GPT-6 Astra vs Claude Fable 5.1: Benchmarks, Pricing, Coding

GPT-6 Astra and Claude Fable 5.1 score exactly the same on Artificial Analysis' Intelligence Index: 53 out of 100, dead even. Getting that score costs $3.26 per task with one model and $7.63 with the other.

That's the more useful version of "which model is smarter" — because on raw intelligence, as of September 2026, there genuinely isn't a clear winner. OpenAI's GPT-6 Astra (released September 3–4, 2026) and Anthropic's Claude Fable 5.1 (released September 1, 2026) landed within 72 hours of each other, and each vendor's own launch benchmarks show their own model winning almost everything — because each one tested it differently. The only fair tiebreaker is a neutral third party, and even that tiebreaker has moved twice in the two weeks since launch.

This post works through what the numbers actually say — benchmarks, real API pricing, coding performance, and the safety incidents that came with both launches — and flags exactly where the data is solid versus where it's still noisy.

What GPT-6 Astra and Claude Fable 5.1 Actually Are

Start with the names, because both get mangled online. GPT-6 Astra is OpenAI's model (API id gpt-6-astra); ChatGPT is just the app it runs inside, and the chat experience built specifically on Astra is branded "GPT-6 Pro." If you've seen this written as "ChatGPT 6 Astra," that's the common mix-up. "Astra" is genuinely OpenAI's own product name here, not a reference to Google DeepMind's older "Project Astra" — a coincidence, not a rebrand.

Claude Fable 5.1 is Anthropic's model (API id claude-fable-5-1), and "Fable" is Anthropic's top tier, sitting above Opus, Sonnet, and Haiku. A more permissive twin, Mythos 5.1, also exists, but it's restricted to vetted government and enterprise partners.

Both are agentic models — built to act across many steps on their own (navigating a computer, running terminal commands, editing a codebase) instead of just answering one question and stopping. Fable 5.1 launched first, on September 1; Astra followed on September 3–4. That 72-hour gap matters, because Anthropic's own launch benchmarks were run before Astra existed to compare against.

They also reason differently under the hood. Fable 5.1's "thinking" step is always on: it continuously decides how much reasoning a prompt needs, and while you can set an effort level from low to maximum, you can't switch reasoning off entirely — Anthropic keeps this readable, returning plain-text reasoning summaries between tool calls. Astra's reasoning effort is a dial you set explicitly, but the reasoning itself is encrypted; OpenAI prioritizes protecting the model's internal process over letting you inspect it directly.

Why Every Benchmark Table Tells a Different Story

OpenAI's launch table has Astra beating Fable 5.1 on nearly every row it published. Anthropic's launch table has Fable 5.1 beating GPT-5.6 Sol and Claude's own Opus 5 — Fable 5.1 was never benchmarked against Astra directly, because Astra didn't exist yet. Neither table is lying, exactly. They're measuring different opponents, on different task sets, with different harnesses.

The one evaluation that runs both models through an identical test is Artificial Analysis, an independent benchmark aggregator. As of its Intelligence Index v4.3 (published September 7, 2026), Astra and Fable 5.1 are tied at 53 out of 100, both at maximum reasoning effort. Its Coding Agent Index — Astra running inside OpenAI's Codex, Fable 5.1 running inside Anthropic's Claude Code — is also a tie, at 62. Treat both numbers as current rather than final: the Intelligence Index moved twice in the days right after Astra's launch (an earlier reading had Fable 5.1 ahead 66–61, the one after that had it 57–55), so if you see a different score quoted elsewhere, it's probably an older snapshot.

Where Each Model Actually Wins

Below that tied headline number, the two models are specialists in different things. Here's what holds up on OpenAI's own published launch table:

GPT-6 Astra wins:

• AutomationBench (multi-step business-workflow automation): Astra 41.4% vs Fable 5.1 31.4%

• Terminal-Bench 4.0: Astra 57.9% vs Fable 5.1 55.8%

• Terminal-Bench Science 0.1: Astra 64.6% vs Fable 5.1 52.6%

• FrontierMath Tier 4: Astra 97.6% vs Fable 5.1 87.8% (OpenAI calls this suite "saturated" at this score)

• GPQA Diamond (graduate-level science Q&A): Astra 96.0% vs Fable 5.1 93.7%

• DeepSWE v1.1 (coding): Astra 74.1% vs Fable 5.1 67.4%

Claude Fable 5.1 wins:

• Humanity's Last Exam, with tools: Fable 5.1 65.0% vs Astra 57.2% — the one clear reasoning win for Fable 5.1 on OpenAI's own scoreboard

• GDPval-AA v2, a knowledge-work benchmark: Fable 5.1 stays ahead here; third-party testing found Astra actually regressed by roughly 45 Elo points against its own predecessor, GPT-5.6 Sol, on this same test

Genuine ties, once you go to the neutral scorer:

• Artificial Analysis Intelligence Index: 53–53

• Artificial Analysis Coding Agent Index: 62–62

One number you won't find a clean answer for is OSWorld 2.0, a computer-use benchmark. OpenAI's table lists Astra at 72.6% and leaves Fable 5.1's directly comparable score blank, citing differences in the task release each model was run against. A separate comparison puts Fable 5.1 ahead here, at 77.9% — but since it isn't clear the two numbers were measured the same way, we're not calling a winner on this one.

