OpenAI Frontier vs Presence: What’s the Actual Difference?

The Platform: What Frontier Actually Is

Frontier is an enterprise platform, launched February 5, 2026, built to give AI agents the same things a new hire gets on day one: context about how the business works, permission to touch the right systems, and a manager watching to see if they’re actually any good at the job. Picture it less like software and more like a workplace full of employees who happen to be made of code — that’s genuinely the pitch.

The architecture reflects that ambition. Frontier links a company’s data warehouses, CRM systems, and ticketing tools into what OpenAI calls a shared business context — essentially a memory every agent can draw from. On top of that sits the part where agents actually do things: running code, opening files, taking real action instead of just suggesting it. And above that, an evaluation layer grades the agents the way a manager would grade an employee, quietly getting stricter as the stakes go up.

Frontier’s architecture: shared business context feeds agent execution, which feeds an evaluation layer, all built on open standards.

Why It’s Open, Not Just Big

What makes Frontier interesting rather than just another walled garden is that it’s explicitly not one. Agents built by OpenAI, by the customer, or by a third-party vendor can all plug into the same shared context using open standards — which raises a question worth sitting with: what actually separates an agent from a chatbot in the first place, once it can move freely between platforms like this. Frontier doesn’t answer that outright, but it does make the question harder to avoid.

The company didn’t launch this quietly, either. Its first cohort of adopters reads like a small piece of corporate America:

• HP
• Intuit
• Oracle
• State Farm
• Thermo Fisher
• Uber

with BBVA, Cisco, and T-Mobile already quietly piloting the approach before the official unveiling. OpenAI also named six “Frontier Partners” — Abridge, Clay, Ambience, Decagon, Harvey, and Sierra — AI-native vendors building directly against the platform rather than around it.

None of this, though, is something you can simply try. Access is sales-led only. There’s no self-serve signup, and OpenAI has said broader availability is coming “over the next few months” — a phrase that, in February, meant sometime this year, and in July, still does.

The Product: What Presence Actually Is

Presence is a managed product, launched July 22, 2026, that deploys voice and chat agents into tightly scoped tasks — resolving a billing dispute, processing an insurance claim, clearing an IT service ticket. If Frontier is an open floor plan, Presence is a locked room with one very specific job to do: nothing broad, nothing improvised, one job done well, over and over. It’s the same instinct behind these 8 AI agent use cases — narrow beats general, at least for now.

Presence scopes each deployment to one job, delivered by a Forward Deployed Engineer, improved through a human-approved Codex loop.

That narrowness isn’t an accident; it’s the entire design philosophy. Each Presence deployment starts with exactly one job, and the agent gets only the system access and knowledge that job needs — nothing more. This is the same instinct behind why so many companies running AI agents can’t actually control them — Presence is built specifically to not repeat that mistake. The company writes the rules: what the agent can do alone, when it needs a human’s sign-off, and when it should just hand the whole thing over.

After launch, the agent keeps learning — sort of. OpenAI’s Codex reads through production sessions and escalations, then proposes changes. A human team tests those changes and approves them before anything actually ships. Nothing updates itself in the wild.

The Delivery Model Is the Real Story

Here’s what separates Presence most sharply from ordinary SaaS: you don’t buy it so much as you’re handed it. It isn’t self-serve. Deployments are led by OpenAI’s own Forward Deployed Engineers — a role modeled on the embedded-engineer function Palantir made famous — alongside a small set of named systems integrators.

You don’t get an API key. You get a project team.

BBVA, SoftBank, and IAG are the named early partners — though all three are still exploring or testing, not running Presence at full scale.

So What’s the Actual Difference?

Strip away the branding and it comes down to this: Frontier is the layer you build on. Presence is the thing you buy pre-built. One is infrastructure for context, permissions, and oversight across any agent doing any job. The other is opinionated — voice and chat, one task at a time — and shows up with OpenAI’s own people already inside it, doing the integration work for you.

Frontier and Presence compared: scope, third-party agent support, and who actually builds the deployment.

Here’s the practical way to think about it. Trying to give a dozen internal teams shared access to company context? That’s Frontier’s problem. Need one voice agent to handle billing disputes starting next quarter, and would rather pay for a finished product than build one from scratch? That’s Presence. Some enterprises may eventually run both at once — Frontier as the foundation, Presence as one specific room built on top of it — though OpenAI hasn’t said the two are formally designed to work together.

Should You Actually Believe the 75% Number?

Here’s the honest answer: believe it as a proof of concept, not as a promise. That 75% figure comes entirely from OpenAI’s own phone support line, graded against OpenAI’s own internal quality standards, on a single English-language channel that OpenAI owns from top to bottom. Nobody outside the company has audited it.

That doesn’t make the number worthless — it’s a legitimate case study, and OpenAI has been unusually candid about the mechanics behind it. But if you’re an enterprise buyer picturing this exact result in your own contact center, adjust your expectations. OpenAI’s deployment has the advantage of clean internal systems and a narrow product surface that most large organizations simply don’t have.

The 75% figure comes from OpenAI’s own line only — BBVA, SoftBank, and IAG are still testing, not running Presence at scale.

