OpenAI Presence is an enterprise product for deploying voice and chat agents, announced on July 22, 2026, and you cannot buy it with an API key. Deployments are led by OpenAI's own Forward Deployed Engineers and selected systems integrators, there is no published pricing, and OpenAI says plainly it is "not yet available as a self-serve product". If you build voice agents for a living, the most important part of the launch is the sales model, not the technology.
I read the announcement, the help centre documentation, and most of the launch coverage so you don't have to. Earlier this month I wrote about OpenAI locking its best voice model away from developers, and Presence is the same story told at a different layer of the stack. The consumer got GPT-Live. The enterprise gets Presence. The developer gets a blog post saying the API "continues to be supported".
What is OpenAI Presence?
Presence bundles the things enterprises keep failing at when they take agents to production: policies and standard operating procedures, guardrails, approved actions, simulations, evaluation graders, and a Codex-powered improvement loop that reads production sessions and proposes changes for humans to approve. Each deployment starts with one narrow job, resolving a billing dispute or handling an insurance claim, and the agent gets only the knowledge and system access that job needs.
The proof point OpenAI leads with is its own phone support line, 1-888-GPT-0090. The company says the agent resolves 75% of inbound issues without human assistance and that the Codex loop cut human handoffs by 15 percentage points in 10 days. Those are OpenAI's numbers, graded by OpenAI's own criteria, on OpenAI's own channel, so hold them loosely. But even discounted, they describe a system that improves in production without an engineer rewriting prompts by hand, and that's the part I'd actually want to steal.
The named customers are earlier in the journey than the launch language suggests. BBVA is exploring voice banking support in Mexico, SoftBank is testing Japanese-language conversations, and Australian insurer IAG is exploring surge support during severe weather events. Exploring and testing. Nobody is named as running this at scale, which sits a little awkwardly next to the phrase "battle-tested" in the second paragraph of the announcement.
Why is Presence not self-serve?
Because the hard part of production agents was never the model. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, and it blames governance, undefined business value, and weak operational discipline rather than model capability. Presence is aimed squarely at that diagnosis. The six-stage deployment process in OpenAI's help docs (scoping, security and legal review, simulation, acceptance testing, staged rollout, post-launch iteration) is a services engagement with a product attached.
Forward Deployed Engineer is a job title borrowed from Palantir, and the borrowing is the tell. Palantir built a very good business embedding engineers inside customers for months because the software only creates value once someone wires it into messy internal systems. OpenAI has now concluded the same thing about agents. It launched the OpenAI Deployment Company with Bain in May, Anthropic launched its Ode consulting arm a week before Presence, and the whole frontier tier is converging on the position that implementation is the business, not inference.
Eligibility for Presence depends on workflow fit, implementation readiness, and "available delivery capacity". That last one is the constraint that matters. Software scales. Engineers cleared into a bank's core systems do not. Whatever Presence costs (nobody knows, VentureBeat asked twice and got silence), it is rationed by headcount, which means it goes to the BBVAs and SoftBanks of the world and nobody else for a long time.
What does OpenAI Presence mean for developers?
The pattern from OpenAI this month is consistent and worth naming. GPT-Live, its best voice model, shipped consumer-only with no API, no pricing, and a waitlist form. Presence, its agent platform, ships enterprise-only with no API, no pricing, and an eligibility review. The API tier, the thing most of us actually build on, got one sentence in the Presence announcement: OpenAI will "continue supporting voice customers with access to our frontier models through the OpenAI API".
I don't think that sentence is a threat. The Realtime API isn't going anywhere. But it does tell you where you sit in the queue when the interesting stuff ships. If your voice agent roadmap depends on OpenAI's newest capabilities arriving as self-serve APIs shortly after announcement, this month should recalibrate that assumption twice over.
There's a genuinely useful lesson buried in here too. What OpenAI is selling to enterprises is mostly not model quality, it's the boring wrapper: simulation before launch, graders checking policy compliance and tool use, guardrails with intervention boundaries, escalation paths that hand humans structured context, controlled rollouts with rollback. None of that requires a Forward Deployed Engineer. It requires treating your agent like production software instead of a demo that escaped. If you're building voice agents on self-serve APIs (I work on one at Speechify, on the SpeechifyAI platform, weigh my bias), that checklist is free to copy and most teams skip it.
The self-serve stack has one structural advantage Presence can't match: you can start today, see the pricing, and leave. SpeechifyAI publishes voice agent pricing from $0.068 per minute all-in, with the models co-located so per-turn latency doesn't leak through vendor seams. OpenAI's offer to the same buyer is a scoping call and a security review. Both are legitimate ways to sell agents. Only one of them is available to you this afternoon.
The comparison that actually matters
| OpenAI Presence | Self-serve agent APIs | |
|---|---|---|
| Access | Limited GA, eligibility review | Sign up, API key |
| Deployment | OpenAI FDEs and systems integrators | Your team |
| Pricing | Undisclosed, per engagement | Published (SpeechifyAI from $0.068/min) |
| Models | OpenAI models, configuration may change | You choose, you pin versions |
| Governance layer | Bundled: simulations, graders, guardrails | Yours to build (and you should) |
| Time to first call | Weeks of scoping, minimum | Today |
The honest read is that these serve different buyers. A bank that failed two agent pilots wants someone accountable sitting inside the building, and Presence is a rational purchase for them. A product team that ships weekly wants control, published pricing, and the ability to swap components, and a managed engagement led by someone else's engineers is the opposite of that.
FAQ
What is OpenAI Presence?
OpenAI Presence is an enterprise platform for deploying AI voice and chat agents, announced July 22, 2026. It bundles policies, guardrails, approved actions, simulations, evaluation tools, and a Codex-powered improvement loop. It is available only through a limited general availability programme, with deployments led by OpenAI Forward Deployed Engineers and selected systems integrators. There is no self-serve option and no published pricing.
Can developers use OpenAI Presence through an API?
No. Presence is not available as an API or a self-serve product. Access depends on an eligibility review covering workflow fit, implementation readiness, and OpenAI's own delivery capacity. Developers building voice agents programmatically still use the OpenAI Realtime API, or a self-serve voice agents platform like SpeechifyAI, which publishes pricing from $0.068 per minute and has been in public beta since June 2026.
How much does OpenAI Presence cost?
OpenAI has not published pricing. Implementation scope and cost are set per customer and per deployment, which is normal for services-led enterprise deals but leaves buyers with no public reference point. For comparison, self-serve voice agent platforms publish rates, with SpeechifyAI voice agents from $0.068 per minute all-in and Simba 3.2 text-to-speech from $6 per million characters on the Scale tier.
What TTS do voice agents use if not OpenAI?
Any agent stack that separates the voice layer can use a dedicated TTS model. Simba 3.2 is the strongest option on current evidence: it ties for first on the Artificial Analysis Speech Arena at 1,234 Elo, streams at 30.2 characters per second, and costs $10 per million characters at list. Co-locating TTS with your ASR and LLM matters more than most teams expect, saving 75 to 150 ms per conversational turn.
Is OpenAI copying Palantir's business model?
Partly. The Forward Deployed Engineer title and the embed-engineers-with-customers delivery model both come from Palantir's playbook. OpenAI also launched a dedicated consulting arm, the OpenAI Deployment Company, with Bain & Company in May 2026. Anthropic did the same with Ode in July. The frontier labs have converged on the view that enterprise value sits in implementation, not raw model access.
Top comments (0)