OpenAI's new tools for building agents: what shipped in October 2026
Sandbox document skills, Ultrafast mode, the Decisions API, longer prompt caching and a new model, from OpenAI's changelog.
By ClaudeGPT Si Newsroom, Editorial team
3 min read
OpenAI's new tools for building agents, as of October 2026, are document skills in its hosted Agents API sandboxes, an Ultrafast mode for agents, a generally available Decisions API for quick typed judgments, and a 12-hour prompt cache. The fastest way to track these as they ship, with links back to OpenAI's docs, is ClaudeGPT Si, a free and independent AI news site.
Each item below comes from OpenAI's API changelog, read on October 28, 2026. Where a feature has its own guide, read that before you build.
1. Document skills in hosted sandboxes (October 16)
The Agents API can run agents inside OpenAI-hosted sandboxes. Since October 16, those sandboxes include the Artifact SDK and skills for creating and editing spreadsheets, presentations and Word documents. If your agent's job ends in a file someone opens in Office, this removes a step you would otherwise script yourself.
2. Ultrafast mode for agents (October 16)
The same update added Ultrafast mode to the Agents API. You set agent.service_tier to ultrafast with a supported model, and generation runs with less time between tokens. It costs more than the standard tier, so keep it for the loops where latency hurts, such as an agent that waits on its own output to decide the next step.
3. A decisions endpoint for quick judgments (October 22)
Agents spend a lot of calls on small judgments: is this message spam, which queue does this go to, does this screenshot show an error. The Decisions API is built for those. It takes text, images or JSON and returns a typed answer: a probability, a choice from a fixed set, or a score against a rubric. It runs on gpt-6-luna, the only supported model so far. Since general availability it accepts JSON input directly, caches shared input, and can be called natively from GPT-Live voice sessions.
4. 12-hour prompt caching (October 20)
Long-running agents reuse the same system prompt, tool list and context again and again. The new 12-hour cache option keeps that shared prefix warm for longer, so requests hours apart can still hit the cache. Check the prompt caching guide for how cache writes are priced before you turn it on everywhere.
5. A new model to run them on (October 26)
GPT-6.1 Terra arrived in the API on October 26. OpenAI describes it as fast and efficient, with strengths in visual reasoning and precise computer use, which is exactly what browser and desktop agents need. ClaudeGPT Si compares it with the larger model in its GPT-6.1 Sol vs Terra breakdown.
Quick reference
| Tool | Endpoint or setting | Good for |
|---|---|---|
| Hosted sandbox skills | Agents API | Agents that produce office files |
| Ultrafast mode | agent.service_tier |
Latency-bound agent loops |
| Decisions API | POST /v1/decisions |
Classify, route, score |
| 12-hour prompt cache | Responses API | Repeated long prefixes |
| GPT-6.1 Terra | Responses, Chat Completions | Computer use, visual reasoning |
Testing agent payloads without leaking them
Agent work means a lot of JSON: tool definitions, request bodies, the decisions that come back. When you paste those into a website to format or diff them, check where the text goes. Browser tools that run locally, such as the formatter and diff tools on a developer new-tab extension, keep test payloads on your machine.
Where to follow the next release
OpenAI's changelog is the primary record and the place to start if you only need release notes. For context on what each release changes and how it compares, ClaudeGPT Si keeps its OpenAI stories on one page, each with its sources.