Building an AI agent able to work autonomously for hours, without coding all the infrastructure behind it, used to be expensive. OpenAI changed that on September 10, 2026, opening up to all developers, through its new Agents API in public beta, the same infrastructure that already runs Codex, its coding agent. For an SME considering a custom AI agent to automate a task, this announcement is worth understanding, even without technical skills.
In brief
- On September 10, 2026, OpenAI launched the Agents API in public beta: it exposes to developers the infrastructure (long-running sessions, execution sandbox, sub-agent coordination) that already powers Codex.
- No additional fee for the API itself: businesses only pay for model tokens, tools used and, where applicable, sandbox compute, at standard rates.
- The API is built on four core objects: agent (model, instructions, tools), environment (sandbox), session (durable state, sometimes spanning several days) and events (stream of results).
- Nine sandbox hosting partners were announced at launch, including Cloudflare, DigitalOcean, Oracle Cloud and Vercel.
- By default, agents built through this API use GPT-6 Astra, OpenAI's flagship model released in early September 2026, though a cheaper model can be selected instead.
What the Agents API actually is
The Agents API is a service that lets a developer create an AI agent able to run code, edit files, search the web and coordinate several sub-tasks, without having to build the most expensive part themselves: orchestrating long sessions, managing context and providing a secure execution environment (the "sandbox"). Until now, this infrastructure was either built in-house by large technical teams or billed by third-party frameworks such as LangChain or CrewAI, on top of the model's own cost.
By making this layer free beyond model usage, OpenAI turns it into a standard building block, similar to how web hosting became a commodity two decades ago. The service remains a developer tool, though: an SME without a technical team won't call it directly, but will benefit from it through the no-code tools and integration agencies that build on top of it.
Define the agent
Open a session
Run inside an environment
Receive events
Why this launch matters for an SME
An SME will almost never call this API directly. What matters is what this launch signals for the ecosystem of tools business leaders already use or are considering.
Before September 10, 2026
Since September 10, 2026
Three concrete consequences worth watching for an SME leader:
- Custom automation quotes should come down. An agency or freelancer building a specific agent for your business (document processing, automated monitoring, augmented customer support) has historically passed on the cost of orchestration infrastructure in its quote. With that layer now free beyond token usage, the "infrastructure" share of a quote should shrink.
- The no-code tools you already use will evolve. Platforms like n8n or Make can integrate this layer to offer agents capable of working autonomously on long tasks, rather than simple event-triggered automations.
- The question to ask your vendors changes. Instead of "how much does your agent framework cost?", ask instead: "how many tokens and tool calls will my agent use per month?" That consumption, not an infrastructure license, will now drive most of the bill.
A table to compare the approaches
| Approach | Who picks it | Main cost | Technical level required |
|---|---|---|---|
| No-code tool (n8n, Make) | SME without a dev team | Monthly subscription + AI tokens | Low to medium |
| Custom agent via Agents API | Agency/in-house dev for the SME | Model tokens + tools + sandbox | High (developer required) |
| Legacy in-house agent | Large enterprises before 2026 | Framework license + tokens + development | Very high |
Worth checking before you dismiss this
Even without technical skills, it's worth asking your current automation provider whether it plans to adopt this new infrastructure. A cost reduction on the vendor's side only reaches your invoice if you ask for it.
Limits worth keeping in mind
This announcement is recent and worth following with some perspective, in the interest of honesty:
- The API is in public beta: OpenAI itself notes that pricing for long sessions and sandbox operations may still change.
- By default, agents use GPT-6 Astra, a premium model that costs more than an entry-level one; the developer must explicitly pick a cheaper model to lower the bill.
- This infrastructure remains developer-only: it does not, on its own, replace guidance on defining a use case that's actually useful for your business.
- OpenAI is not alone here: Anthropic and Google offer comparable approaches for their own models, so comparing vendors still matters before any commitment.
FAQ
What is OpenAI's Agents API?
The Agents API is a service OpenAI launched on September 10, 2026, in public beta, giving developers access to the infrastructure (long-running sessions, execution sandbox, sub-agent coordination) that already powers Codex, OpenAI's coding agent.
Should an SME use this API directly?
No, in most cases. This API is aimed at developers and integration agencies. An SME will benefit from it indirectly, through no-code tools or custom agents built by a provider relying on it.
How much does an agent built with the Agents API cost?
OpenAI charges no additional fee for the API itself. The cost depends on model tokens consumed, tools used and, if needed, sandbox compute, at the API's standard rates.
Will this lower the price of automation tools I already use?
That's likely for custom agents and for future no-code tool features, but nothing is automatic. It's worth asking your current provider directly whether it plans to build on this new infrastructure and pass the cost reduction on to you.
AI agent infrastructure is standardizing fast in 2026. To learn how to make the most of it without blowing your budget, check out our other resources on AI automation or see how SMEs like yours have structured their agent projects in our customer stories.
