OpenAI's Agents API puts the Codex harness behind one API call
Public beta to every dev, billed only for tokens and compute. Also today: inclusionAI's free MIT multimodal model, and a freeze on new $200 Pro sign-ups.
Copy markdown
The Codex harness, now one API call
The same harness that runs Codex — context management, tool use, subagents — is in public beta for every developer. No extra fee: you pay only for the tokens, tools, and container time your agents burn, with gpt-6-astra shown in the sample code.
Sandboxes: hosted, self-hosted, or bring your own
Agents execute in an OpenAI-hosted sandbox by default, or you self-host with codex exec-server, or plug into nine partner runtimes — Modal, Cloudflare, E2B, Vercel, DigitalOcean, Daytona, Runloop, Blaxel, Oracle. Sandboxes run code, edit files, and return artifacts.
The plumbing you were hand-rolling, built in
You get automatic context compaction as a session nears its limit, tool search and programmatic tool calls to cut token spend, parallel subagent orchestration, and native MCP plus custom functions — the wiring most agent stacks reinvent.
The numbers OpenAI is quoting
Early customers report a 60% drop in cost per case, 4x lower latency, 86% fewer failed agent responses, and an eval score climbing from 0.71 to 0.85. Vendor figures, but that's the pitch.
The catch: US-only, no zero-retention
Data stays US-only and Zero Data Retention isn't supported yet, so EU-resident and regulated workloads can't move onto it today. Check that before wiring it into anything compliance-bound.
Build this weekend: a hosted fix-it agent
Point the Agents API at a repo inside a hosted sandbox and you have a cloud coding agent — triage an issue, branch, run tests, open a PR — with no harness, queue, or container fleet of your own to maintain.
Elsewhere: Ling 3.0 Flash VL goes free and open
inclusionAI released an MIT-licensed 124B MoE (5.5B active) that takes text, image, and video in, with a 260K context window at ~139 tok/s. Free on OpenRouter, weights on Hugging Face, and it lands #2 on Artificial Analysis' open-model index.
Elsewhere: OpenAI freezes new $200 Pro sign-ups
Astra demand pushed OpenAI to pause new Pro subscriptions on Sept 10; existing Pro accounts keep working, and Plus, Go, and the API stay open to new users. No end date given.