OpenAI DevDay 2026: Agents API Adds Computer Use, GPT-6.1 Sol Arrives at One-Fifth of Astra's Price
At DevDay on 2026-09-29, OpenAI shipped GPT-6.1 Sol, an Ultrafast speed tier, and computer use in the Agents API, moving hosted agent execution to the center of its developer platform.
OpenAI DevDay 2026 (September 29, 2026) centered on the Agents API and GPT-6.1 Sol: the Agents API has OpenAI host the agent loop, sandboxes, and memory and now adds computer use, while GPT-6.1 Sol offers near-GPT-6 Astra capability at $2 input and $10 output per million tokens; it signals OpenAI moving from selling models to selling a hosted agent execution layer.
According to OpenAI's official Developer Community DevDay 2026 roundup and the API Changelog, 2026-09-29 brought GPT-6.1 Sol, computer use in the Agents API, and Ultrafast mode for GPT-6 Astra. The DevDay summary says Ultrafast delivers up to 8x faster token generation in Codex and up to 6x in the API. It also lists a Decisions API in limited preview that uses Luna for classification and request routing, and Private Intelligence features combining Zero Data Retention with Private Safety Processing.
The Agents API itself entered public beta on 2026-09-10. OpenAI's announcement says it runs the agent loop on OpenAI infrastructure, coordinating model calls, tool use, and context operations, with sandboxes hosted by OpenAI, self-hosted, or provided by partners such as E2B, Modal, and DigitalOcean. There is no extra fee for the Agents API itself; customers pay for tokens, tools, and hosted container usage. The GPT-6.1 Sol docs list $2 input and $10 output per million tokens with a 1,050,000-token context. GPT-6 Sol and GPT-6 Luna were released on 2026-09-22, with GPT-6 Luna priced at $0.10 input and $0.50 output.
The direction is clear: beyond models, OpenAI wants to own the agent execution layer — session orchestration, context compaction, sandboxes, memory, and computer use all hosted by the platform. For developers that lowers the barrier; for enterprises it deepens platform lock-in, because agent state and execution history stay with the vendor.
Caveats: 'near-Astra' capability and the speed multiples are OpenAI's claims, and the Agents API is still in beta, so interfaces and billing may change. A notable detail: the GPT-6.1 Sol docs state that Chat Completions does not support tool calling for this model — tools require the Responses API — and GPT-6 Luna supports function calling in Chat Completions only when reasoning_effort is set to none. On the Codex side, OpenAI added cloud execution while devices are offline, two-way voice in the CLI, and cloud-based Code Review.
When evaluating the Agents API, enterprises can start with a few concrete questions: whether sandboxes should be OpenAI-hosted, self-hosted, or provided by partners such as E2B, Modal, or DigitalOcean; whether hosted container costs are budgeted, since a community moderator explicitly warned new users to calculate container costs first; when agent memories are created and where they are stored; and whether ZDR with Private Safety Processing covers their use case. On model choice, GPT-6 Luna costs $0.10 input and $0.50 output per million tokens versus $2 and $10 for GPT-6.1 Sol, a 20x gap that favors splitting high-volume steps such as classification and routing from complex coding work.
What to watch: when the Agents API reaches general availability, the scope and conditions of ZDR with Private Safety Processing, and whether the Decisions API makes small-model routing a built-in OpenAI platform feature that overlaps with the routing layer of enterprise-built gateways.
By turning agent execution into a hosted service, OpenAI makes 'where data and agent state live' a core selection question. Enterprises running their own private AI platform should check how well their gateway handles Responses API tool calling, and decide which agent workflows can run on on-prem open models such as gpt-oss in their own sandboxes and which genuinely need a cloud frontier model.
