Glonce

GPT-6 Astra cuts unintended actions 89% versus GPT-5.6 Sol

modelsThe Glonce editorsOpenAI🔥81Glonce rating81Impact82Novelty72Relevance100Surprise50Players98Credibility87
GPT-6 Astra cuts unintended actions 89% versus GPT-5.6 Sol

OpenAI says it introduced GPT-6 Astra, its most capable model, last week, and the system is now available in ChatGPT Work, Codex and the API. The company describes Astra as the world's most intelligent and aligned model to date and says it is state of the art on computer use, browsing, professional work, software engineering, cybersecurity and science. Unlike AI systems that OpenAI says require businesses to prepare their data, redesign workflows or build custom integrations before delivering value, it says Astra can write code and operate the same everyday applications people already use, even ones without an API, so companies can put it to work inside existing workflows from day one.

Within the first few days of rollout, OpenAI says customers were already using Astra for tasks ranging from optimizing GPUs to spotting discrepancies in financial statements to producing more on-brand slide decks. OpenAI itself rolled Astra out internally weeks before the public launch: its developer and marketing teams used Astra together with Codex to turn three hours of multicamera footage into a "GPT-6 Astra Developer First Impressions" video that drew more than 550,000 views within four days, and its engineering team used Astra to find and fix a memory-allocation bottleneck that was slowing Codex sessions in a test environment; switching memory allocators cut turn latency 25×, at the cost of roughly 30% higher peak memory use. OpenAI also says Astra is better at following a company's voice, templates and design standards, so its first output lands closer to something a team can put to use directly.

On cost, OpenAI says Astra was trained to complete tasks using fewer tokens and fewer retries, which the company says means less rework and a lower cost per task. It says that with Astra, OpenAI now holds the majority of the cost-efficiency frontier on professional-work and coding evaluations, including Terminal Bench 4.0 and the Artificial Analysis Intelligence Index, though the post cites no specific scores or rankings. API pricing starts at $10 per million input tokens and $50 per million output tokens.

On safety, OpenAI calls Astra its most aligned model yet, with stronger adherence to human intent and authorization. During training, the company tested Astra on an internal computer-use safety benchmark that checks models against difficult business scenarios such as exposing confidential information, sharing a dashboard too broadly, or deleting data. In that evaluation, OpenAI says Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1, and that additional confirmation steps and automated review improve results further for both GPT-6 Astra and GPT-5.6 Sol, though it gives no updated figures for that combined state.

To manage that access, OpenAI is adding enterprise admin controls that let organizations restrict Astra to approved websites and desktop applications, manage uploads and downloads, and control browsing history, so teams can start with a limited configuration and expand it over time. ChatGPT Work and Codex also ship with confirmation policies that can require approval before consequential actions, plus automated review of potentially unsafe or unauthorized tool calls. Alongside Astra, OpenAI is launching new ChatGPT Desktop plugins, built on its latest browser-use capabilities, for Oracle Analytics, Power BI (a Microsoft Fabric service), Navan and Avalara. Separately, OpenAI says Astra is the first model to reach the "Critical" cybersecurity capability threshold under its Preparedness Framework, and that it responded by training Astra to respect safety and security boundaries, improving its resistance to jailbreak attempts, and deploying automated checks meant to block harmful responses. Zero Data Retention is available to eligible API customers on supported endpoints, subject to approval, and enterprise access to Astra stays off by default at launch until an administrator enables it under the organization's applicable rate card and agreement.

Key facts

  • OpenAI made GPT-6 Astra, introduced the previous week, available in ChatGPT Work, Codex and the API, and says it can operate everyday business applications directly, even ones without an API, without prior integration work.
  • API pricing starts at $10 per million input tokens and $50 per million output tokens; OpenAI says Astra completes tasks in fewer tokens and retries and now holds the majority of its cost-efficiency frontier on evaluations including Terminal Bench 4.0 and the Artificial Analysis Intelligence Index.
  • On OpenAI's internal computer-use safety benchmark, covering scenarios like leaking confidential data or deleting files, Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1.
  • New enterprise admin controls can restrict Astra to approved sites and apps and manage uploads, downloads and browsing history; OpenAI says Astra is the first model to reach the "Critical" cybersecurity threshold under its Preparedness Framework, prompting added safeguards.
  • New ChatGPT Desktop plugins connect Astra to Oracle Analytics, Power BI, Navan and Avalara, and OpenAI's own engineers say switching memory allocators cut Codex turn latency 25× (at roughly 30% higher peak memory) in a test environment; enterprise access stays off by default until an administrator enables it.

Why it matters

GPT-6 Astra's pitch here is specifically for business use: OpenAI says it can operate inside a company's existing applications and workflows, even ones without an API, without custom integration work first. That claim, paired with concrete API pricing, a cost-efficiency claim tied to named benchmarks, and a safety comparison against both OpenAI's own prior model and a named rival, gives enterprise buyers specific numbers to weigh rather than general marketing language.

Who it affects

Enterprise IT and security administrators who will configure Astra's access controls and decide rollout scope; developers using Codex; business teams doing the kind of work OpenAI's own examples name, including GPU optimization, financial-statement review and deck production; and OpenAI's plugin partners Oracle Analytics, Power BI (a Microsoft Fabric service), Navan and Avalara, whose enterprise tools now connect to ChatGPT Desktop through Astra's browser-use capabilities.

How to use it

Astra is available now in ChatGPT Work, Codex and the API. API pricing starts at $10 per million input tokens and $50 per million output tokens. Enterprise access is off by default at launch; an administrator must enable it under the organization's applicable rate card and agreement, and can then use the new admin controls to restrict Astra to approved websites and desktop applications, manage uploads and downloads, and control browsing history, alongside confirmation policies that require approval before consequential actions. Zero Data Retention is available to eligible API customers on supported endpoints, subject to approval. The new ChatGPT Desktop plugins for Oracle Analytics, Power BI, Navan and Avalara offer another route in for teams that already use those tools.

How solid is it

Every figure here comes from OpenAI's own blog post and its own internal testing, with no independent benchmark cited. The safety comparison, 89% and 74.7% fewer unintended outcomes than GPT-5.6 Sol and Claude Fable 5.1, covers an internal computer-use safety benchmark that OpenAI describes only by three example scenario types, without stating how many scenarios or trials were run. The cost-efficiency claim names Terminal Bench 4.0 and the Artificial Analysis Intelligence Index but gives no actual scores. The memory-allocator fix that OpenAI says cut Codex turn latency 25×, at roughly 30% higher peak memory, is scoped to an unspecified internal test environment, with no baseline latency figures and no allocators named, so the absolute numbers behind the ratio are not known.

Risks and caveats

The safety and efficiency numbers are self-reported and self-graded by OpenAI, comparing its own models, and in one case a competitor's, on a benchmark it designed itself. Reaching the "Critical" cybersecurity capability threshold under OpenAI's own Preparedness Framework means Astra is capable enough in that domain that OpenAI says it strengthened protections against misuse and unauthorized action, including resistance to jailbreak attempts and automated checks to block harmful responses, but the post does not explain what specifically qualified the model for that threshold. Zero Data Retention is conditional, available only to eligible API customers on supported endpoints and subject to approval, rather than a default guarantee, and broad computer-use access to business systems is exactly why OpenAI pairs the launch with confirmation policies and admin controls rather than shipping it open by default.

“Astra produced unintended outcomes 89% less often than GPT-5.6 Sol and 74.7% less often than Claude Fable 5.1.”

— OpenAI