
Model comparison
Qwen3.8 Max vs MiniMax M3 API price, context and cost.
Qwen3.8 Max (Alibaba) and MiniMax M3 (MiniMax) side by side: per-1M prices, cached-input rates, context, output limits and capability flags from the live RouterPlex catalog, with three worked costs. A coding-agent session costs $2.52 on Qwen3.8 Max and $0.29 on MiniMax M3. Both run on one prepaid key with the same OpenAI-compatible request.
Side by side
Catalog fields, one row each.
| Catalog field | Qwen3.8 Max | MiniMax M3 |
|---|---|---|
| Provider | Alibaba | MiniMax |
| Model ID | qwen3.8-max | MiniMax-M3 |
| Input / 1M | $2.00 | $0.30 |
| Output / 1M | $6.00 | $1.20 |
| Cached input / 1M | $0.70 | $0.03 |
| Context window | 1M (1,000,000 tokens) | 1M (1,000,000 tokens) |
| Max output | 131K tokens | 512K tokens |
| Image input | Yes | Yes |
| Tool calling | Yes | Yes |
| Reasoning | Yes | Yes |
| Long-prompt tier | None | None |
| Price basis | Vendor list price, 0% markup | Vendor list price, 0% markup |
- Qwen3.8 Max: RouterPlex bills $2 input / $6 output per 1M tokens, Alibaba's list price. Cached prompt tokens are billed at $0.70 per 1M, above Alibaba's own cache-hit tier: our cache price is set from what our upstream charges us for a cached token, so it does not follow the lab's rate.
What the catalog says
MiniMax M3 is 85% cheaper on input ($0.30 against $2.00 per 1M tokens). MiniMax M3 is 80% cheaper on output ($1.20 against $6.00 per 1M tokens). MiniMax M3 is 96% cheaper on cached input ($0.03 against $0.70 per 1M tokens). On the coding-agent session below, MiniMax M3 costs $0.29 against $2.52 for Qwen3.8 Max. Both accept a 1M context window. MiniMax M3 can return up to 512K tokens per response, against 131K for Qwen3.8 Max. On LMArena Text Leaderboard (2026-09-02), Qwen3.8 Max has an overall score of 1479.6 and MiniMax M3 1442.9. That is one public leaderboard, not a test of your workload.
Generated from catalog prices, limits and capability flags. Price does not measure output quality.
Published benchmarks
Where real data exists.
Raw score as the source publishes it, percentile in brackets. Only rows with a cited source are shown; "Not listed" means the source has no entry for that model.
| Source / category | Qwen3.8 Max | MiniMax M3 |
|---|---|---|
| overallLMArena Text Leaderboard | 1479.6 (95) | 1442.9 (80) |
| codingLMArena Text Leaderboard | 1519.7 (93) | 1498.0 (83) |
| reasoningLMArena Text Leaderboard | 1497.3 (97) | 1438.0 (79) |
| writingLMArena Text Leaderboard | 1466.5 (96) | 1407.1 (79) |
| generalLMArena Text Leaderboard | 1476.7 (95) | 1435.0 (82) |
LMArena Text Leaderboard, version 2026-09-02, retrieved 2026-09-10. A leaderboard measures its own prompt set, not your workload. See the head-to-head benchmark page.
Worked cost examples
Same workload, both bills.
Each line multiplies the catalog rate per 1M tokens by the tokens shown. Cached tokens use the route's cached-input rate, or the full input rate when it has no cache discount. Reasoning tokens bill as output, so a model that thinks at length raises the output line beyond these figures.
40 requests of 50K prompt tokens, 80% of each prompt a cache hit, 2.5K output each.
0.4M input x $2.00 + 1.6M cached x $0.70 + 0.1M output x $6.00 = $2.52
0.4M input x $0.30 + 1.6M cached x $0.03 + 0.1M output x $1.20 = $0.29
10,000 conversations of 2K prompt tokens and 500 output tokens, no caching.