If your systems are fragmented, your documentation is uneven, and your compliance requirements are heavier than a phone support line’s, expect a lower number at launch — one that climbs over time as the agent gets tuned, not a guaranteed 75% out of the box.

The Uncomfortable Question Nobody’s Asking

Here’s the part that doesn’t get said out loud enough: OpenAI is, in both of these deployments, both the vendor and the builder. The same company selling you the model is also the company configuring exactly how it behaves once it’s live in your business.

That’s a genuinely different arrangement than the one enterprises are used to — buying software from one company, then hiring an independent integrator to actually install it, where responsibility for something going wrong is usually traceable to one clear party. When OpenAI plays both roles at once, and a policy misfires in production — an action approved that shouldn’t have been, an escalation that should have triggered and didn’t — the question of whose fault that actually is gets a lot murkier, a tension flagged directly when Presence first launched.

When the model vendor is also the implementer, the usual line of accountability gets harder to trace.

None of this is new, exactly. Cloud consulting arms and systems integrators have blurred the line between vendor and implementer for decades. What’s different this time is the compressed timeline — enterprises are being asked to sign off on live AI agent behavior before the governance norms for this exact arrangement have had time to catch up.

What Happens to the People Whose Jobs This Touches?

Slower hiring, most likely — not a wave of layoffs, at least not yet. That’s the read from the analysts who actually examined Presence’s workforce implications, rather than taking OpenAI’s own framing at face value.

• Tulika Sheel, Senior Vice President at Kadence International, expects the first ripple to hit hiring and team growth rather than existing headcount — roles shifting gradually toward the complex cases and escalations that still need a human touch.
• Pareekh Jain, CEO of Pareekh Consulting, points to Tier-1 support agents handling predictable, repetitive queries as the most exposed group — though broader cuts would only follow once companies actually reorganize around the technology, not before.
• Lian Jye Su, chief analyst at Omdia, offers the calmest take of the three: this kind of automation isn’t new. Companies have leaned on similar tools from Genesys, NiCE, Five9, and AWS for years. The likely near-term reality is AI handling the routine, humans keeping the work that needs judgment and empathy.

Sheel, Jain, and Su each read Presence’s workforce impact differently — but none forecasts a cliff.

Read across all three, the pattern is consistent even where the tone differs: nobody credible is predicting a cliff, and everybody credible is predicting a shift.

Can You Even Get In?

Not easily, no. Access to both Frontier and Presence currently runs through direct enterprise sales — there’s no self-serve path for either one, and both are built strictly for large organizations.

Frontier is available to a limited set of customers as of its February launch, with broader access promised “over the next few months.” Presence runs through what OpenAI calls a “limited general availability program” — a narrower framing than Frontier’s, and one that fits its more consulting-style delivery. Eligibility for Presence depends on workflow fit, implementation readiness, and OpenAI’s available delivery capacity — a constraint that scales with engineering headcount, not licensing tiers.

Enterprise interest funnels down to sales-led eligibility — there’s no self-serve signup for either product.

If you’re a mid-market company without an existing OpenAI enterprise relationship, neither door is currently open to you. The realistic way in is an existing ChatGPT Enterprise or API relationship, followed by a direct conversation with an OpenAI account team — one worth having only after you’ve weighed what agentic AI actually costs against what you’re hoping to get back.

And How Does This Stack Up Against Anthropic?

OpenAI’s strategy compares to Anthropic’s on one key axis: packaging versus platform. Presence brands agent-governance as a standalone product with implementation services baked in, while Anthropic’s parallel enterprise push centers on helping companies deploy Claude directly, with no separate branded deployment layer sitting on top. The two companies are, in other words, telling slightly different stories right now.

Packaging versus platform: how OpenAI’s Presence and Anthropic’s enterprise approach diverge.

For an enterprise weighing the two ecosystems, the real difference has less to do with model quality and more to do with who ends up accountable for what. How ChatGPT and Claude actually compare for work and every Claude model, compared are worth reading before evaluating either vendor’s enterprise offering specifically.

Neither company has published a true head-to-head on enterprise agent deployment costs or outcomes. For now, anyone comparing the two is working from vendor-supplied claims on both sides — the same grain of salt that applies to OpenAI’s own 75% figure applies here too.

Naming the Pattern: The Orchestration-Deployment Split

Once you see it, you can’t unsee it: OpenAI has actually built a two-tier strategy for enterprise AI agents, and naming it makes the buying decision simpler.

The Orchestration-Deployment Split: one fork for shared context, one for scoped, engineer-delivered jobs.

Conclusion

Frontier and Presence are OpenAI’s two-tier answer to enterprise AI agents — the Orchestration-Deployment Split, orchestration versus deployment — and a quiet signal of the broader dashboard-to-agent shift already reshaping B2B software.

The number worth watching isn’t the 75% resolution rate — it’s whether OpenAI can scale Forward Deployed Engineers as fast as it’s signing enterprise logos, because that’s the actual bottleneck on both products right now.

Map your own workflow against the Split before you evaluate either one.