Follow the AI industry.

The downloadable coding model supports local customization, but its hardware needs and commercial license matter as much as its benchmark scores.

The new models target different deployment sizes: multimodal enterprise agents and local assistants, with a smaller 4GB edition scheduled for October 23.

The model cutoff took effect October 14 for developers signed in through ChatGPT accounts, while direct API-key connections remain active. Teams must manually update workspace defaults, command-line flags, and scheduled scripts to supported replacement models.

xAI has updated its Voice Agent API to Grok Voice 3.0, automatically moving the default alias to the new speech-to-speech generation model while holding pricing at $0.08 per minute.

The API is generally available with introductory rates through November 30. Pro subscribers gain access in the Gemini app and AI Mode, while Antigravity access requires Ultra or Enterprise.

Character Technologies Inc. is keeping its platform online while seeking court approval for a sale, but says the platform will wind down if the sale does not provide for continued operations.

The standard Muse Wonder model enters the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers ahead of planned consumer rollouts.

The maximum remains 1 million tokens. Sottiaux says subscription usage will carry no long-context fees.

Agents can write programs that divide large tasks into phases and combine results, with spending and completion conditions developers need to handle.

Already-claimed offers remain active while Anthropic re-reviews applications; existing members keep other program benefits.

The company’s October 9 report describes unintended interactions with live government and other websites. Its broader internet restriction will remain until security and monitoring measures reliably catch such behavior.

The current help page says deeper prompts may reach Flash or Pro; most free-account prompts go to Flash-Lite.

The enterprise agent combines model choice, shared skills and Workspace access. Google describes controls for permissions, network traffic and the token and sandbox costs of delegated work.

The lightweight model introduces adaptive thinking and lower starting rates, but requests with prompts beyond 100,000 tokens jump to a fivefold pricing bracket.

Answers can include interactive tools, diagrams and comparisons. The rollout starts with paid plans, with specific app and thinking-level restrictions.

Defense Access opens applications to eligible individual researchers as well as organizations, while higher tiers require stronger vetting. Data retention and cloud-provider restrictions remain material conditions.

Mistral AI has launched an API preview of its natively multimodal mixture-of-experts model, offering a 1-million-token context window alongside promotional inference pricing. Standalone weights for self-hosting remain unreleased, with the French AI lab scheduling a public download for late October.

OpenAI has launched an invisible text-watermarking system called `textGrain`, making it available as an opt-in feature for API customers and preparing a phased rollout for eligible ChatGPT and Codex outputs in the European Union. Detection access remains restricted to approved researchers, and vendor evaluations show that detection reliability degrades significantly when text is edited or lacks sufficient length.

Reflection has unveiled Beam, a sparse mixture-of-experts model offering 23 billion active parameters and a planned open-weights release, alongside estimated inference-compute claims that exclude real-world serving overhead.

The nomination-only program pairs a four-day intensive with 12 weeks of deployment support, targeting 10,000 Frontier Deployed Engineers across enterprise partners by the end of 2027.