FinOps
Kimi K3, GLM-5.3, Qwen3.8 Max and Muse Spark pricing: an LLM Stats comparison
Per-million-token rates from Moonshot, Z.ai, Alibaba and Meta compared with other models, using updated third-party indices and a reproducible monthly example.
In this article
Three of the four new models deliver between 2.2 and 3.4 times more index points per dollar than Claude Opus 5.5: in the support example, Muse Spark 1.3 costs $357.50 a month versus $1,400, but the index does not measure your case: evaluate it with your data.
1. The prices of the four new models#
We compare Muse Spark 1.3 (Meta), Kimi K3 (Moonshot AI), GLM-5.3 (Z.ai) and Qwen3.8 Max (Alibaba Cloud) with four other models. Rates come from each provider; indices were observed in the public LLM Stats table on November 4, 2026. This is a model selection, not the current top 15.
| Model | Provider | Input (USD/M) | Output (USD/M) | Cache read | Blended 3:1 | LLM Stats index |
|---|---|---|---|---|---|---|
| Muse Spark 1.3 | Meta | $1.25 | $4.25 | $0.15 | $2 | 52.0 |
| Kimi K3 | Moonshot AI | $3 | $15 | $0.30 | $6 | 51.4 |
| GLM-5.3 | Z.ai (Zhipu AI) | $1.40 | $4.40 | $0.26 | $2.15 | 50.5 |
| Qwen3.8 Max | Alibaba Cloud | $2 | $6 | — | $3 | 50.4 |
| Claude Opus 5.5 | Anthropic | $4 | $20 | — | $8 | 59.9 |
| Claude Sonnet 5.5 | Anthropic | $2 | $10 | — | $4 | 57.9 |
| Gemini 4 Argon | $2 | $10 | — | $4 | 54.4 | |
| GPT-6.1 Sol | OpenAI | $2 | $10 | — | $4 | 53.4 |
Standard rates are USD per million tokens. Blended pricing assumes three input tokens for every output token. The composite index belongs to LLM Stats: it is neither a Codifly-run benchmark nor a measurement of your task.
AI API prices in November 2026: rates and real 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#
In this snapshot, the four models score between 50.4 and 52.0, versus 59.9 for Claude Opus 5.5 and 57.9 for Sonnet 5.5. Blended pricing ranges from $2 to $6 per million, versus $4 and $8 for those two models. Positions and scores can change between reviews.
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 Nov 4, 2026), USD per million tokens
"Claude Opus 5.5": (4.00, 20.00, 59.9),
"Claude Sonnet 5.5": (2.00, 10.00, 57.9),
"Gemini 4 Argon": (2.00, 10.00, 54.4),
"GPT-6.1 Sol": (2.00, 10.00, 53.4),
"Muse Spark 1.3": (1.25, 4.25, 52.0),
"Kimi K3": (3.00, 15.00, 51.4),
"GLM-5.3": (1.40, 4.40, 50.5),
"Qwen3.8 Max (Singapore)": (2.00, 6.00, 50.4),
}
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")| Model | Monthly cost (100,000 conversations) | Index | Index points per $100 |
|---|---|---|---|
| Muse Spark 1.3 | $357.5 | 52.0 | 14.55 |
| GLM-5.3 | $386 | 50.5 | 13.08 |
| Qwen3.8 Max (Singapore) | $540 | 50.4 | 9.33 |
| Claude Sonnet 5.5 | $700 | 57.9 | 8.27 |
| Gemini 4 Argon | $700 | 54.4 | 7.77 |
| GPT-6.1 Sol | $700 | 53.4 | 7.63 |
| Kimi K3 | $1,050 | 51.4 | 4.90 |
| Claude Opus 5.5 | $1,400 | 59.9 | 4.28 |
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 1.14 times. And the metric hides that 52 and 60 points can mean very different error rates on your case.
What 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.
| Region (Qwen3.8 Max) | Input (USD/M) | Output (USD/M) |
|---|---|---|
| Singapore and US (US scope) | $2 | $6 |
| Virginia, Frankfurt, Tokyo and Hong Kong (global scope) | $1.65 | $4.951 |
- 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.
| If your priority is… | Look first at | Candidates to evaluate |
|---|---|---|
| Lowest cost with close quality | The blended price and the free quota | GLM-5.3 or Muse Spark 1.3 |
| Long context | Each model's published window | Kimi K3 (1,048,576 tokens) |
| Highest index in this group | Third-party index and evaluation on your task | Claude Opus 5.5 or Sonnet 5.5 |
| Data residency or compliance | The region and the provider's data policy | The provider and region your contract allows |
| Workloads with a lot of repeated prompt | The cache-read rate | Kimi K3 ($0.30), GLM-5.3 ($0.26) or Muse Spark 1.3 ($0.15) |
To compare these models with those from OpenAI, Anthropic, Google and xAI, use Codifly's AI model comparator and the November price guide; to estimate your volume, the AI cost calculator.
How to Choose an LLM for Production Without Relying on Leaderboards
6. Keep your numbers current#
Rates and indices were reviewed on November 4, 2026. The calculation retains the hypothetical workload described here; confirm regional rates and evaluate models with your own data before deciding.
- 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.