GPT-6 Astra vs Claude Fable 5.1: Same Headline Price, a Very Different Bill

GPT-6 Astra vs Claude Fable 5.1: Same Headline Price, a Very Different Bill

On 1 September 2026, Anthropic released Claude Fable 5.1. Two days later, on 3 September, OpenAI released GPT-6 Astra and called it the world's most intelligent and most aligned model. Each lab says its model is the strongest. That much is expected.

What matters more to a business owner is not who wins the title, but one number: what you pay each month. And here is the surprise — the headline price of both models is literally identical, $10 per million input tokens and $50 per million output. Yet the final bill differs substantially. The difference is buried in a line item that never makes a headline.

What OpenAI actually shipped in Astra

Astra's announcement centres on something different from previous generations: computer use. The model doesn't only answer you — it operates software on your behalf. It fills in forms, updates customer records in a CRM, organises a calendar, runs research and drafts the summary in your email or document editor, builds a website and then QA-checks its own front end.

The numbers OpenAI published on this:

• Speed before accuracy. In a latency simulation on OSWorld 2.0, Astra scored 72.6% at roughly forty minutes per task, versus 65.7% at roughly seventy-five minutes for GPT-5.6 Sol — about 47% less time per task.
• With the updated Codex harness, OpenAI reports 1.9x faster task completion compared with the GPT-5.6 Sol experience on the Mind2Web benchmark.
• Staying inside task boundaries. On an evaluation built after the Hugging Face incident, GPT-5.6 Sol went beyond the authorized target 48% of the time without production safeguards. Astra did so in 0% of cases.
• Maths and reasoning. 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.

In cybersecurity, Astra meets the Critical threshold under OpenAI's Preparedness Framework. The version shipping today therefore refuses advanced offensive tasks such as building proof-of-concept exploits, while defensive work like secure code review and patching stays available.

What Anthropic shipped in Fable 5.1

Fable 5.1 pushes a different angle: coding, knowledge work, and stamina on long-running tasks. Its published numbers against Opus 5 and GPT-5.6 Sol:

• Agentic scientific research (Terminal-Bench-Science): 52.6% versus 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol.
• Agentic coding (Terminal-Bench 4.0): 55.8% versus 52.3% for Opus 5 and 37.3% for GPT-5.6 Sol.
• Knowledge work (GDPval-AA): 1853 points versus 1824 for Opus 5.
• Business workflows (AutomationBench): 31.4% versus 26.9% for Opus 5 and 19.6% for GPT-5.6 Sol.

Anthropic also added three effort levels — low, medium and high — and says low or medium effort matches or beats Fable 5 at a much lower cost.

Price: the line item that makes the difference

Published rates, per million tokens:

Item	GPT-6 Astra	Claude Fable 5.1	Gemini 3.8 Flash
Input	$10	$10	$0.75
Output	$50	$50	$3.75
Cache read	$1	$0.25	—
Context window	~1.1M tokens	1M tokens	1M tokens

Look at the third row. A cache read on Fable 5.1 costs a quarter of what it costs on Astra. That single line is what decides your bill if you run the model seriously.

If cache reads are new to you: when a model re-reads context it has already processed — your project files, your standing instructions, a long conversation history — you are billed at a reduced rate because it isn't reprocessing from scratch. A single agentic task re-reads the same context dozens of times. Cache reads therefore dwarf fresh input in agentic work, and a 4x gap on that line shows up at month end, not in the price table.

Anthropic estimates typical workloads cost about 25% less than Fable 5, with savings reaching roughly 45% on heavily agentic work. OpenAI, meanwhile, flags a line item that runs the other way: requests above 272,000 input tokens move to a higher rate for the entire request — $20 input and $75 output. Very long context on Astra is not free, which matters most if you push large documents in one shot.

Where Google sits in all this

Google did not ship a new frontier model in this wave. Its generally available flagship is still Gemini 3.1 Pro from February 2026. But it is playing a different game entirely: it launched Gemini 3.8 Flash on 2 September at an introductory $0.75 input and $3.75 output.

