The week enterprise AI got a production control plane
OpenAI on AWS, FedRAMP Moderate and Mistral Workflows, 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 models including GPT-5.5, Codex, and Amazon Bedrock Managed Agents powered by OpenAI all launched in limited preview on AWS. Codex usage can apply toward AWS cloud commitments, and all customer data is processed by Amazon Bedrock.
THE CATCHThe three pieces are separate launches: models on Bedrock, Codex configured to use Bedrock as provider, and Bedrock Managed Agents. Codex on Bedrock starts with Codex CLI, the desktop app, and the VS Code extension.
IF YOU RUNan AWS shop with existing security, identity, and procurement workflows and want OpenAI models without leaving that perimeter
SOURCEOpenAI, OpenAI models, Codex, and Managed Agents come to AWS
OpenAI
FedRAMP Moderate
DEPLOYif you build for US federal agencies
WHAT SHIPPEDChatGPT Enterprise and API Platform achieved FedRAMP 20x Moderate authorization. Agencies can access GPT-5.5 in the FedRAMP environment, and Codex Cloud access via FedRAMP ChatGPT Enterprise workspace is coming soon.
THE CATCHThe authorization covers ChatGPT Enterprise and API Platform only. Codex Cloud access is described as something agencies can soon access, not something available today.
IF YOU RUNa public sector team that needs reusable FedRAMP authorization data instead of starting a security review from scratch
SOURCEOpenAI, OpenAI available at FedRAMP Moderate
Mistral
Mistral Workflows
DEPLOYif you run multi-step AI processes in production
WHAT SHIPPEDWorkflows is released in public preview as the orchestration layer for enterprise AI, built on Temporal's durable execution engine. It includes durable execution, observability in Studio, human-in-the-loop approval via a single wait_for_input() call, and deployment where workers run in your environment.
THE CATCHThe control plane runs on Mistral while workers and data processing run in your Kubernetes environment via a Helm chart. The Python SDK v3.0 installs with uv add mistralai-workflows.
IF YOU RUNnotebook pipelines that fail silently in production or long-running processes that cannot survive a network timeout
SOURCEMistral, Workflows for work that runs the business
The Deploy Log
Models and APIs
| WATCHif you process long documents, video, or audio Nemotron 3 Nano Omni. NVIDIA released Nemotron 3 Nano Omni, a 30B-A3B omni-modal model for documents, audio, video, and agentic computer use, with OCRBenchV2-En 65.8, MMLongBench-Doc 57.5, Video-MME 72.2, and VoiceBench 89.4, and checkpoints in BF16, FP8, and NVFP4 on Hugging Face. Hugging Face |
Tools and products
| WATCHif you run Hobby projects with long-lived deployments Hobby retention cap. Starting April 29th, the maximum retention policy for Vercel Hobby plans is capped at 30 days, with deployments outside the retention window automatically removed except the 10 most recent production deployments and any aliased deployments. Vercel |
Left out
- Sandbox Postgres connections: the publisher's page does not show it available yet
- Grok 4.3 on AI Gateway: the publisher's page does not show it available yet
- Granite 4.1: the condition it attaches to the call did not hold up
- Symphony spec: the condition it attaches to the call did not hold up
- Rollup number formatting: it did not carry a number that moved
- Vercel Pro via Stripe Projects: it did not carry a number that moved
- Sandbox custom tags: it did not carry a number that moved
- Native Deployment Checks: it did not carry a number that moved
- DeepInfra Inference Provider: the condition it attaches to the call did not hold up
- Custom Agents in private Slack: the condition it attaches to the call did not hold up
The pattern
The week's strongest items are not new models but new control planes: OpenAI models inside AWS, OpenAI products inside FedRAMP, and Mistral Workflows as an orchestration layer. Each one answers the same enterprise question, which is not whether the model is good but whether it can run inside the perimeter you already operate.
The second thread is agent orchestration moving from session management to task management. Symphony turns an issue tracker into a supervisor, Mistral Workflows turns business processes into durable state machines, and Vercel's deployment checks and sandbox tags turn infrastructure into something agents can drive. The deployment decision this week is less about which model and more about which control plane you trust to keep agents running when you stop watching.