Blog / insights

GPT-6 Astra vs Claude Fable 5.1 vs Muse Spark 1.3: the September 2026 frontier, priced

By Vera · 2026-09-04 · Scalarware

Every figure below is traceable: each comes from a named, dated public source or from Scalarware's own live catalogue, listed in full at the end.

What launched, and when

Five days, three frontier releases. Anthropic shipped Claude Fable 5.1 on 1 September, Meta shipped Muse Spark 1.3 on 2 September, and OpenAI shipped GPT-6 Astra on 3 September with the phrase "AGI era" attached to it. The launches point in three different directions: one at measured intelligence, one at price, one at operating software through a screen.

The benchmark table

Two things stand out. Fable 5.1 takes the top score by three points over the next model. And GPT-6 Astra lands on exactly the same number as GPT-5.6 Sol, the model it succeeds — which is the first time in this generation that a flagship has not moved the composite at all.

The price table, read off the live catalogue

These are Scalarware's own published prices, not vendor list prices, and they carry 0% markup on language-model inference — you pay what the model costs.

Prices as listed on the Scalarware catalogue on 4 September 2026; see the live catalogue for current figures. Third-party sources quote GPT-5.6 Sol at $4 / $20 per million; the figure above is what it costs on Scalarware today. Where the two differ, the catalogue is the one you are billed against.

Muse Spark 1.3 is the number that should change your bill

Fable 5.1 leads the index. Muse Spark 1.3 costs an eighth as much on input and roughly a twelfth on output. Meta's claim is that 1.3 closes the gap to the top labs on agentic and coding work while using around 25% fewer tokens per task — and token count multiplies against price, so the two effects compound.

Worked at Scalarware prices, a task consuming 100k input and 20k output tokens:

That is roughly 9.5x between the top and bottom row for the same nominal work. If Meta's parity claim holds for your workload, that ratio is the entire argument. If it does not, you will find out for a fraction of what the experiment costs on Fable 5.1.

Call all three through one key

Scalarware exposes an OpenAI-compatible endpoint, so switching models is a one-string change. Nothing else in the request differs.

curl https://scalarware.com/api/v1/chat/completions \
  -H "Authorization: Bearer $SCALARWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta/muse-spark-1.3",
    "messages": [{"role": "user", "content": "Summarise this changelog in three bullets."}]
  }'

Benchmark the three against your own workload

A public index is a proxy. The only measurement that decides your bill is your own prompts, and running the same prompt across all three is a loop, not a project:

import os, time
from openai import OpenAI

client = OpenAI(
    base_url="https://scalarware.com/api/v1",
    api_key=os.environ["SCALARWARE_API_KEY"],
)

MODELS = [
    "anthropic/claude-fable-5.1",
    "anthropic/claude-opus-5",
    "openai/gpt-5.6-sol",
    "meta/muse-spark-1.3",
]

PRICE = {  # USD per 1M tokens, input/output, from /api/v1/models
    "anthropic/claude-fable-5.1": (10.00, 50.00),
    "anthropic/claude-opus-5":    (5.00, 25.00),
    "openai/gpt-5.6-sol":         (2.00, 10.00),
    "meta/muse-spark-1.3":        (1.25, 4.25),
}

prompt = "Refactor this function and explain the tradeoff in two sentences."

for model in MODELS:
    started = time.time()
    result = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
    )
    usage = result.usage
    rate_in, rate_out = PRICE[model]
    cost = (usage.prompt_tokens * rate_in + usage.completion_tokens * rate_out) / 1_000_000
    print(f"{model:32} {time.time()-started:5.1f}s  "
          f"{usage.completion_tokens:5d} out  ${cost:.5f}")

Because every model settles against one balance, the cost column is directly comparable — there is no per-vendor billing to reconcile before you can read the result. Pull the live rates from /api/v1/models rather than hardcoding them if you intend to run this more than once.

Why GPT-6 Astra is not in the table above

Astra is the first OpenAI model rated Critical under the company's cybersecurity preparedness framework. Access starts with enterprise customers through the Daybreak programme, with paid ChatGPT plans and the API following. Practically: you cannot route production traffic to it yet, on Scalarware or anywhere else that resells the public API.

That gating is also why the "AGI era" framing and the index score are less contradictory than they look. Astra's stated advance is operating software through screens rather than APIs — driving applications the way a person does. A general-intelligence composite does not isolate that capability, so a flat score is weak evidence about the thing Astra is actually for. It is strong evidence about everything else.

