OpenAI Frontier

What it is

OpenAI Frontier is OpenAI’s enterprise platform for building, running and governing AI agents across a company’s systems. It launched on 5 February 2026 and is a different purchase from a ChatGPT Enterprise seat or a per-user Dots agent: those give one employee an assistant, Frontier gives the company a place to deploy agents that run business processes and are managed like staff.

OpenAI’s product page splits it into four layers:

• Business Context connects data warehouses, CRMs, ticketing tools and internal apps into what OpenAI calls a semantic layer, so every agent reads the same picture of how the business works and builds memory over time.
• Agent Execution gives agents an environment to work with files, run code and use tools. Agents can run in local environments, in the customer’s own cloud infrastructure or on OpenAI-hosted runtimes.
• Evaluation and optimization loops that show managers which agent behaviour works and which does not.
• Identity, permissions and audit. Each agent gets its own identity with explicit permissions, under the same IAM that covers employees, plus monitoring and action logs. The page lists SOC 2 Type II, ISO/IEC 27001, 27017, 27018, 27701 and CSA STAR.

The launch post says Frontier works whether agents are “developed in-house, acquired from OpenAI, or are integrated from other vendors,” and that they can be reached through ChatGPT, Atlas or inside existing business apps. That claim is the core of the “do we govern everything here” question, and it is the claim to test first (see Watch-outs).

Who can actually buy it

Nine months after launch, Frontier is still sold only through OpenAI’s sales team. The launch post said it was “available today to a limited set of customers,” with broader availability “over the next few months.” As of 10 November 2026 the product page offers only Contact sales: no price list, no self-serve signup, no public admin documentation.

The named customers are large enterprises. HP, Intuit, Oracle, State Farm, Thermo Fisher and Uber were first to adopt, and BBVA, Cisco and T-Mobile piloted the approach. The surrounding programme tells you the intended buyer:

• Frontier Partners, AI-native vendors building on the platform: Abridge, Clay, Ambience, Decagon, Harvey and Sierra.
• Frontier Alliances, multiyear deals announced 23 February 2026 with Accenture, BCG, Capgemini and McKinsey.
• AWS was named the exclusive third-party cloud distribution provider for Frontier on 27 February 2026.
• The OpenAI Deployment Company (11 May 2026), a majority-OpenAI-owned services business launched with more than $4 billion of initial investment, led by TPG with 18 other firms. Its forward deployed engineers staff the Enterprise Frontier Program. OpenAI agreed to buy the consultancy Tomoro to bring about 150 of them on day one.

Frontier comes with a services engagement. It is not a tool an ops team switches on.

Pricing reality

OpenAI publishes no Frontier price. Expect a bill in three parts. The platform contract is negotiated per deal. Model usage comes on top. The deployment engagement, with OpenAI or Deployment Company engineers or an Alliance consultancy, is the third part, and its scope is set by how many workflows you take to production. Third-party guides estimate six-figure annual contracts. Those are estimates, not reported deals, and should not anchor a budget.

Without a price, benchmark against what the alternatives charge for the same job:

• Governance only. Microsoft Agent 365 is $15 per user per month on top of E5 or Business Premium. For 2,000 employees that is $360,000 a year, and it governs agents without building or running any.
• Execution inside the CRM. Agentforce Flex Credits work out to $0.10 per standard agent action, so 1 million actions a year is $100,000 before the Salesforce edition underneath.
• Build it yourself in AWS. Bedrock AgentCore publishes its meters: Runtime at $0.0895 per vCPU-hour and Gateway at $0.005 per 1,000 tool calls, plus tokens.

If a Frontier quote for comparable scope sits far above those numbers, the difference is paying for the engineers and the context layer. Ask for those as separate lines.

Best for

A CIO, head of AI or business-operations leader at a large enterprise already standardised on OpenAI models, who wants agents to run cross-system processes end to end and has budget for an embedded deployment team. OpenAI’s own page names revenue operations, customer support and procurement as target processes. The scoped case where it fits: a RevOps process like quote approval or renewal risk that reads from the warehouse, writes to the CRM and opens tickets, where nobody inside the company has built production agents before.

Not for

A team of 5-30 that wants agents running this quarter. With no self-serve path and an engagement model built around forward deployed engineers, the AI agent ops stack on n8n gets an owner per agent, approval gates and execution logs in place in days. It is also the wrong choice if your control requirement is a single registry over agents from several vendors, because that is the least proven part of Frontier’s pitch.

Versus the alternatives

• Microsoft Copilot Studio plus Agent 365, Microsoft being one of the two market-share leaders. Pick Microsoft when identity lives in Entra and the agents ship into Teams and Microsoft 365. It splits build and govern into two published SKUs, and Agent 365 already registers Copilot Studio, Foundry and n8n agents.
• Salesforce Agentforce, the other share leader at $1.2B ARR in May 2026. Pick it when the agents that matter act on Salesforce data. Governance next to the record is easier to prove than governance that federates to it.
• Claude Cowork and Claude Enterprise. Pick Claude when the work is human-driven sessions, not unattended processes, and you want to self-serve now. Enterprise is $20 per seat plus usage with organisation-level model controls.
• n8n, the fastest-growing entrant in agent automation, with 10x revenue growth in the year to its October 2025 Series C and a $5.2B valuation in May 2026. Pick it when workflows are mostly deterministic, you need to self-host, or you want per-execution pricing you can forecast. Its agents also register in Agent 365.
• Amazon Bedrock AgentCore. Pick it when the security team wants agents inside your AWS account with published meters and a choice of Claude, GPT-5.5 and open-weight models.

Watch-outs

• “Limited availability” has lasted nine months. What you sign may include capabilities that are not generally available. Guard: get a written list of which Frontier capabilities are GA on your contract start date, and tie acceptance milestones to that list, not to the launch demos.
• Governing other vendors’ agents is a claim, not yet a demonstrated feature. OpenAI says Frontier integrates agents from other vendors through open standards. At launch, VentureBeat reported that OpenAI had not said whether other vendors’ models and tools would be allowed. Guard: in the pilot, register one non-OpenAI agent, such as a Claude-based or n8n agent, and confirm that its identity, permission scoping and action logs appear in Frontier the same way OpenAI agents’ do. If they do not, keep Agent 365 or Okta as the cross-vendor registry.
• The context layer becomes the switching cost. The semantic layer, agent memories and evaluation sets are the most valuable things a Frontier deployment builds, and they live on OpenAI’s platform. Guard: write an export clause into the contract that covers context-layer definitions, memories, eval datasets and audit logs, in a documented format, and test one export before the renewal date.
• The deployment knowledge sits with the vendor’s engineers. If forward deployed engineers design the architecture and governance, your team can end up unable to change it after they leave. Guard: name an internal owner per workflow from week one, and make a handover milestone, with your staff shipping one agent change unassisted, a condition of the services statement of work.

Related: the agent access governance stack covers the identity and permission layer around any agent platform, and the enterprise AI rollout stack prices the layers around it.