OpenAI says Astra beats Claude Fable 5.1 badly on its own safety benchmark

[FairDos96]

OpenAI published a detailed look at GPT-6 Astra's business capabilities, highlighting that it can work through everyday applications, spreadsheets, decks, dashboards, even ones without an API, without requiring businesses to first prepare data or build custom integrations. Customer testimonials cited in the release include Cognition integrating Astra into its Devin coding agent on launch day, Databricks reporting significantly better cost per task on its OfficeQA benchmarks, and Hebbia saying Astra followed client briefs 17 percent more faithfully than the next best model while sourcing claims to the right document 19 percent more often

On OpenAI's internal computer use safety benchmark, which tests scenarios like exposing confidential information or deleting data, Astra produced unintended outcomes 89 percent less often than its own predecessor GPT-5.6 Sol, and 74.7 percent less often than Anthropic's Claude Fable 5.1. Astra also set a new record of 74 percent on the DeepSWE v1.1 coding benchmark according to partner Datacurve, and reached 57.9 percent on Terminal-Bench 4.0 compared to 55.8 percent for Fable 5.1, at roughly 63 percent lower estimated API cost per task

New enterprise admin controls let organizations restrict Astra to approved websites and applications, manage uploads and downloads, and require approval before consequential actions, while enterprise access remains off by default at launch. OpenAI confirmed Astra is the first model to cross its Critical cybersecurity capability threshold under its Preparedness Framework, prompting strengthened protections against both misuse and unauthorized model behavior. Curious what people think about OpenAI publishing a direct safety comparison against a specific named competitor, does that kind of head-to-head safety claim carry real weight, or does it need independent verification before anyone should treat it as settled

Gpt 6 Astra Next Generation Workopenai.com

[FinnBalor]

A 74.7 percent gap against a specific named competitor on a safety benchmark is a genuinely bold claim to publish, worth remembering this is OpenAI's own benchmark design being applied to a rival's model rather than a neutral third party's test

[Mia86]

The 25x lower turn latency OpenAI's own engineering team found by switching allocators is a nice concrete internal example, shows real practical value beyond just the marketing-oriented customer testimonials featured elsewhere in this release

[Brandon_62]

Enterprise access being off by default despite all these capability claims is a sensible, cautious rollout choice, forces organizations to make a deliberate decision to enable it rather than getting swept in automatically

[Protocol]

Hebbia's specific numbers on brief-following and source accuracy are genuinely useful because they're concrete and falsifiable, that's a meaningfully more specific claim than most vague AI vendor testimonials usually offer

[Scarlett]

Curious how Anthropic responds to this specific safety benchmark comparison, a direct public claim naming a competitor's model this explicitly usually invites some kind of rebuttal or its own competing benchmark data eventually

[Cobalt Sophie]

Crossing the Critical cybersecurity threshold under OpenAI's own framework right as this business-focused release rolls out is worth sitting with, that's a serious internal safety classification landing at the same time as an aggressive commercial push