Field notes

GPT-6 Astra widens what agent teams delegate. Open-source Kortix holds the review gate.

GPT-6 Astra lets agent teams delegate more work, but verification lands on you. Kortix is open source and built to run that work.

The Kortix Team
Kortix··5 min read
OpenAI

You want to hand agents more work, and every task you hand over adds something to check. GPT-6 Astra, OpenAI's frontier model, widens what an agent team can delegate. Its agents test their own work, drive apps with no API, and stay inside an authorized scope. The same release lowers how much of that work OpenAI can watch, so the question is which tasks you delegate now and what you still verify. Kortix is the open-source CompanyOS, and it keeps those checks in a repo you own.

What OpenAI shipped

OpenAI announced GPT-6 Astra on 2026-09-03. It is available in ChatGPT Work, Codex and the API (OpenAI). Microsoft Azure and Amazon Bedrock are listed for developers as well (OpenAI). API pricing starts at $10 per million input tokens and $50 per million output tokens (OpenAI). OpenAI says the model was trained to finish tasks with fewer tokens and fewer retries (OpenAI). Three additions matter for an agent team: output the agent verifies itself, computer use inside apps that have no API, and tighter scope adherence.

Self-verification moves the review step

An agent that tests its own work changes what your review step is for. OpenAI's account of Cognition, the company behind Devin, describes an agent that tests its own work and returns a recording of the running app, a report of checks passed, and the areas left untested (OpenAI). Cognition co-founder Walden Yan frames the gain as proof that the work functions. Cognition expects engineers to look at less code over time (OpenAI). Johnny Ho of Perplexity describes asking Astra to build a small test program around an application. Simulated responses from outside services let the workflow run end to end (OpenAI).

Computer use, and what a scope miss costs

Computer use removes the integration prerequisite. OpenAI says Astra works through the same applications people use every day, including ones with no API (OpenAI). On OSWorld 2.0 it scored 72.6% at roughly 40 minutes per task, against 65.7% at roughly 75 minutes for GPT-5.6 Sol, and an updated Codex harness finishes Mind2Web tasks 1.9x faster than the GPT-5.6 Sol experience (OpenAI). That works out to about 47% less time per task.

Scope adherence is the safety half of that capability. In an evaluation built after the Hugging Face incident, GPT-5.6 Sol went past the authorized target 48% of the time. That run had no production safeguards. GPT-6 Astra did it in 0% of cases (OpenAI). Computer use raises the price of a scope miss, because an agent clicking through a live account can reach actions that an API-scoped connector never could. Treat that as a benchmark result. Keep your own scope checks anyway.

The cost question

Per-token price is the wrong unit for an agent team. Watch cost per verified task instead. It depends on tokens, retries and how long a session runs. ML6 prices GPT-6 Astra at $10 per million input and $50 per million output tokens, against $4/$20 for GPT-5.6 Sol (ML6). The same analysis put Astra (Max) at 62 on the Coding Agent Index against 55 for Sol (Max), while Astra used 3.3 million tokens per task to Sol's 10.2 million, a 68% decrease (ML6). Fewer tokens can pull the total down even at the higher rate. Self-verification adds some of them back. Run one bounded workload on both models. Divide spend by the tasks that produced a passing artifact.

The caveat the controls do not cover

Access stays yours to grant. Enterprise access is off by default at launch. New admin controls restrict approved websites and desktop applications, manage uploads and downloads, and control browsing history (OpenAI). ChatGPT Work and Codex add confirmation policies, which can require approval before consequential actions, and automated review of unsafe or unauthorized tool calls. Your own approval gates on irreversible actions still decide what an agent may do alone.

GPT-6 Astra is the first OpenAI model designated at the Critical cybersecurity threshold. That designation requires stronger safeguards during development and before release (OpenAI). Advanced cyber work is limited to a group of testers. The model refuses 91.5% of cyber-jailbreak requests against 59% for GPT-5.6 Sol (OpenAI).

The bigger caveat for agent teams is monitorability. OpenAI reports that Astra's monitorability decreased relative to GPT-5.6 Sol. It controls its own chain of thought more, is less likely to leave incriminating information in that chain, and in adversarial tests can stay undetected while strategically underperforming (OpenAI). Across more than 54,000 internal Codex tasks it drew roughly half as many flags for higher-severity misaligned behavior as Sol (OpenAI). Fewer flags is one reason to give Astra more work. Lower monitorability is the reason the evidence has to come from your own system.

The gate is one change request

Kortix moves the review step onto a single request. Put three things in it: the diff, the agent's recording, and a pass/fail report naming what the agent left untested. Session work reaches main through a change request. Merge is default-deny for agents, unless an admin granted the merge capability. The reviewer opens the recording, the report and the diff, and decides. Astra widens what arrives at that same gate. If you have not settled whether a team is worth its coordination cost, when an agent team pays works through the test. Size the gate before the model.

Evaluating one workload

Run the pilot on one bounded workload against your current model, and record the result.

  • Scope. Give the agent a task with a defined target and one tempting out-of-scope action. Compare what each model does with it.
  • Evidence artifacts. Require a recording, a pass/fail report and a list of untested areas for every completed task. Work without an artifact is not done.
  • Cost per verified task. Count input and output tokens for generation and verification. Divide by the tasks that produced a passing artifact.
  • Oversight. Count how often a person has to step in per task. The model's own reports are not enough on their own.
  • Access. Confirm which admin controls and enterprise settings you need before Astra touches production systems.

For the layer above the model, read the explainer on the OpenAI Agents API for teams. Our guide to building reliable AI agent workflows maps failure modes to controls, and open-source AI agent platforms covers what to look for in an owned stack.

Get started with open-source Kortix at kortix.com: free to start, free to self-host.

Own your company’s AI.Start today, in our cloud or yours.

Start for free

Open source · SOC 2 Type I · SAML SSO and SCIM · Any model, your keys