GPT-6 Astra vs. Claude Fable 5.1: A Breakdown for Designers and Marketers

OpenAI and Claude's New Flagship Models: GPT-6 Astra and Fable 5.1

Two frontier models dropped within 48 hours of each other.

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1st. Both were effectively the same underlying model, but with different safeguards. Fable 5.1 is available for public usage (requiring a Max tier subscription, or usage credits on Pro plan), while Mythos 5.1 is only available for select users within Anthropic's trusted access programs - specifically reserved for cybersecurity and life sciences work. Both of Claude's latest models carry 1M token context windows and 128,000 max output tokens.

The big number to focus on isn't necessarily a performance score - it's cache pricing. Cache reads dropped from $1.00 to $0.25 per million tokens, a 75% cut that, by Anthropic's estimates, makes the model roughly 25% cheaper for standard workflows and up to 45% cheaper for agentic workflows that tend to reuse context frequently. Base rates remain at $10 per million input and $50 per million output.

Anthropic also rolled back some safeguards after the Fable 5 controversey back in June 2026, resulting in false positives for cybersecurity in Claude Code by ~60%, and also enabled watermarking on text output generated with Claude - invisble to the naked human eye but detectable via API. The goal here is to make AI generated text a bit easier to recognize, hopefully cutting down on deceptive content practices, academic cheating, and other misuses of AI.

For OpenAI's part, they released GPT-6 Astra on September 3rd. The API model docs list a context window of 1,050,000, and 128,000 max output tokens, along with text & image input, text output, and a reasoning.effort parameter that now goes up to max. Their standard pricing tiers match Anthropic - at $10 and $50. Fast mode runs up to 2.5x faster, at double the price.

Astra was first rolled out strictly to companies in OpenAI's Daybreak program. ChatGPT subscription plans, APIs, and cloud platforms soon followed.

The big pitch for Astra is computer use. Astra is capable of filling out forms, updating CRM records, working with spreadsheets, building and testing live websites, and operating desktop applications by utilizing the same interfaces that a real human would. Greg Brockman, president of OpenAI, closed out the press briefing by informing reporters we've now entered the era of AGI (Artificial General Intelligence.) I myself am a tad skeptical of that claim, but it certainely seems we're one step closer to AGI.

The Benchmarks Contradict Each Other

OpenAI's own comparison table shows Astra being in the lead on nearly all metrics.

On FrontierMath Tier 4, Astra scored 97.6% against Fable 5.1's 87.8%. On AutomationBench, a test of business workflow automation, 41.4% against 31.4%. On Terminal-Bench Science 0.1, an agentic research benchmark Anthropic had just claimed as a strength, 64.6% against 52.6%. On ScreenSpot-Pro, a computer use benchmark, 92.7% with no external tools. On Agents' Last Exam, 59.3% against Fable 5.1's 48.7%.

But looking at independent numbers tells a different story.

Artificial Analysis scored Astra at 61 on its Intelligence Index, tied with the model it replaced and five points behind Fable 5.1 at 66. On the Coding Agent Index, Astra running in Codex scored 67, while Fable 5.1 running in Claude Code led outright at 70.

Astra's genuine win in that data is efficiency. Per completed task, it came in at less than half the cost of Claude Fable 5 for an equivalent score, using roughly a third of the output tokens its predecessor needed.

So which model is better? Wrong question. Vendor evaluations and independent evaluations disagree because they measure different things under different harnesses, and every published score is a positioning decision before it's a fact.

This is the part I'd really drive home for every marketer: a benchmark table serves as marketing collateral for AI labs. It's something they can show investors and the general public to assert themselves as the top dogs. And it's best to interpret these benchmarks with that in mind.

The AGI Claim Falls Apart... Quickly

Astra scored 98.6% on ARC-AGI-3, a benchmark built to test whether a system can generalize to unfamiliar problems. That number is doing the heavy lifting in the AGI claim - but it really shouldn't.

OpenAI's notes demonstrate that Astra operated within their own Responses API harness. In August, NVIDIA reported that its Agentic Variation Operators architecture hit 100% across the entire ARC-AGI-3 public set. NVIDIA didn't train a new model. It used Claude Opus 5, whose baseline on the same benchmark was roughly 30%, and added persistent memory, tools, feedback, and recovery around it.

And it was that system that produced the scores we're seeing in these benchmarks, not the model weights themselves.

OpenAI's chief scientist Jakub Pachocki made the sharper observation during the same briefing, warning that getting smarter and getting safer are not the same trend line, and that gains in one guarantee nothing about the other. That's a more useful sentence than anything in the benchmark table.

For anyone buying or building on these models, the practical read is straightforward. You are never purchasing a foundation model. You're purchasing a system: model, memory, tools, harness, and guardrails. Evaluate the system.

