Ling-3.1-flash vs MiniMax-M3.1-Flash: API pricing, deals, and off-peak rates

How they differ

Three workloads at volume show what these rate cards mean in practice. A ticket triage service that sends 2,000,000 requests a month, each with 1,500 input tokens and 200 output tokens and no cache hits, costs $1,260 on Ling-3.1-flash and $690 on MiniMax-M3.1-Flash at standard rates. A support assistant with 300,000 conversations a month, each carrying 8,000 input tokens and 600 output tokens, with three in four input tokens read from cache, costs $450 and $225. A nightly tagging job over 150,000 documents of 25,000 input tokens and 800 output tokens each, with no cache hits, costs $1,233 and $634.50. Ling's deal takes 75% off, so its rates become $0.075 input, $0.015 for cache reads and $0.225 output, and the nightly job costs $308.25 instead of $1,233. A deal is the provider's call and can end, so don't build a yearly budget on it. MiniMax's off-peak rates are $0.075 input, $0.0075 for cache reads and $0.30 output. Run the nightly job inside that window and it costs $317.25 instead of $634.50. The hours belong to MiniMax, so read the current window on its pricing page before you schedule anything around it. With both specials live, Ling's deal rates are $0.075 input and $0.225 output, against MiniMax's off-peak $0.075 and $0.30. Put those side by side for your own mix, since the cheaper side can differ between input and output. Only the off-peak window needs scheduling. A deal applies to any request, which makes it easier to budget for interactive traffic. These examples leave out cache writes and tool fees.

Verdict

Sort your traffic into three piles before you read any rate card. Interactive traffic can't wait for an off-peak window, so it pays standard rates on MiniMax, while Ling's deal applies whenever it's live. Batch work that can run at night is where off-peak hours help. Anything that sends images or video rules out Ling, because its API listing takes text only. On cache reads, MiniMax charges $0.015 per million tokens against Ling's $0.06, which matters for assistants that resend a long prompt every turn. Hosting your own copy changes the picture again. Ling's memory budget is almost all weights, because its cache stays small even at long context. MiniMax keeps a large cache at long context, so the length you serve matters far more there. Run a few hundred of your own requests through both before you commit.