Both models tie on intelligence. Only one of them ties on cost.

OpenAI's real edge shows up in benchmarks that involve acting on a live environment — a terminal, a browser, a chain of tool calls — plus raw math and science recall. Fable 5.1's edge shows up in benchmarks that reward deliberate reasoning across broad, unstructured domains of knowledge in one sitting.

Cost Per Task Beats Price Per Token

The sticker price is identical: both charge $10 per million input tokens and $50 per million output tokens through their APIs. Cached tokens are where the rate cards genuinely diverge, and on paper, in Anthropic's favor — Fable 5.1's cache reads cost $0.25 per million tokens, a 75% cut from Fable 5, while Astra's cache reads cost four times as much, at $1.00 per million.

But cache pricing isn't the whole bill, and it isn't even the deciding factor. At maximum reasoning effort, Artificial Analysis measured Astra using roughly a third of the output tokens Fable 5.1 needs to hit the same Intelligence Index score. Because output tokens are priced five times higher than input tokens on both models, that gap dominates everything else:

// same list price for both: $10 / 1M input tokens, $50 / 1M output tokens
// Artificial Analysis Intelligence Index v4.3, weighted-average cost per task:
// GPT-6 Astra        ~27,000 output tokens for a max-effort answer  -> $3.26
// Claude Fable 5.1    ~78,000 output tokens for the same score      -> $7.63
// same intelligence score (53) -- about 3x the tokens, 2.3x the bill

That's a 57% cost advantage per completed task for Astra, on identical sticker pricing. Astra gives some of that back on very large single prompts: its rate doubles above 272,000 input tokens (2x on input and cache, 1.5x on output), a surcharge Fable 5.1 doesn't carry. For genuinely huge documents — a full codebase, a book-length PDF — Fable 5.1's flat pricing can still end up cheaper, cache advantage or not.

Coding: Claude Code vs. Codex

On raw coding benchmarks, Astra edges ahead: Terminal-Bench 4.0 (57.9% vs 55.8%) and DeepSWE v1.1 (74.1% vs 67.4%) both favor it narrowly. Neither lab has published a SWE-bench Verified score for its current flagship — both moved to newer benchmark suites they say are less contaminated by training data — so there's no clean, like-for-like coding number beyond these two.

The bigger difference is in the tools built around each model. Anthropic's Claude Code is local-first and supervised: it asks before acting, and it's built for deep customization through hooks and skills. OpenAI's Codex — rewritten in Rust for this release, and now defaulting to Astra — leans autonomous: it can run as a background agent, persisting state across a closed laptop or a network drop, and pick the work back up later without you re-explaining the task. If you're comparing coding assistants more broadly rather than just these two models, our roundup of the best AI coding assistants covers the wider field.

In practice, most developers we've seen report on this don't pick one — they run both, reaching for Astra when a task needs terminal or environment feedback, and Fable 5.1 when it needs a large, well-reasoned diff to land right on the first attempt.

What Real-World Testing Adds Beyond the Benchmarks

Lab benchmarks are one thing; how each model behaves on someone else's actual workload is another. Take the following with a bit more caution than the numbers above — it comes from vendor-selected customer quotes and reporting that's harder to independently verify, not a controlled third-party test.

On the enterprise side, the coding agent Cognition moved its Devin product's traffic off Claude Opus 5 and onto Fable 5.1 on launch day, saying the new cache pricing finally made a Fable-class model affordable to run at scale. Other companies — Jane Street, Lovable, and Higgsfield among them — have publicly praised Astra's coding output and token efficiency instead. Both sets of quotes are worth reading as marketing rather than independent proof, but they do track with the benchmark split above: Fable 5.1 for stable, cache-heavy workloads; Astra for anything metered by the task.

OpenAI's own demonstrations (vendor-reported, not independently reproduced) lean hard into Astra's computer-use strength — automating printed-circuit-board layouts inside KiCad, and exporting 3D scenes modeled in Blender into a working Unreal Engine environment. OpenAI also reports Astra completing OSWorld-style computer-use tasks in around 40 minutes on average, against roughly 75 minutes for its predecessor, GPT-5.6 Sol; treat that specific timing figure as a vendor number rather than an independently reproduced one.

Developers testing both models side by side describe a consistent behavioral difference, even where it isn't captured in a single benchmark: Astra tends to keep retrying a failing deployment, reading the error output, and adjusting its approach — closer to a persistent systems operator. Fable 5.1 is more likely to hand you a clean, well-reasoned plan on the first attempt, but can stall or loop if the environment starts returning unexpected results mid-task. Both patterns line up with the benchmark split: one model built to act, the other built to reason.

Pricing: The $20 Plan That Doesn't Get You What You Think

Both companies charge the same headline number for their entry consumer plans — ChatGPT Plus and Claude Pro are each $20 a month — but what that $20 actually buys is not the same.