That is not a challenge to Astra or Fable 5.1 on intelligence. It is a challenge on something else — a 13x price gap. Most of what a mid-sized company needs daily — classifying a customer message, summarising a conversation, extracting fields from an invoice, drafting a reply — does not need a frontier model at all. Running those on a model that charges $50 per million output tokens means paying for intelligence you never use.

Why you should not trust any direct benchmark comparison

Here is the point most of what you will read this week skips. The two models were never compared to each other:

• Each lab benchmarked against itself. OpenAI compared Astra to GPT-5.6 Sol; Anthropic compared Fable 5.1 to Opus 5 and GPT-5.6 Sol. There is not one official figure placing Astra against Fable 5.1 on the same test.
• Even the shared benchmark doesn't compare. Both cite OSWorld 2.0, but OpenAI reports 72.6% in a latency simulation while Anthropic reports 77.9% partial-credit and 41.7% strict. Three numbers under three scoring methods. Putting them in one table compares different things.
• The labs ran their own measurements. That isn't an accusation, it is how every launch works. Anthropic itself cites a standard error of 3.5 to 4.5 points on the science benchmark, and independent verification takes weeks.
• Safeguards lower the score. Anthropic notes Fable 5.1 was evaluated with production safeguards on, and tasks where safeguards intervened scored zero — which depresses its published number without implying weakness.

The practical takeaway: the only benchmark that matters is your own workload. Run both models on fifty real cases from your business and measure quality, cost and latency together. One week of that testing is worth every published comparison table.

What this means for your business

The principle we build on at Origami has not changed despite how fast models turn over: don't tie your system to one model. Make the model a replaceable component in your design, not a foundation you build on. In practice, for this wave:

• Split work across three tiers. High-volume daily tasks on a cheap Flash-class model, mid-weight work on a balanced model, and only genuinely top-tier tasks on a frontier model. This alone cuts the bill more than any price negotiation.
• Budget on cache reads, not the headline rate. If your work is agentic — an assistant reading your files and replaying context — cache reads are your real bill, and a 4x gap there outweighs any benchmark delta.
• Test the lowest effort level first. Raise it only when measurement proves the need, not on impression.
• Before handing a model the keys to your systems, define its limits in writing: which systems it can reach, which actions require human approval, and where it logs what it did. Astra's computer-use capability means the model now acts rather than merely answers — which turns governance from a nicety into an operating requirement.
• If you process personal data, data retention is an architectural decision, not a legal footnote. Anthropic announced Enterprise Frontier Safeguards, storing data in cloud infrastructure the customer controls entirely, rolling out in phases starting this autumn. Any obligation under Saudi Arabia's Personal Data Protection Law is measured against what is in force today, not what is announced.

The Origami view

When a client asks us for an AI-backed system, we don't start from which model is strongest. We start by splitting their tasks into tiers by what each actually deserves, then build a layer that lets the model be swapped without rebuilding the system. The reason is simple: the model you pick today will not be the best one in six months. Anthropic and OpenAI both shipped frontier models within forty-eight hours of each other — that alone tells you any system bound to a single vendor has risk in its foundation, not at its margin.

Conclusion

Astra leads on computer use, speed, and staying inside task boundaries. Fable 5.1 leads on coding, scientific research and long-task stamina, and undercuts on the cache-read line that agentic work lives on. Gemini 3.8 Flash wins a completely different fight — daily running cost. The right question is not which is strongest, but which task deserves which tier. Whoever answers that with a measurement table from their own workload ends up with the best quality at the lowest bill.

Sources

• OpenAI — official GPT-6 Astra announcement, source of the OSWorld 2.0 and Mind2Web figures, the task-boundary evaluation, the cybersecurity findings and the rollout date.
• Anthropic — official Claude Fable 5.1 and Claude Mythos 5.1 announcement, source of the benchmark table, cache-read pricing, savings estimates and Enterprise Frontier Safeguards.
• OpenAI developer documentation — GPT-6 Astra model, for API pricing and context limits.

Figures as stated in the sources above on 4 September 2026. Benchmarks were run by each lab on its own models, and independent verification is still in progress.