Kimi K3, GLM-5.3, Qwen3.8 Max and Muse Spark pricing: the AI APIs rising on LLM Stats

1. The prices of the four new models#

The LLM Stats leaderboard of October 17, 2026 includes in its top 15 four models from providers that Codifly's price guides did not cover: Muse Spark 1.3 (Meta), Kimi K3 (Moonshot AI), GLM-5.3 (Z.ai) and Qwen3.8 Max (Alibaba Cloud). The table gathers their prices, read on each provider's official page, alongside four leaders for reference.

Prices are in dollars per million tokens and are standard rates. The blended price assumes 3 input tokens for every output token. The index is the LLM Stats composite, taken from its public table of October 17; it is a third-party measurement, not Codifly's.

Go deeperAI API prices in October 2026: what each model costs

Documentation: Moonshot AI: Kimi K3 pricing ↗ · Z.ai: GLM-5.3 pricing ↗ · Alibaba Cloud: Model Studio pricing (Qwen3.8 Max) ↗ · Meta: Muse Spark pricing and rate limits ↗ · LLM Stats: model leaderboard ↗

2. Measured quality versus price#

The four new models cluster between 51.0 and 52.4 index points, six to ten points below Claude Opus 5.5 (60.7) and Sonnet 5.5 (58.7). Their blended price runs from $2 to $6 per million, versus $4 and $8 for those leaders.

Two readings. First, the composite index averages very different tests: a one-point difference can be invisible on your task or decisive. Second, the price difference is real and accumulates with volume, as the worked example shows.

3. Worked example: the same support bot#

We reuse the example from the earlier guides: 100,000 conversations a month, with 1,500 input tokens and 400 output tokens each (150 million input and 40 million output). The code calculates the monthly cost and how many index points you get per $100 spent; it is a way to compare efficiency, not quality.

MODELS = {  # (input, output, LLM Stats index of Oct 17, 2026), USD per million tokens
    "Claude Opus 5.5": (4.00, 20.00, 60.7),
    "Claude Sonnet 5.5": (2.00, 10.00, 58.7),
    "Gemini 4 Argon": (2.00, 10.00, 55.4),
    "GPT-6.1 Sol": (2.00, 10.00, 54.0),
    "Muse Spark 1.3": (1.25, 4.25, 52.4),
    "Kimi K3": (3.00, 15.00, 52.2),
    "GLM-5.3": (1.40, 4.40, 51.3),
    "Qwen3.8 Max (Singapore)": (2.00, 6.00, 51.0),
}
CONVERSATIONS, INPUT, OUTPUT = 100_000, 1_500, 400   # tokens per conversation

def monthly(input_price, output_price):
    return CONVERSATIONS * INPUT / 1e6 * input_price + CONVERSATIONS * OUTPUT / 1e6 * output_price

for model, (input_price, output_price, index) in sorted(MODELS.items(), key=lambda m: monthly(m[1][0], m[1][1])):
    cost = monthly(input_price, output_price)
    print(f"{model}: {cost:,.2f} a month | {index / cost * 100:,.2f} points per $100")

Muse Spark 1.3 is the cheapest in the example ($357.50 a month) and Claude Opus 5.5 the most expensive ($1,400): a 3.9-times difference. Per index point, Muse Spark 1.3, GLM-5.3 and Qwen3.8 Max deliver 2.2 to 3.4 times more than Claude Opus 5.5; Kimi K3, with $15 output, only 1.15 times. And the metric hides that 52 and 60 points can mean very different error rates on your case.

Go deeperWhat an AI API really costs in Latin American currencies (September 2026)

4. Conditions that change the price#

These are the conditions we verified on the official pages; others (discounts, quotas, subscription plans) may exist.

• Qwen3.8 Max has regional rates: $2 input and $6 output in Singapore and in the US scope of Virginia, and $1.65 and $4.951 in the other global regions. Choose the region by your compliance needs, not only price.
• Qwen3.8 Max offers 1 million free tokens for 90 days in Singapore, and its page marks every row with a caching discount without giving the rate.
• Kimi K3, GLM-5.3 and Muse Spark 1.3 publish a cache-read rate ($0.30, $0.26 and $0.15 per million): if your prompt repeats, effective input drops a lot.
• Prices exclude taxes and may change without notice; note the date you read them.

Before sending customer data to any provider, review its data policy and the region where it is processed. That decision usually weighs more than a few cents per million tokens.

5. How to choose#

Price and index are filters, not the final decision. The evaluation on your own cases is what rules.

To compare these models with those from OpenAI, Anthropic, Google and xAI, use Codifly's AI model comparator and the October price guide; to estimate your volume, the AI cost calculator.

Go deeperHow to Choose an LLM for Production Without Relying on Leaderboards

6. Keep your numbers current#

This group of providers changes rates and models quickly. The figures are those on the official pages on October 17, 2026, and the index is from LLM Stats' public table that day.

• Note the date you read each rate and each index.
• Recalculate the example with your real input-to-output ratio.
• Confirm region, data policy and terms of use before committing.