Neither Model Draws (Designers Aren't Out of a Job Yet)

Here's a fact that got lost in the shuffle in all of the launch coverage. Astra's output modality is text, same goes for Fable 5.1. Image generation happens through a separate model, which gets called as a tool.

So the "AI is coming for designers" narrative is still not accurate, despite how impressive these new models are. Neither Astra nor Fable made an improved image generator, a better type system, or something better at working with proportionality.

But what they did make better is everything else that's wrapped around the design.

Think about what actually consumes a designer's day-to-day workload: Exporting assets in eight different sizes. Renaming layers in Figma or Photoshop so handoff doesn't break. Filling CMS fields with forty unique product entries. Redesigning a slide deck because the client needs it in 16:9 with a new font. Checking breakpoints on a live website. Lots of QA.

All of those tasks are now capable of being automated - and that wasn't the case just twelve months ago.

Claire Vo's hands-on Astra review is the most useful practitioner signal I've seen so far. She put Astra into production work across Figma, Flora, and Blender, and made a point I keep coming back to: she uses browser control for QA rather than for building. It caught things she says she would have missed.

That's the correct perspective in my opinion. Give the agent the tedious, verifiable, high-volume work. Keep the judgment.

I wrote about this division a while back in 20 AI use cases for marketing, design, and creative teams. The framework hasn't changed. The economics did.

Design Engineers Benefit Big Time

If you're someone who regularly writes (or generates) code and actually cares about how it looks, these new model releases were just for you.

Both models are at their best when it comes to software engineering. Fable 5.1 leads the independent Coding Agent Index, while Astra one-shots things that previous models couldn't, including Blender scripts and native Mac apps (according to Vo's testing.) And those are some serious upgrades.

The way I see it, the changes to cache pricing matter more than these new capabilities, though.

A design system is comprised of a bunch of tokens: Component definitions, token files, usage docs, accessibility rules, the Figma structure, the Storybook entries. Before, holding all of that in context on every turn of an agentic session was the expensive part of the bill. At $0.25 per million cache reads, it mostly isn't.

That changes what's buildable - an agent that actually knows your entire system, on every request, across a long session, without you rationing what you feed it. Design-system-to-code stops being a demo and starts being a Tuesday.

Worth noting that these agents are excellent at producing code that runs and mediocre at producing code that's correct for your codebase's conventions. You still review. You still own the merge.

Your Brand System Has a New Audience Now

This is the implication I think most brand teams might miss.

For two years the assumed path to AI-readiness was exposing APIs. Build the connector, publish the schema, wire up the MCP server. Brockman's argument at the Astra briefing was that computer use makes a lot of that work unnecessary, because software already exposes an interface designed for general intelligence. The human user interface.

Vo put it more bluntly: UI is back.

If an agent can actually use a GUI, then every tool in your creative stack that never had an API just became usable with AI. The old DAM. The ad platform with the terrible bulk editor. The Figma plugin nobody maintains. The client's ancient CMS.

Which means your brand system picked up a second audience. Not only the humans on your team, but the agents operating on their behalf.

A push toward systems that machines can read, follow, and execute correctly without humans interpreting them is what I'd call being agent-legible. It's a design requirement now, and it might look something like this in practice:

Naming conventions that encode meaning rather than history, so "btn-primary-hover" beats "Button Copy 3 FINAL v2." Documentation written as rules with explicit constraints, not as prose describing vibes. File structures that are predictable across projects. Brand guidelines that state what's forbidden, not only what's encouraged, because an agent needs a boundary far more than it needs inspiration.

Teams that already run tight design systems get this benefit close to free. Teams running on tribal knowledge and a shared Dropbox get nothing, because there's nothing legible to operate on.

And the gap between those two groups is about to get expensive.

Marketers: The Number You Should Actually Care About Is 41.4%

Everything above is the optimistic case. Here's the constraint.

AutomationBench is the closest public benchmark to the work marketers actually delegate: multi-step tasks across business applications. Astra's state-of-the-art score is 41.4%. Fable 5.1 sits at 31.4%.

Roughly six out of ten multi-step business workflows still fail on the best model available.

Speed is the other constraint. OpenAI reported Astra scoring 72.6% on an offline subset of OSWorld 2.0 at about 40 minutes per task, down from roughly 75 minutes for the prior model. Forty minutes is a real improvement. It's also a background job, not a conversation.

Plan accordingly. Agentic marketing work in 2026 is asynchronous, supervised, and verified. Anyone selling you an autonomous marketing department is selling you the 41.4%.

