Hey guys, Mr. Technology here.
The OpenAI changelog for September 10, 2026 published three substantial entries in a single day. The Agents API moved to public beta — a managed Codex harness that handles session orchestration, context compaction, and recovery so builders can focus on tools and prompts. GPT-Live 1 went generally available on the API at $0.05 per minute of voice (billed per second), with full-duplex voice that can run alongside a backend model or agent. And project API keys now support an expiration date at creation, with an organization- or project-level maximum-key-lifetime policy that admins can enforce in Platform settings. All three are documented on the primary source (OpenAI changelog — September 10, 2026, fetched 2026-09-11 14:09 UTC).
This is a documentation-surfacing report. Every claim below is verifiable verbatim in the changelog entry and the linked reference pages. I have not run the Agents API, GPT-Live 1, or the new project key expiration flow against a live organization; treat the operational details as documentation comparison, not firsthand test.
From the OpenAI changelog, dated September 10, 2026 (verified verbatim 2026-09-11 14:09 UTC), three entries:
"Released the Agents API in public beta. Build agents with a managed Codex harness while OpenAI handles session orchestration, context compaction, and recovery."
"Use durable sessions to continue work across turns, stream progress, and connect your own tools and MCP servers. Run agents in OpenAI-hosted sandboxes or connect a sandbox from your own infrastructure or a supported provider."
The Agents API is the OpenAI-hosted equivalent of "spin up a Codex CLI as a managed service." Builders do not have to provision, monitor, or recover sessions themselves; OpenAI runs the harness, the compaction, and the recovery machinery on the developer's behalf. Sessions are durable across turns, progress is streamed, and the agent can be configured to call MCP servers or the developer's own tools. The sandbox can be OpenAI-hosted or bridged into the developer's own infrastructure or a supported provider.
This is the managed-Codex product that the prior Codex CLI releases were building toward. The build-vs-buy question for autonomous-agent infrastructure gets a clear "buy" option today.
"GPT-Live 1 is now generally available in the API. Build full-duplex voice conversations that can continue while a backend model or agent handles reasoning and tools."
"Use Responses delegation with an OpenAI model, or client delegation to connect your own backend. Voice sessions cost $0.05 per minute, billed per second; backend model and tool usage is charged separately."
GPT-Live 1 is OpenAI's first generally available full-duplex voice model on the API. Two integration shapes are documented: Responses delegation (the voice session uses an OpenAI Responses-API model for reasoning) and client delegation (the voice session is connected to the developer's own backend). Voice itself is billed at $0.05/minute, per-second billing; backend model and tool usage is charged separately under the usual pricing tables.
The $0.05/minute number is the editorial signal: per-second billing and a flat per-minute rate make cost modeling trivial — a 10-minute voice session is $0.50 of voice, plus whatever backend model and tool tokens it consumes.
"You can now set expiration dates when creating project API keys. Administrators can also enforce a maximum key lifetime at the organization or project level in Platform settings, requiring newly created keys to expire within the configured limit."
Two distinct controls:
The combination closes the long-standing practice of issuing non-expiring API keys that then leak through stale CI secrets, deprecated deployments, or former employees. Both controls are documented on the production best practices guide.
The Agents API public beta is the bigger story. A managed Codex harness with durable sessions, OpenAI-hosted or your-own-infrastructure sandboxing, and built-in MCP / custom-tool integration is the missing piece for builders who want the operational shape of Codex without operating it themselves. For a founder evaluating "do I build my agent product on a hand-rolled runtime, on Codex CLI in our own fleet, or on the Agents API," the Agents API moves "buy" from a hypothetical to a real option. The cost is the cost of OpenAI sandboxes + model + tool usage; the operational savings is "no harness, no compaction, no recovery."
GPT-Live 1 GA at $0.05/minute unblocks production voice products that needed a vendor-billed voice layer rather than a self-hosted STT/TTS pipeline. Two integration shapes matter: Responses delegation (one-call integration with an OpenAI model) and client delegation (integration with your own backend) let builders pick their backend story. Per-second billing keeps cost predictable. The first generally available voice model is also the first OpenAI voice product where the cost of the voice itself is small relative to the cost of the backend reasoning — for an agent that's mostly talking to tools, the voice bill will be a rounding error on the tool bill.
Project API key expiration closes a security practice that's been overdue. Most enterprise API-key handling guidance says "rotate keys on a schedule"; OpenAI's prior surface required the developer to remember to rotate. The new surface lets the developer set an expiration at creation and the admin enforce a maximum lifetime at the org or project level. For teams in regulated environments (SOC 2, ISO 27001, FedRAMP), the admin-enforced max-lifetime control is the path to compliance evidence without relying on team discipline.
The Agents API is the new top-of-stack option for managed agent infrastructure. The Codex CLI releases over the past two months — v0.152.0 (Vim mode, MCP package names, app-server timeouts), v0.153.0 (Vim undo/redo, plugin CLI, experimental context), v0.154.0 (experimental worktrees, inline ask, Windows server sharing) — were all building toward the harness that the Agents API now exposes as a service. Builders who want the harness without operating the CLI can switch to the Agents API; builders who want the CLI's local-first ergonomics keep that path.
Voice gets a per-second billing surface. Prior OpenAI voice products (Realtime API, Audio API) billed input tokens and output tokens; GPT-Live 1 bills the voice session itself at $0.05/minute on top of the usual model and tool rates. For a builder modeling a 1-hour voice session: $3.00 of voice + whatever the backend model and tools cost. The per-second billing means the rate scales with actual use, not provisioned capacity.
Non-expiring API keys are no longer the default for projects with a managed policy. If your Platform admin sets a maximum key lifetime, any newly created key in that project must expire within the window. Existing keys are not retroactively expired; the policy applies to new key creation. The combination of per-key expiration at creation and org-level policy enforcement is the new baseline for project API key management.
Primary sources verified at 2026-09-11 14:09 UTC:
No firsthand test was run against the Agents API, GPT-Live 1, or the project key expiration flow. The article is a documentation comparison, not a benchmark or field report.
Cost. Two distinct cost surfaces:
Risk.
Limitations of this report.
Three substantial OpenAI API entries in a single day. The Agents API public beta is the bigger story — a managed Codex harness with durable sessions, MCP integration, and OpenAI-hosted or your-own-infrastructure sandboxing is the build-vs-buy answer for autonomous-agent infrastructure. GPT-Live 1 GA at $0.05/minute voice with per-second billing is the operational unlock for production voice products. Project API key expiration is the overdue security control for regulated environments.
For a founder evaluating "do I build my agent product on a hand-rolled runtime, on Codex CLI in our own fleet, or on the Agents API" — the Agents API moves "buy" from a hypothetical to a real option. The cost is OpenAI sandbox + model + tool usage; the operational savings is "no harness, no compaction, no recovery."
For a builder shipping voice products, GPT-Live 1 is the first vendor-billed voice layer with predictable per-second pricing. For a security team reviewing API key hygiene, the new project API key expiration control is the path to compliance evidence without relying on team discipline.
Today
This week
Skip if not in scope
Originally published: 2026-09-11 14:08 UTC Last verified: 2026-09-11 14:09 UTC No corrections at this time.