Official news from the last 30 days, ordered by announcement date. Preview, general release, hosting expansion and product updates are labelled separately. Same-date items use provider/title ordering.
Mistral released improved Le Chonk checkpoints after the October preview. The modified-MIT license includes material commercial conditions, and the weights target substantial GPU infrastructure.
What to watch: An API preview and downloadable weights are different access routes. Read the exact license and runtime requirements before choosing self-hosting.
Runway’s October 27 announcement connects ad-performance inputs to branded creative variants and publishing workflows. The feature is described for paid plans.
What to watch: A more connected creative loop may reduce handoffs. No return-on-investment claim has been independently verified.
OpenAI introduced GPT-6.1 Terra for visual understanding, computer use and writing. Its positioning distinguishes these tasks from the hardest coding workloads assigned to Sol and Astra.
What to watch: Evaluate it on the visual or writing task you actually need, and tightly bound any computer-use permissions.
The Decisions API is generally available for bounded classification and scoring workflows. It is a distinct interface from general conversational chat.
What to watch: A focused decision interface can fit narrow tasks, but calibration and correctness still require task-specific evaluation.
GLM 5.3 Flash became available on Mistral. This is a new hosting option, not the original checkpoint launch.
What to watch: Check the host’s access and pricing separately from the original MIT checkpoint.
Anthropic announced the donation of Agent Skills to the Agentic AI Foundation, supporting portable instructions across agent systems.
What to watch: Portability can reduce duplicated instructions. It does not guarantee identical behavior or security across runtimes.
Anthropic announced a proactive product powered by Opus 5.5, with briefs, connected-app work and its own cloud computer. The research preview starts through a US Pro/Max waitlist.
What to watch: Background work and monitoring share plan usage. Access arrives in waves; no 27 October rollout change was verified.
Source update: 2026-10-23 · announcement date retained above.The previews cover text-to-speech, transcription, real-time transcription and a larger transcription model. The Voice Agents SDK is a separate product.
What to watch: Match the specific speech task and preview access instead of treating the family as one interchangeable model.
Google announced additional Gemini tools for small-business workflows through a staged rollout.
What to watch: Check your account, region and rollout status before relying on an announced feature.
The Agent SDK supports typed tools, skills, MCP, local evaluations and deployment to Cursor or your own infrastructure.
What to watch: It gives developers more explicit building blocks for agents. Evaluate behavior locally and keep credentials narrowly scoped.
New controls let users shape slide and video outlines and producer notes before rendering source-derived outputs.
What to watch: Reviewing the plan earlier can catch missing context before generating a polished artifact.
The October 17 version updated the API alias and web/app model. The changelog says the calling method and pricing are unchanged.
What to watch: Use the current changelog: it reverses an earlier V4-Pro retirement plan, so an old shutdown notice is no longer a safe assumption.
Cursor announced cloud-agent canvases and environment controls. The environments API is beta.
What to watch: Environment setup is part of agent reliability. Beta APIs need migration and failure handling before becoming a critical dependency.
Timestral is a four-size forecasting family with CPU support and modified MIT terms. It is not a chatbot LLM.
What to watch: A forecasting model belongs in a different evaluation workflow from conversational assistants.
The speech-to-speech API moved grok-voice-latest to Voice 3.0 automatically. Comparisons in the announcement are not independently tested by PrisAIm.
What to watch: A moving latest alias can change application behavior; validate the model version and your own audio workflow.
Google announced broader Gemini 4 Argon access for complex coding, professional and scientific workflows. Availability and introductory API terms vary by product.
What to watch: An access announcement is not uniform access across every plan. Check the specific product and pricing terms.
Wonder is offered through the Meta Model API and Muse Code; Full is for Muse Code subscribers. Consumer Muse rollout is forthcoming.
What to watch: These are Muse models, not Llama. Public weights were not verified in this source check.
Source update: 2026-10-13 · announcement date retained above.The image-model update is available through checkpoint and API routes. Its default checkpoint license is research-only, with separate commercial licensing.
What to watch: Do not treat model availability as unrestricted commercial permission.
Anthropic positioned Haiku 5.5 for high-volume tasks. The announcement’s performance and efficiency claims are provider results.
What to watch: High-volume use magnifies small errors. Evaluate representative tasks, quality thresholds and cost together.
The October 6 announcement introduced a public preview and promised downloadable weights. Improved open weights were subsequently confirmed on October 29; see the separate release update.
What to watch: The preview date and the later checkpoint release are different events. Current API access, license terms and hosting requirements are explained in the model profile.
The image generation and editing model is generally available. It is distinct from the text-output Gemini 4 Argon model.
What to watch: Choose the model by its documented output modality, not only the Gemini family name.
Sonnet 5.5 is a commercial cloud model for coding and knowledge work, with documented text/image input and text output. It is distinct from Opus, Haiku and the Duet product.
What to watch: Check migration details, adaptive thinking and current pricing against your workload. Provider capability statements are not independent PrisAIm benchmarks.