Frontier models move into enterprise rails; smaller models take the edge
OpenAI on AWS, ChatGPT Dreaming memory and GPT-Rosalind, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
OpenAI
OpenAI on AWS
DEPLOYif you run workloads on AWS
WHAT SHIPPEDOpenAI frontier models and Codex are generally available on AWS through Amazon Bedrock, in both Commercial and GovCloud regions. Codex is used by more than 5 million people every week.
THE CATCHThe AWS path removes procurement, security review, and production readiness friction, but Daybreak cyber models and Codex Security are not yet available and remain a future expansion.
IF YOU RUNAWS with existing security, compliance, procurement, billing, and governance workflows
SOURCEOpenAI, OpenAI frontier models and Codex are now available on AWS
OpenAI
ChatGPT Dreaming memory
DEPLOYif you pay for Plus or Pro
WHAT SHIPPEDOpenAI began rolling out a new dreaming-based memory synthesis system for ChatGPT, available to Plus and Pro users in the US today, with additional countries and Free and Go users over the coming weeks.
THE CATCHThe new system is built to fix staleness, correctness, and scalability across hundreds of millions of users and multi-year time horizons, and memories are reviewable through a memory summary page.
IF YOU RUNlong-running ChatGPT projects where context carries across chats
SOURCEOpenAI, Dreaming: Better memory for a more helpful ChatGPT
OpenAI
GPT-Rosalind
DEPLOYif you work in life sciences research
WHAT SHIPPEDGPT-Rosalind came out of research preview and is now available globally to eligible organizations through OpenAI's trusted-access program, with published pricing taking effect on October 5, 2026.
THE CATCHThe model scores 27.5% on MedChemBench versus GPT-5.5 at 25.1% while using 7.2% fewer tokens, and 63.2% on LabWorkBench versus 55.8%, but access is gated to organizations with strong governance and enterprise-grade security.
IF YOU RUNdrug discovery, genomics, or wet lab workflows needing sourced evidence and bioinformatics execution
SOURCEOpenAI, Introducing new capabilities to GPT-Rosalind
The Deploy Log
Models and APIs
| DEPLOYif you run high-volume customer claims operations Travelers AI Claim Assistant. Travelers expanded its OpenAI Realtime API voice assistant countrywide within two months, with 85 to 90% of customers using the AI Assistant now completing their claim filing through AI. OpenAI |
| DEPLOYif you run multilingual content moderation pipelines Nemotron 3.5 Content Safety. NVIDIA released Nemotron 3.5 Content Safety, a 4B-parameter model that unifies multimodal input, 12-language explicit coverage, custom policy enforcement, and auditable reasoning traces in one inference call. Hugging Face |
| DEPLOYif you run high-throughput routing or RAG workloads Mellum2. JetBrains released Mellum2, a 12B-parameter Mixture-of-Experts model that activates only 2.5B parameters per token and delivers more than 2x faster inference than similar-sized models under the Apache 2.0 license. Hugging Face |
Tools and products
| WATCHif you generate video with synchronized audio Grok Imagine Video 1.5. Grok Imagine Video 1.5 from xAI is now available on AI Gateway, generating video from an input image with synchronized audio in a single pass and stronger face accuracy and character consistency. Vercel |
| DEPLOYif you build commerce storefronts Shopify on Vercel. You can now create a Shopify store directly from Vercel, install the Shopify integration to configure credentials, build with v0, and deploy a production-ready storefront in minutes. Vercel |
| DEPLOYif you serve private files from Blob storage Vercel Blob signed URLs. Vercel Blob now supports time-bound signed URLs scoped to a single operation, pathname, and expiry up to 7 days, with upload URLs supporting multipart for direct browser streaming. Vercel |
| WATCHif you search open-source skills programmatically skills.sh API. The skills.sh API is now available, letting you authenticate with your project's Vercel OIDC token and query more than 600,000 skills from the open-source ecosystem at a rate limit of 600 requests per minute. Vercel |
Left out
- Holo3.1: the publisher's page does not show it available yet
- Nemotron 3 Ultra on AI Gateway: the publisher's page does not show it available yet
- Qwen 3.7 Plus: the publisher's page does not show it available yet
- MiniMax M3: the publisher's page does not show it available yet
- Chat SDK Twilio adapter: the publisher's page does not show it available yet
- Chat SDK Velt adapter: the publisher's page does not show it available yet
- Chat SDK AgentPhone adapter: the publisher's page does not show it available yet
- Chat SDK Lark Feishu adapter: the publisher's page does not show it available yet
- Elastic build machines OOM protection: a figure in the draft could not be found on the publisher's own page
- Monorepo Git settings: it did not carry a number that moved
- Vercel Sandbox drives: it did not carry a number that moved
- Chat SDK Slack primitives: it did not carry a number that moved
- Chat SDK Vue and Svelte: the condition it attaches to the call did not hold up
- Vercel CLI Session Traces: it did not carry a number that moved
- Vercel Blob OIDC: it did not carry a number that moved
The pattern
OpenAI is moving its frontier models into enterprise procurement rails while simultaneously pushing smaller, specialized models to the edge. The AWS Bedrock launch puts OpenAI frontier models and Codex inside existing Commercial and GovCloud security and billing workflows, and Travelers shows the operational result: 85 to 90% of customers complete claim filing through the Realtime API voice assistant. Yet the same week, NVIDIA ships Nemotron 3.5 Content Safety as a 4B-parameter model and JetBrains ships Mellum2 as a 12B Mixture-of-Experts model activating 2.5B parameters per token under Apache 2.0. These are not frontier replacements; they are edge components that enforce policy, audit reasoning traces, and run 2x faster inference without a procurement event.
The infrastructure layer is converging on signed, scoped, and rate-limited access. Vercel Blob signed URLs bind an operation, pathname, and expiry up to 7 days, while the skills.sh API authenticates with a project Vercel OIDC token and serves 600,000 skills at 600 requests per minute. GPT-Rosalind exits research preview with published pricing on October 5, 2026 and gated trusted access, scoring 27.5% on MedChemBench versus GPT-5.5 at 25.1% while using 7.2% fewer tokens. ChatGPT Dreaming memory rolls out to Plus and Pro users in the US to fix staleness and correctness across hundreds of millions of users. Grok Imagine Video 1.5 generates video with synchronized audio in a single pass on AI Gateway, and a Shopify store can be created directly from Vercel and deployed in minutes. The pattern is not one model winning; it is every model becoming a component behind a signed URL, a rate limit, a governance gate, or a memory summary page.