In this briefing
  1. 01Salesforce's roadmap proposes bringing identity, behaviour, outcomes and cost into the control plane
  2. 02Cost records enter an API, but their granularity still has boundaries
  3. 03Training data paths begin to use short-lived workload identities
  4. 04Reproducible evaluation and physical lifecycle add Day-2 evidence
  5. What to watch next
  6. Sources and verification
Key points
  1. Salesforce placed AI discovery and registration, identity, policy, lifecycle, evaluation, observation of behaviour and outcomes, and cost control in a new AI Control Plane. The availability of existing foundation technologies does not mean that the whole product is available; the new capabilities and unified experience are only planned to begin rolling out from the start of Salesforce FY28.
  2. Baseten's two updates separately cover daily cost subtotals for Model APIs, and OIDC or AWS AssumeRole for pulling private images, weights and data during the training job setup stage. They cannot be extrapolated into a per-request cost ledger or a claim that all runtime access is free of long-lived credentials.
  3. Claude Code provides plugin evals that are scored and reproducible and output JSON and HTML, while Dell Private Cloud manages the physical lifecycle across four classes of virtualisation platform. The former is a vendor tool rather than an independent standard; the latter's deployment-efficiency figures come from Dell internal testing or commissioned research.
Signal 01

Salesforce's roadmap proposes bringing identity, behaviour, outcomes and cost into the control plane

At 11:09:31 on 11 September (Beijing time), Salesforce announced the Trusted Enterprise AI Harness, listing context, agency, action, governance, security and models as six categories of capability. The accompanying new AI Control Plane is planned to discover and register each Agent and AI capability, establish identities and policies, manage lifecycles, evaluate performance, observe behaviour and outcomes, and control costs across both Salesforce and third-party AI. The model layer also proposes routing according to accuracy, performance, cost and business requirements.

The release sets an explicit maturity boundary at the same time: many of the technologies underpinning the Harness are already available, but the new capabilities and unified experience are only planned to begin rolling out from the start of Salesforce FY28. Availability, packaging, pricing and upgrade paths will be announced closer to GA and may vary by region and customer agreement. Salesforce also explicitly tells customers to base purchasing decisions on products and services that are currently available. The six-layer architecture can therefore serve as a checklist of controls, but cannot be described as a complete AI Control Plane that is already GA.

What this may mean for enterprise adoption

Architectural commitments, existing products and future rollout are three different states in an enterprise platform catalogue. The unit that can genuinely be accepted for a control plane is not a conceptual diagram, but queryable Agent identities, policy versions, model resolution, evaluation batches, records of behaviour and business outcomes, cost attribution and lifecycle actions. Fields that have yet to be released remain in an unknown state and cannot be replaced by the roadmap.

Signal 02

Cost records enter an API, but their granularity still has boundaries

At 00:07:15 on 12 September (Beijing time), Baseten released Model API costs. The new Management API route `v1/billing/model_apis`, called with HTTP GET, returns a subtotal for each calendar day after usage-based pricing, and callers can provide start_date and end_date. The official description says that cost history begins at 20:45 UTC on 5 August 2026. Compared with viewing an invoice only at month end, this provides a machine interface for retrieving daily Model API costs automatically.

The announcement promises only a daily subtotal for Model APIs. It does not say that training jobs, Dedicated Inference, storage, networking or other tool costs will enter the same endpoint, and provides no per-request costs, business tags, budget thresholds or anomaly alerts.

What this may mean for enterprise adoption

A daily API can become one layer of the underlying cost ledger, but cannot independently explain why a particular Agent task became more expensive. Baseten's exact subtotals are the provider's billing result; they do not mean that the enterprise has already completed attribution to a department, customer or business process. A complete analysis still depends on request volume, token usage, tool calls, retries, the model actually routed, internal cost centres, original invoices and price snapshots.

Signal 03

Training data paths begin to use short-lived workload identities

At 05:53:45 on 12 September (Beijing time), Baseten released support for OIDC and AWS AssumeRole for training jobs. During setup, a training job can use these methods to pull private container images and download model weights or training data without storing long-lived cloud credentials in Baseten. OIDC covers Amazon ECR, S3, Google Artifact Registry, Container Registry and Cloud Storage, and a token's access scope can be limited to a Baseten organization, team, training project or single job.

For AWS resources, users can instead avoid registering an OIDC provider and allow Baseten to use an organization-unique external ID to assume an IAM role in the customer's account. This path covers private ECR images and weights or data in S3. The announcement explicitly scopes the capability to training job setup; it does not promise that inference runtime access, arbitrary external access during training or other cloud services use the same mechanism.

What this may mean for enterprise adoption

The AI platform identity ledger therefore expands from a human API key to workloads. It covers the issuer of a token or role, the job, resource scope, lifetime, actual access and revocation status, and needs to distinguish setup, training runtime and inference runtime. Short-lived identities reduce the storage of long-lived keys, but the customer's responsibility for configuring the IAM trust policy, least privilege, external ID and audit logs remains. Without those records, it is not possible to establish whether a short-lived identity crossed its intended data boundary.

Signal 04

Reproducible evaluation and physical lifecycle add Day-2 evidence

Claude Code v2.1.269 was released at 03:17:55 on 12 September (Beijing time), adding `claude plugin eval`, which can run a plugin eval suite and generate scored, reproducible JSON and HTML results. The same version also added repository-level OpenTelemetry labels and made the Agent concurrency limit for a single workflow configurable from 1 to 256.

At 23:00:02 on 11 September (Beijing time), Dell Private Cloud published its account of physical lifecycle management, covering server discovery, firmware and driver alignment, cluster readiness, upgrade pre-checks, capacity expansion, safe reclamation, drift and compliance. The solution is built on PowerEdge, Dell Storage and bring-your-own licensing, can be reused across VMware vSphere/VCF, Red Hat OpenShift, Nutanix AHV and Microsoft Azure Local, and is supported by Dell through a single point across the stack. The fastest time-to-ready figure of 2.5 hours comes from Dell 2025 internal testing: measurement began after hardware installation, configuration and platform onboarding had been completed, and no human interaction was required after initiation. The claim of up to 90% fewer manual deployment steps also comes from Dell internal testing. The claim of up to 65% cost savings is relative to HCI and comes from a Dell-commissioned 2026 analysis by Omdia; actual results may vary.

What this may mean for enterprise adoption

Day-2 evidence for AI spans both the application and physical layers. The upper layer includes evaluation inputs, scores, reports and tool versions; the lower layer includes hardware, firmware, drivers, virtualisation platforms, drift and upgrade state. Claude's figure of 256 is a configurable concurrency limit, not validated throughput, and a vendor evaluation cannot establish coverage of an enterprise's own data, permissions and failure modes. Dell's efficiency figures are likewise bounded by the test start point, comparison target and commissioned-research methodology. Neither replaces the enterprise's own acceptance standard; reproducibility first depends on whether the environment has been fixed.

Verification

Sources and verification

  1. Salesforce Introduces the Trusted Enterprise AI HarnessSalesforce · 2026-09-11T03:09:31+00:00 · Official announcement
  2. Model API costsBaseten · 2026-09-11T17:07:15+01:00 · Official announcement
  3. OIDC and AWS AssumeRole for training jobsBaseten · 2026-09-11T22:53:45+01:00 · Official announcement
  4. Claude Code v2.1.269Anthropic / GitHub · 2026-09-11T19:17:55Z · Project release
  5. The Hardest Part Of Private Cloud Isn’t The HypervisorDell Technologies · 2026-09-11T15:00:02+00:00 · Official announcement

Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.

← Back to AI Daily Briefing