20M input x $2.00 + 5M output x $6.00 = $70.00
20M input x $0.30 + 5M output x $1.20 = $12.00
200 requests of 150K prompt tokens and 2K output tokens, no caching.
30M input x $2.00 + 0.4M output x $6.00 = $62.40
30M input x $0.30 + 0.4M output x $1.20 = $9.48
Switching
One key, one line.
Both models sit behind the same RouterPlex key and the same OpenAI-compatible endpoint. Keep the request shape and change the model string. Claude Code and the Anthropic SDK use the same ids on https://api.routerplex.com.
Read the quickstartimport osfrom openai import OpenAIclient = OpenAI(base_url="https://api.routerplex.com/v1",api_key=os.environ["ROUTERPLEX_API_KEY"],)reply = client.chat.completions.create(model="qwen3.8-max", # switch: "MiniMax-M3"messages=[{"role": "user", "content": "Hello!"}],)print(reply.choices[0].message.content)
# Qwen3.8 Maxcurl https://api.routerplex.com/v1/chat/completions \-H "Authorization: Bearer $ROUTERPLEX_API_KEY" \-H "Content-Type: application/json" \-d '{"model": "qwen3.8-max", "messages": [{"role": "user", "content": "Hello!"}]}'# MiniMax M3: same request, one field changedcurl https://api.routerplex.com/v1/chat/completions \-H "Authorization: Bearer $ROUTERPLEX_API_KEY" \-H "Content-Type: application/json" \-d '{"model": "MiniMax-M3", "messages": [{"role": "user", "content": "Hello!"}]}'
Questions
Before you pick one of the two.
01Is Qwen3.8 Max cheaper than MiniMax M3?+
Qwen3.8 Max costs $2.00 input and $6.00 output per 1M tokens on RouterPlex; MiniMax M3 costs $0.30 and $1.20. On a coding-agent session (40 requests of 50K prompt tokens, 80% of each prompt a cache hit, 2.5K output each) Qwen3.8 Max comes to $2.52 and MiniMax M3 to $0.29, so MiniMax M3 is cheaper for that workload. A chat app month (10,000 conversations of 2K prompt tokens and 500 output tokens, no caching) costs $70.00 on Qwen3.8 Max and $12.00 on MiniMax M3.
02Which has the larger context window, Qwen3.8 Max or MiniMax M3?+
They match: both accept 1M tokens of context on RouterPlex. Maximum output is 131K for Qwen3.8 Max and 512K for MiniMax M3.
03Do Qwen3.8 Max and MiniMax M3 support image input and tool calling?+
Per the RouterPlex catalog, Qwen3.8 Max: image input, tool calling, reasoning. MiniMax M3: image input, tool calling, reasoning. These flags come from the labs' published model metadata, not from a quality test.
04How do I switch from Qwen3.8 Max to MiniMax M3?+
Change one string. Both run on the same RouterPlex key and the same OpenAI-compatible endpoint (https://api.routerplex.com/v1), so the request stays identical and only model changes from "qwen3.8-max" to "MiniMax-M3". The same ids work on the Anthropic-compatible /v1/messages endpoint for Claude Code.
05What do benchmarks say about Qwen3.8 Max vs MiniMax M3?+
LMArena Text Leaderboard (2026-09-02, retrieved 2026-09-10) lists an overall score of 1479.6 for Qwen3.8 Max and 1442.9 for MiniMax M3. A public leaderboard measures its own prompt set; test both on your prompts before moving traffic.
06When is Qwen3.8 Max worth the higher price?+
Not on price, context or catalog capabilities: MiniMax M3 matches or beats it on each. The reason to pay more would be output you prefer on your own prompts; this page does not rank quality, so run both in the RouterPlex Arena before moving traffic.
Provider pricing
Try both
Both models behind one key.
Top up from $5 with no subscription, then send the same prompt to each.