OpenAI just turned its internal ChatGPT Work agent infrastructure into a public API — and quietly set a shutdown clock on the testing tool many teams rely on to keep agents from breaking in production. If you’re shipping anything agentic this quarter, both facts change your roadmap, not just your reading list.
Why it matters
Until now, the orchestration, tool-calling, and long-running session handling that powered ChatGPT Work lived behind OpenAI’s own product wall. Builders had to stitch together similar behavior themselves with LangGraph, custom state machines, or homegrown retry logic. That plumbing is now exposed as an API primitive, alongside a new full-duplex voice model called GPT-Live-1. At the same time, OpenAI’s Evals — the tool many teams use for regression testing on prompts and agent behavior — is being phased out on a fixed timeline. Together, these moves mean the practical work of shipping AI products this quarter shifts from “which model is smartest” to “how do I test, price, and maintain an agent stack that OpenAI is actively rebuilding underneath me.”
What actually shipped
- Scaled agent infrastructure, publicly available. The orchestration layer behind ChatGPT Work — designed for long-running assistants and multi-step workflows rather than single-turn chat — is now callable through the standard API, with a documented fast-setup path.
- GPT-Live-1, a speech-to-speech model. This is a full-duplex voice model that can run alongside a reasoning model and tool-calling stack, priced at $0.05 per minute of voice, billed per second. That per-second billing detail matters for cost modeling on real call volume, not demo minutes.
- Evals is being sunset. Per OpenAI’s API docs, existing Evals become read-only on October 31, 2026, with full shutdown on November 30, 2026. That’s over a year out, but migration plans for automated testing don’t build themselves in a weekend.
- Large-context agentic APIs are being discussed in the ecosystem — one report cites an Astra-style agentic API with a 1.05 million-token context window and up to 128,000 output tokens. Treat this as an emerging claim to verify against your own provider’s documentation before you architect around it, not a confirmed OpenAI spec.
What this means if you ship product this quarter
Three concrete decisions should be on your sprint board, not your someday-list.
1. Stop treating voice as a bolt-on
If your product roadmap includes a voice assistant, GPT-Live-1’s full-duplex, per-second pricing changes the build-vs-buy math. Full-duplex means it can listen and speak simultaneously, which matters for interruption handling — a detail that’s easy to demo well and expensive to get wrong in production call centers or voice agents. Before committing, model your actual expected call duration and interruption rate against the $0.05/minute rate, not a best-case scripted call.
2. Build an Evals exit plan now, even though shutdown is 2026
Read-only starts October 31, 2026; full shutdown is November 30, 2026. That sounds far off, but if your CI/CD pipeline depends on automated eval runs for prompt regressions or agent behavior checks, you need to know what replaces that dependency before it becomes urgent. Options worth evaluating now include open-source eval frameworks, third-party observability platforms, or in-house regression harnesses. The mistake to avoid is discovering this dependency the week Evals goes read-only.
3. Measure cost per successful workflow, not cost per token
This is the theme builders keep coming back to, and for good reason. A cheaper model that requires three retries, extra tool calls, or fails on interruption handling can cost more per completed task than a pricier model that finishes cleanly on the first try. Before switching production traffic to any new model or agent stack, run the comparison properly.
| Old mindset | New mindset for this quarter |
|---|---|
| Compare price per 1K tokens | Compare cost per successfully completed workflow |
| Assume routing evals from last quarter still apply | Rerun routing evals against updated models and agent infra |
| Manual QA before launch | Automated eval pipeline with a documented migration plan past 2026 |
| Voice as a separate vendor integration | Voice as a first-class API layer alongside reasoning and tools |
The honest caveat
Public agent infrastructure and cheaper voice APIs lower the barrier to building agentic products, but they also mean more of your product’s reliability depends on infrastructure you don’t control and that OpenAI can change again. Teams that skip building their own eval and cost-tracking layer — assuming the vendor’s tools will always be there — are the ones who will scramble in late 2026.