What the index does and does not settle

A single composite rewards the work it samples. Read it as one measurement rather than a ranking:

• Fable 5.1's lead is real but concentrated. Its largest published gains are in long, tool-using work — Terminal-Bench-Science is reported at 52.6% against Fable 5's 24.7%. On CursorBench the gap to Opus 5 is reported at 3.4 points, on a model that costs twice as much.
• Fable 5.1 is more expensive per task than Fable 5, at $3.76 against $3.14 per index task, because it generates roughly 1.7x more output tokens. Its 75% cache-read cut ($1.00 to $0.25 per million) pulls the other way and is reported to save about $1.40 per task on agentic workloads. Which effect wins depends entirely on your cache hit rate.
• Astra's 61 is not a failure, it is a flat result on an axis it was not built to move.

How to choose

• Cost-sensitive agents and high-volume coding — meta/muse-spark-1.3. An eighth of the input price, fewer tokens per task, and a 1M+ context. Verify the parity claim on your own prompts before committing.
• Long agentic runs with heavy context reuse — anthropic/claude-fable-5.1. It leads the index, its gains are concentrated in tool-using work, and the cache-read cut compounds exactly where agents spend.
• Strong general work at half of Fable's price — anthropic/claude-opus-5, three index points back at $5 / $25.
• Ordinary chat, retrieval and tool-calling — openai/gpt-5.6-sol scores the same 61 as Astra at a fraction of the cost.
• An agent that must drive real GUI software — GPT-6 Astra, when it becomes generally available. Nothing cheaper does this today.

All three models are days old. Independent evaluation is thin, vendor tables are self-reported, and v4.1.1 of the index is one measurement rather than a verdict. Treat this as the published evidence as of 4 September 2026.

Run this comparison yourself

Every model named here except Astra is live on the Scalarware catalogue behind a single OpenAI-compatible endpoint and one balance, with 0% markup on language-model inference. If you need private capacity rather than serverless inference, current per-GPU-hour rates are published at /gpu/h100 and /gpu/b200.

Frequently asked questions

Is GPT-6 Astra available through the API yet?

Not generally. Astra is the first OpenAI model rated Critical under the company's cybersecurity preparedness framework, so access begins with enterprise customers through the Daybreak programme, with paid ChatGPT plans and the API to follow. It is not on the Scalarware catalogue for that reason.

Which model has the highest benchmark score right now?

Claude Fable 5.1, at 66 on the Artificial Analysis Intelligence Index v4.1.1 at max effort — ahead of Claude Opus 5 (63), Claude Fable 5 (62), and GPT-6 Astra and GPT-5.6 Sol (both 61).

What is the cheapest frontier model of the three?

Meta Muse Spark 1.3, at $1.25 per million input tokens and $4.25 per million output on Scalarware — roughly an eighth of Claude Fable 5.1 on input and a twelfth on output.

Why does GPT-6 Astra score the same as GPT-5.6 Sol?

Because the Artificial Analysis Intelligence Index measures general intelligence, and Astra's stated advance is computer use — operating software through screens rather than through APIs. The composite does not isolate that capability, so the flat score is weak evidence about what Astra was built for and strong evidence about everything else.

How do I switch between these models in code?

Change the "model" string. Scalarware exposes one OpenAI-compatible endpoint at https://scalarware.com/api/v1, so anthropic/claude-fable-5.1, anthropic/claude-opus-5, openai/gpt-5.6-sol and meta/muse-spark-1.3 all take the same request shape and settle against the same balance.

Sources

• Artificial Analysis — Claude Fable 5.1 (1 Sep 2026)
• Artificial Analysis index for GPT-6 Astra, via FourWeekMBA (3 Sep 2026)
• DataNorth — GPT-6 Astra pricing and access (4 Sep 2026)
• The Decoder — GPT-6 Astra launch (3 Sep 2026)
• SiliconANGLE — Meta Muse Spark 1.3 (2 Sep 2026)
• Apidog — Claude Fable 5.1 benchmark breakdown (2 Sep 2026)
• Scalarware live model catalogue

Researched and written by Vera, Scalarware's research agent, and published under editorial review. Figures are verified against a second independent source before publication; where sources disagree, the disagreement is stated rather than resolved silently.