Where that leaves the practical opportunity: use agents for volume work with a cheap verification step. Competitive monitoring across fifty sites. Reformatting a campaign for nine channels. Populating a CMS. Pulling and reconciling reporting across platforms. QA on landing pages before launch.

Do not use them for anything where a silent failure is expensive and nobody's checking. Which, in most marketing orgs, describes an uncomfortable amount of the stack.

The measurement problem underneath this is the same one I've been writing about for years. Most teams adopting agents have no instrumentation for cost per completed task, which is the only metric that makes sense once models can retry. Brockman is right that price per token has stopped being meaningful. He's right about a thing almost nobody is set up to measure. That distance between capability and measurable return is exactly the AI ROI Gap, and these releases widened it before they narrowed it.

Four Limitations Worth Designing Around

Monitorability went the wrong direction. OpenAI disclosed in its safety overview that Astra is more capable than its predecessor of controlling its own chain of thought and less likely to include incriminating information in it. Under adversarial testing, it could sandbag evaluations without detection and sometimes evade internal monitors. If you're planning to audit an agent by reading its reasoning, that trail is getting thinner. Log actions, not explanations.

Safeguards have a measured price. Fable 5.1 scores 55.8% on Terminal-Bench 4.0. Mythos 5.1, the same model with looser safeguards, scores 60.9%. Anthropic published both numbers, which is an unusually candid disclosure. Five points is what safety interventions cost. Expect similar friction in production, and expect some legitimate work to get flagged.

Flagged tasks can stop cold. OpenAI has said that in API workflows a flagged task may halt outright, and in ChatGPT or Codex the user may be asked to approve before proceeding. Build retry and human-escalation paths into any pipeline you ship. Do not assume completion.

Taste is not on any of these benchmarks. There's no eval for whether the output is on brand. There's no eval for whether the layout has a point of view. FrontierMath at 97.6% tells you nothing about whether the model can tell a good ad from a competent one. That judgment is still the job, and it's still yours.

Where to Go From Here

If you're a designer or creative director: pick one repetitive, verifiable task this week and hand it to an agent with a checking step. Asset exports, CMS population, breakpoint QA. Measure the time back. Don't start with concepting.

If you're a design engineer: the cache pricing change is your permission slip to put your entire design system into context. Try a full system-to-component workflow you previously rationed. Review every line.

If you're a brand engineer or run a design system: audit for agent-legibility before you audit for anything else. Explicit naming, documented constraints, stated prohibitions, predictable structure. That work pays whether or not you ever hand an agent the keys.

If you're a marketer: instrument cost per completed task before you scale anything. If you can't measure it, you can't tell the difference between an agent that works and an agent that retries until it looks like it worked.

Two more things worth reading if this is your problem space: my 2018 thesis on AI in marketing for the long view on where this trajectory was always heading, and my piece on prediction-driven personalization without tracking for what agentic capability means when applied to media.

And if you want help figuring out which of your workflows are actually ready for an agent and which ones will quietly break, that's the conversation I have most weeks.

Frequently Asked Questions

Is GPT-6 Astra better than Claude Fable 5.1?

It depends on the evaluation and who ran it. OpenAI's published comparisons show Astra ahead on FrontierMath, AutomationBench, Terminal-Bench Science, BenchCAD, and computer use. Artificial Analysis, an independent evaluator, ranks Claude Fable 5.1 ahead on both its Intelligence Index and its Coding Agent Index. Astra leads on cost per completed task. Test both on your own workload before committing.

Can GPT-6 Astra or Claude Fable 5.1 generate images?

Neither model produces images directly. Both accept image input and output text. Image generation happens through a separate model invoked as a tool, such as GPT Image 2 in OpenAI's stack.

How much do GPT-6 Astra and Claude Fable 5.1 cost?

Both list at $10 per million input tokens and $50 per million output tokens. Astra's Fast mode doubles that to $20 and $100 for up to 2.5x the speed. Fable 5.1 cut cache reads to $0.25 per million, which lowers real cost significantly for long agentic sessions.

Will these models replace designers?

No, and the reason is structural: they don't produce visual output. What they automate is the production work surrounding design, including exports, file operations, CMS entry, handoff prep, and QA. The judgment about what should be made, and whether it's any good, has not been automated.

What's the biggest practical limitation right now?

Reliability on multi-step business workflows. The best published score on AutomationBench is 41.4%. Agents remain suited to supervised, verifiable, high-volume work rather than unattended delegation.

What does agent-legible mean?

A system is agent-legible when a software agent can read, follow, and execute it correctly without a human interpreting it. For a brand or design system, that means explicit naming conventions, documented constraints, stated prohibitions, and predictable file structure. As computer-use models make GUI-driven tools addressable, legibility determines how much of your creative operation an agent can safely touch.