DeepSeek V4.1 Flash vs MiniMax M3.1 Flash: Identical Cards, Different Models

Two vendors, two models, one price. Both DeepSeek V4.1 Flash and MiniMax M3.1 Flash cost $0.15 per million input tokens and $0.60 per million output on OrcaRouter. Both serve a 1M-token context. Neither is a flagship, and both are the tier their vendor wants volume to land on. The sticker price cannot separate them, which makes this the cleanest comparison on the site: every other column decides, and none of them are hidden.

The cards

• DeepSeek V4.1 Flash — $0.15 in, $0.60 out; $0.003 cached input; doubling to $0.30 and $1.20 in two weekday windows; 1M context; 384K maximum output; 2,500 concurrent requests; vision supported; measured p50 first token 1.49 seconds.

• MiniMax M3.1 Flash — $0.15 in, $0.60 out; 1M context; 524K maximum output; measured p50 first token 2.66 seconds; no published cache-hit line on the card as we read it.

Three differences matter, and all three are measurable before you commit: output ceiling, first-token latency, and whether a cached prefix gets a discount.

Ceiling: 524K against 384K

MiniMax's output ceiling is 36% larger. On a single call, that is the difference between a 350,000-token generation fitting and not — and 350,000 tokens of output at $0.60 per million is 21 cents either way, so the ceiling is a capability line rather than a cost line. If your workload never writes more than 300K in one request, this column is a tie with a marketing number attached.

If it does, MiniMax is the only one of the two that will accept the job, and the identical price stops being the point.

Speed: 1.49 against 2.66 seconds

DeepSeek's measured first token is 1.49 seconds; MiniMax's is 2.66. Both are fast enough that a chat interface feels responsive, and both are slow enough that an agent making twenty turns per task pays attention to the difference — about 23 seconds per task at these figures, against a token bill of cents. On throughput-bound pipelines the faster model is simply more work done per hour on the same concurrency budget.

The cache line, which only one card shows

DeepSeek publishes a cache-hit price: $0.003 per million input tokens, against $0.15 for a miss. That is a fifty-fold discount on the part of a prompt that repeats, and it is the single largest lever on either card. MiniMax's card, as we read it, carries no equivalent row — which does not mean caching is absent, only that the discount is not publishable from the card.

For an agent that re-sends a stable system prompt and a growing history on every turn, that asymmetry can outweigh both the ceiling and the latency. It is also the first thing to test rather than assume: send the same prefix twice and compare the two bills.

The peak window, which only one card has

DeepSeek doubles every rate between 01:00–04:00 and 06:00–10:00 UTC on weekdays, Chinese public holidays excluded. MiniMax's $0.15 and $0.60 are flat. A European or Asian workload that runs during business hours is on DeepSeek's peak rate for most of its traffic — $0.30 and $1.20, which is twice MiniMax's sticker for the same tokens.

So the identical cards are not identical in practice. Off-peak, with a stable cached prefix, DeepSeek is materially cheaper. In a weekday peak window with no caching, it is twice as expensive. The schedule decides which sentence is true.

What to run

Both models are on OrcaRouter at their vendors' published rates, on one OpenAI-compatible endpoint with automatic failover, which is what makes this test cheap: route a slice of traffic at each and compare output tokens per completed task, p95 latency, and the share of requests landing in DeepSeek's peak windows.

What you should not do is pick on the sticker. It is the one column where these two models agree, which means it is the one column that tells you nothing.