AI news videos

OpenAI Launches the Agents API in Public Beta

OpenAI has released the Agents API in public beta, opening the harness and infrastructure behind Codex to developers. Read it for what it is: the vendor's own product announcement, not an independent evaluation.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• Every factual claim in this script comes from a single first-party source, OpenAI's own announcement "Introducing the Agents API", published on openai.com on September 10, 2026 under the Product and API tags, captured at 2026-09-10T20:10:04.629Z. SENTENCE AUDIT FOLLOWS. Hook sentence one, "OpenAI launched the Agents API today in public beta", is supported by "Today, we're introducing the Agents API in public beta" and by "Agents API is available in public beta today to all developers"; I changed "we're introducing" to "OpenAI launched" only to name the vendor explicitly for a viewer who has no page context. Hook sentence two, "It runs cloud agents on the harness behind Codex", is supported by the article subtitle "Build and run cloud agents with the Codex harness, fully managed by OpenAI" and by "bringing that same harness and infrastructure that powers Codex to developers through a simple, flexible API"; I dropped the word "same" and the phrase "fully managed" for time and kept the plain factual relationship. Scene one sentence one, "You create an agent in one API call by naming the task model tools and environment", is supported by "With the Agents API, you can create a production-ready agent in a single API call by specifying the task, model, tools, and environment"; I replaced "a single API call" with "one API call", dropped the hyphenated compound "production-ready" because the text-to-speech engine stutters on hyphenated compounds, replaced "specifying" with "naming", and removed the serial commas to reduce narration pauses. Scene one sentence two, "OpenAI hosts the harness", is supported verbatim in substance by "OpenAI hosts and maintains the harness"; I dropped "and maintains" purely for the word budget and it removes no meaning that the script later relies on. Scene two sentence one and two, "You choose where agents run. An OpenAI managed sandbox your own infrastructure or a sandbox partner", are supported by "You choose the agent's compute environment: in an OpenAI-managed sandbox, on your own infrastructure, or with one of our sandbox partners" and by "The Agents API lets you choose a sandbox that fits your application"; I removed the hyphen from "OpenAI-managed" for the text-to-speech engine, compressed "with one of our sandbox partners" to "a sandbox partner", and split the list into a second sentence to avoid commas. Scene two sentence three, "Sessions can keep working for days", is supported by "They also need infrastructure that keeps them running reliably for days" together with "Today, we're introducing the Agents API in public beta, bringing that same harness and infrastructure that powers Codex to developers" and by "To support models working for hours, we've built context management that helps agents carry relevant information across longer sessions"; I hedged with "can" rather than promising duration, and I deliberately did not repeat the article's word "reliably" because that is a vendor performance claim I cannot verify. Scene three sentence one, "Agents can delegate parts of a task to subagents working in parallel", is supported by "the Agents API can break complex tasks into independent pieces and delegate them to subagents that work in parallel"; I wrote "parts" instead of "independent pieces" and used the article's own noun "subagents" rather than the hyphenated form "multi-agent", which the article also uses, because hyphenated compounds are barred from narration. Scene three sentence two, "One customer reports a sixty percent cost reduction per case", is supported by the OpenAI-published testimonial of Bhavyansh Sabharwal, Member of Technical Staff at SafetyKit, "After migrating our case review workflow to the Agents API, we saw a 60% reduction in cost per case, lower latency, and significantly improved token efficiency while maintaining existing performance"; I wrote the figure as words, "sixty percent", because digit-and-symbol tokens are garbled by the text-to-speech engine, and I attributed it in the narration as "One customer reports" so no viewer can mistake it for an independent measurement. Scene four sentence one, "This is OpenAI announcing its own product", is supported by the article's own byline and structure, author "OpenAI", and by the first person vendor voice throughout, for example "we're introducing the Agents API" and "OpenAI hosts and maintains the harness". Scene four sentence two, "The customer numbers are testimonials OpenAI selected with no independent testing", is supported by the article's section heading "What our customers are saying about Agents API" and by the seven named quotes collected under it, which are the only source of every performance figure on the page. Scene four sentence three, "You pay for tokens and tools used", is supported by "There are no additional fees for using the Agents API - you simply pay for the tokens and tools your agents use, as outlined on our pricing page"; I did not restate or extrapolate any price level, because the article gives none. The endcard narration is Inboxsmith promotional copy and makes no claim about OpenAI. HONESTY CAVEAT: this story is a vendor announcement in which OpenAI announces OpenAI's own commercial product, so every capability statement in the script is the vendor describing itself; the seven quoted performance figures on the page, including the evaluation score moving from 0.71 to 0.85, the four times latency reduction, the sixty percent cost reduction per case, and the eighty six percent reduction in failed agent responses, are testimonials that OpenAI selected and published from named customers, with no independent verification, no published benchmark, no methodology, and no third-party testing of any kind, and the script states this explicitly in scene four rather than burying it, and uses only one of those figures with a spoken attribution. CORROBORATION STATUS: zero of five YouTube transcripts were captured for this story because the local yt-dlp build is out of date and failed to retrieve any of them, so there is no third-party corroboration whatsoever; timeline probes surfaced only dated titles for related but different earlier OpenAI products, namely "Intro to Agent Builder" from October 6 2025, "Build Hour: Responses API" from October 14 2025, and "Build Hour: Agents SDK" from May 28 2026, and surfaced no earlier Agents API launch, which confirms this hosted API is genuinely new today rather than a re-promotion of an older feature, and this first-party article is therefore the only source behind every sentence of the narration. DELIBERATE OMISSIONS: I stripped the article's figurative and promotional language rather than carrying it into narration, including the customer phrase "a huge out-of-the-box lift", the phrase "fan out work across hundreds of agents", "stand up agents in hours", "connect them to the physical world", "mission-critical logistics operations", "first-class integrations", "strong foundation", "powerful harness", and "production-ready", and for time I dropped six of the seven customer quotes entirely along with all their named speakers and companies, the remaining three headline testimonial numbers, the full ecosystem partner list of Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel, both JSON code samples and every identifier in them including the model name string, the vault and capability directory fields, the OpenAI hosted sandbox section and its file, package, skill and plugin configuration, the versioned harness access point about new model capabilities at each model launch, the automatic context compaction detail for sessions approaching the context limit, the tool search and programmatic tool calling details including parallel calls and cache preservation, the support for MCP and custom functions and built-in tools like web search, the open source Codex harness and public codebase point, the VPC and CPU, GPU and memory configuration options, and the closing invitation to send feedback during the beta ahead of general availability. PROMPT INJECTION CHECK: the captured page contained a large amount of non-article interface text around the body copy, including "Skip to main content", the top navigation items "Research", "Products", "Business", "Developers", "Company" and "Foundation", the repeated marker "(opens in a new window)", "Log in", "Login", "Try ChatGPT", "Loading...", "Share", the duplicated in-page section link list, the "1 of 7" carousel counter, the trailing customer logo names, the "Keep reading" and "View all" blocks with three adjacent September 10 2026 headlines, and a long site footer covering research, safety, product, business, developer, company, support and terms links plus "OpenAI (c) 2015-2026", "Your privacy choices", "English" and "United States"; all of these are ordinary navigation, chrome and boilerplate strings, none of them was phrased as an instruction to an agent or reader, none attempted to make me run a command, edit a file, change configuration or ignore my rules, no instruction-like content of any kind was present anywhere in the article body or the customer quotes, and nothing in the captured text was treated as an instruction or followed.(OpenAI's own first party announcement, "Introducing the Agents API", published on openai.com on September 10, 2026 under Product / API)