GPT-6 Astra vs Claude Fable 5.1: same price, 2.4x the bill

GPT-6 Astra and Claude Fable 5.1 have the same price on FastMetal: $10.55 per million input tokens and $52.75 per million output tokens. A price table says they cost the same. Run the same prompts through both and Fable 5.1's bill comes out about 2.4x Astra's, because it returns far more tokens.

The short version:

• Fable 5.1 used more output tokens on all 43 prompts, more than double on 30 of them, and about 2.4x in total.
• The readable answer grew only 1.4x. Most of the extra tokens are reasoning you pay for but never see.
• Fable 5.1 was about twice as slow: 45.5 s on average against 22.3 s.

How we measured

We sent the 43 prompts from FastMetal's comparison set to both models under the same conditions in September 2026: debugging, code review, SQL optimization, business email, article outlines, decision write-ups and similar tasks. No system prompt, no max_tokens. This is what each model does when you don't tune anything.

We recorded response time and output tokens (completion_tokens). Input is the same prompt for both, so it costs about the same on either model. The difference is all on the output side.

It's one run on a small set. Read it as a tendency, not a law.

Results

Tokens went up 2.4x; the text you can actually read went up 1.4x. The gap is mostly thinking tokens. They are billed as output tokens, but they don't appear in the answer.

What that costs

With identical output rates, the bill ratio is the token ratio. At $52.75 per million output tokens, the averages above work out to about $0.066 of output per request on Astra and $0.157 on Fable 5.1. Across all 43 prompts, that's roughly $2.84 against $6.73.

A few cents per request doesn't matter much on its own. An agent that calls the model dozens of times per task multiplies the 2.4x straight into its cost. The latency doubles too, which you will feel in any interactive UI.

One pricing detail the headline rates hide: Astra's input rate doubles to $21.10 per million on prompts over 272K tokens, while Fable 5.1 keeps one rate across its 1M-token window. If you routinely send very long prompts, check the pricing page before assuming the two stay level.

Which one to use

• GPT-6 Astra when the shape of the answer is known. Conversion, classification, summarization, extraction: shorter, faster and cheaper for the same per-token price.
• Claude Fable 5.1 when the reasoning is the point. It spends more tokens because it thinks longer. For root-cause analysis or design decisions, where the process matters more than the conclusion, that can be worth paying for.

A price table can't make this choice for you. Two models at the same rate can still differ by more than 2x on the invoice. The reliable test is to run a few dozen of your own prompts through both and compare completion_tokens.

Try both

Same key, same balance; only model changes:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-6-astra", "messages": [{"role": "user", "content": "Hello"}]}'

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "anthropic-claude-fable-5-1", "messages": [{"role": "user", "content": "Hello"}]}'

Both models' answers to the same prompts are side by side on the comparison page. Details are on the GPT-6 Astra and Claude Fable 5.1 pages.