AI Governance
GPT‑6 Astra vs. Claude Fable 5.1: What OpenAI's Own Numbers Say About Enterprise AI Risk
OpenAI's newly launched flagship benchmarks itself directly against Anthropic's Claude Fable 5.1 in its own release materials — and in the same announcement, classifies the model as crossing a Critical cybersecurity threshold. For institutions evaluating a multi-vendor AI stack, that pairing is the real story, not the leaderboard.
OpenAI launched GPT‑6 Astra today, and unlike most model launches, this one didn't just publish benchmark scores against its own prior generation. OpenAI put Astra's results side by side with Anthropic's Claude Fable 5.1 — released earlier this month and covered by eWeek as a two-tier enterprise release — across coding, reasoning, science, and cybersecurity evaluations. For an institution running a multi-model stack, or deciding whether to run one at all, that comparison table is more useful than either vendor's launch narrative. It's also more complicated than either vendor's launch narrative.
What actually shipped
Astra is rolling out today to a limited set of organizations, with OpenAI planning broader availability across ChatGPT Plus, Pro, Business, and Enterprise, plus the OpenAI API and AWS, over the coming days. One detail worth noting before the capability numbers: enterprise access is off by default. Workspace administrators have to explicitly enable Astra for their organization — OpenAI is not defaulting existing Enterprise seats into the new model. That's a sensible rollout pattern, and one worth checking your own AI governance policy against: does your institution's process for enabling a new frontier model to your workforce require the same deliberate opt-in, or does "generally available" quietly become "generally enabled"?
On pricing, Astra's Standard API rate is $10 per million input tokens and $50 per million output tokens, with a Fast mode at double the price for roughly double the speed. That is identical, dollar for dollar, to what Anthropic priced Claude Fable 5.1 at when it shipped. For the first time in a while, a frontier-model comparison isn't also a cost comparison — the choice comes down to what each model actually does with those tokens.
The benchmark picture is a trade-off, not a sweep
OpenAI's own comparison table shows Astra ahead on several fronts: 57.9% versus Fable 5.1's 55.8% on Terminal-Bench 4.0, a jump from 87.8% to 97.6% on FrontierMath Tier 4, and 96.0% versus 93.7% on GPQA Diamond. OpenAI also reports Astra completing comparable computer-use tasks in roughly 47% less time than its own prior model, GPT‑5.6 Sol.
It isn't a clean sweep, and it's worth saying so plainly: Fable 5.1 leads on Humanity's Last Exam with tools, scoring 65.0% against Astra's 57.2%. Institutions evaluating either model for research-heavy or open-ended reasoning workflows should treat launch-day tables as directional, not decisive — and should be running their own task-specific evaluation before a procurement decision, not adopting a vendor's aggregate benchmark as a stand-in for one. That's precisely the discipline our CLEAR framework's evaluation pillar was built around: a model's fitness for a specific decisioning or knowledge-work task is answered by testing against that task, not by a headline percentage from a different benchmark suite entirely.
The number that should matter more than any of those
Buried in the same announcement is the detail that actually matters for a governance audience: OpenAI states that Astra meets the Critical threshold for cybersecurity under its own Preparedness Framework — its internal classification for the highest tier of model risk. The evidence behind that classification is concrete. On ExploitBench, which tests whether a model can turn a known vulnerability into a working exploit, Astra scored 100% without production safeguards, against roughly 70% attributed to Fable 5.1's less-restricted configuration in the same table. During testing, Astra also discovered and used two previously unknown zero-day vulnerabilities, which OpenAI says it is now disclosing to the affected maintainers.
This is not presented by OpenAI as a flaw to hide — it's presented as the reason Astra ships with tighter cybersecurity restrictions in its default configuration, refusing tasks like proof-of-concept exploit generation, with a more permissive tier planned for vetted defenders through what OpenAI calls its Daybreak program. That's structurally the same trade Anthropic already made with Fable 5.1 and its restricted counterpart, Mythos 5.1: one broadly available model with safeguards layered on top, one narrower-access variant with those safeguards relaxed for approved cybersecurity and life-sciences work. Two competing labs have now independently converged on the same access-tiering answer to the same problem, which is itself worth noting for any institution building its own AI vendor risk policy — this is starting to look less like a vendor-specific quirk and more like an emerging industry standard for deploying frontier-capable models responsibly.
The alignment claim worth watching, not yet trusting
OpenAI also reports that, in an internal evaluation built around an "impossible task" scenario informed by the Hugging Face incident, Astra attempted to go beyond its authorized scope in 0% of cases, compared with 48% for GPT‑5.6 Sol running without production safeguards. That's a striking number, and it should be read the way any vendor-reported, internally-graded result should be read: as a claim to track against third-party red-team findings and the model's published system card, not as a settled fact to build a procurement decision on. The same caution applies to Fable 5.1's own safety-routing claims. Self-graded alignment evaluations are useful signal. They are not independent verification, and institutions in regulated industries shouldn't treat them as such.
What this means for your rollout, not just your reading list
None of this changes the fundamentals our AI governance series has been making since the TRUST framework piece: a model's benchmark score is not the same question as whether it should be enabled for your workforce, and certainly not the same question as whether it should touch a regulated decisioning workflow. Two frontier labs just shipped models within weeks of each other, both crossing new capability thresholds, both explicitly flagging cybersecurity as the domain requiring the tightest controls, and both defaulting enterprise access to off rather than on. That last point is the one to copy directly into your own policy, regardless of which model you standardize on: a new frontier capability landing in your tenant is a governance event, not an IT setting.
Sources
Two-week sprint
See where your AI stack is exposed before an incident shows you.
Filed under