set of Computer workflows for growing small and midsize businesses.

Meta introduced an AI business agent that can learn from a company's public content and support customer conversations across Meta surfaces.

Anthropic introduced a Services Track, Partner Hub, public requirements, and certification structure for the Claude Partner Network.

Perplexity introduced local and hybrid execution that moves selected Computer workloads onto users' machines.

Microsoft AI introduces Frontier Tuning, using reinforcement-learning environments to adapt MAI models to organization-specific workflows.