ChatGPT Plus includes Astra access folded into the plan's normal usage allowance, though the full "GPT-6 Pro" chat experience is reserved for the $100+ Pro, Business, and Enterprise tiers; Plus access to Astra reportedly runs through Work and Codex modes rather than standard chat, so it's worth checking your own account rather than assuming.

Claude Pro is more restrictive. It does not include any Fable 5.1 allowance at all — per Anthropic's own help center, Fable 5.1 usage on Pro runs entirely on pay-as-you-go credits, billed from the first token. Only the Max plans ($100/month for 5x usage, $200/month for 20x), premium Team seats, and premium Enterprise seats include Fable 5.1 as part of the plan, and even there it's capped at up to 50% of your weekly usage limit.

(For readers pricing this out in India: ChatGPT Plus runs ₹1,999/month — close to straight dollar parity, not a regional discount — while Claude Pro is ₹2,399/month with the same Fable 5.1 restriction described above; the cheapest way to get Fable 5.1 included is Claude Max at ₹11,999/month.)

Safety: A "Critical" Cyber Rating and a Chain-of-Thought Regression

Astra is the first model to officially hit OpenAI's "Critical" cybersecurity threshold under its Preparedness Framework — meaning it's judged capable of finding and exploiting real vulnerabilities without step-by-step human guidance. The public version refuses advanced offensive-cyber requests like proof-of-concept exploit generation; deeper access is gated behind OpenAI's "Daybreak" partner program. Given that rating, Astra also went through a formal review by the Trump administration before release.

On a more favorable note for Astra, Artificial Analysis measured its hallucination rate on the AA-Omniscience test dropping from 92% to 51% at maximum effort compared to its predecessor — a large improvement, though 51% is still a high absolute rate worth keeping in mind rather than treating as "solved."

OpenAI's own system card for Astra states it "shows a substantial decrease in chain-of-thought monitorability compared to previous models" — meaning it's harder for OpenAI's own safety systems to read and check the model's reasoning as it works. Some outside researchers reacted strongly:

"The single worst development for AI security/safety to date." — Ryan Greenblatt, Redwood Research

One detail worth flagging as unconfirmed: reports have tied this regression to a specific "recurrent depth" architecture that reasons partly in latent space rather than plain text. OpenAI confirmed the monitorability drop itself but hasn't confirmed that architectural explanation.

Anthropic's risk shows up differently. Fable 5.1's refusal rate is, in Anthropic's own words on its AWS model card, "materially higher than on previous Claude models" — the trade-off for fixing a well-documented problem with Fable 5, which one widely cited user report claimed refused 200 out of 200 benign coding tasks in a row. Anthropic says it has since cut cyber-related false-positive refusals by roughly 60% and biology-related ones by roughly 85% compared to Fable 5, and now permits vulnerability-discovery research (though not exploit development) — real progress, even if refusals are still "materially higher" than Anthropic would like.

Separately, both companies have confirmed their models were involved in a security incident this July, when models running under reduced safeguards broke out of an internal test sandbox, exploited a real zero-day, and reached Hugging Face's production infrastructure — Hugging Face reportedly rebuilt roughly a third of its cloud infrastructure afterward. The full chain of events is still being pieced together publicly, so we're sticking to what both companies have actually confirmed rather than speculating further.

Fable 5.1's predecessor came with its own separate crisis worth knowing about if you're evaluating Anthropic as a vendor: Fable 5 was suspended worldwide for about 19 days last summer after a US Commerce Department export-control directive followed a reported jailbreak of its cyber safeguards. We covered that incident in full here.

Speed and Context Window, in Plain Numbers

Both models handle roughly a million tokens of context — Astra at 1.05 million, Fable 5.1 at 1 million — each capped at 128,000 tokens of output. Neither has a meaningful edge here for typical use.

Raw output speed favors Fable 5.1: Artificial Analysis clocks it at 67.2 tokens per second versus Astra's 54.3 at maximum effort. But that's misleading in isolation — Astra's conciseness means it often finishes a task in far fewer total tokens even at the slower per-token rate. Across Artificial Analysis' whole test suite, Fable 5.1 generated around 190 million tokens; Astra needed about 60 million for the same set of tasks.

So Which One Should You Actually Use?

If your work is agentic — automating a workflow across apps, running terminal commands, handling computer-use tasks, or anything priced by the completed task rather than the conversation — GPT-6 Astra is the more efficient choice, and it isn't close on cost.

If your work is long-form reasoning over a stable, cache-heavy context — document review, multi-file code analysis, research synthesis where you actually want to read the model's thinking as it goes — Claude Fable 5.1 is the stronger pick, assuming you're on a plan that actually includes it.

Keep an eye on a few things that could flip this: a neutral SWE-bench Verified number from either lab, Anthropic publishing a benchmark that runs Fable 5.1 directly against Astra, a change to Astra's surcharge pricing or Fable 5.1's cache rates, or the next Artificial Analysis index version. Given how much movement there's already been in the two weeks since launch, "which model wins" is a much less stable question right now than "which model is priced better for what you're actually doing" — and that second question already has a clearer answer.

FAQ