Three frontier coders now cost under a dollar a million. The cheapest sticker is not the cheapest bill.
DeepSeek V4-Pro, Qwen3.7 Plus, and MiniMax M3 all landed in the last two months under $0.60 per million input tokens. If you sort the price list by input rate, Qwen looks like the winner. Run an actual coding month through all three and DeepSeek comes out ahead by a wide margin, because output and cache are where the money goes. Here is the math, and the two reasons you might still pay more.
Photo by Pawel Czerwinski on Unsplash
The three rate cards
| Model | Input / 1M | Output / 1M | Cache read | Context | Inputs |
|---|---|---|---|---|---|
| Qwen3.7 Plus | $0.40 | $1.60 | $0.08 | 1M | Text, image, video |
| DeepSeek V4-Pro | $0.435 | $0.87 | $0.0036 | 1M / 384K out | Text only |
| MiniMax M3 | $0.60 | $2.40 | $0.12 | 1M | Text, image, video |
List prices from each provider, June 2026. MiniMax M3's 7-day launch promo (half price) expired June 8, so these are the standing rates. Qwen and MiniMax cache rates apply to cache hits; DeepSeek cache hit is $0.003625/M, rounded here.
Why the lowest input price misleads you
Qwen3.7 Plus has the lowest input rate of the three at $0.40 per million. DeepSeek V4-Pro is a hair more at $0.435, and MiniMax M3 is the priciest at $0.60. So if you glance at a sorted price list, Qwen wins and you move on.
The ranking flips on output. DeepSeek charges $0.87 per million output tokens. Qwen is $1.60, almost double. MiniMax is $2.40, about 2.76x DeepSeek. Output tokens are the ones the model generates: every line of code, every reasoning trace, every tool call argument. On a coding agent those add up fast, and they bill at the output rate, not the input rate.
So the question is not "who has the lowest sticker" but "what is your input-to-output ratio." Below a certain output share, Qwen's cheaper input wins. Above it, DeepSeek's cheaper output takes over. Here are four workload shapes priced per the token mix shown, running from almost-all-input at the top to output-heavy at the bottom.
| Workload shape | DeepSeek V4-Pro | Qwen3.7 Plus | MiniMax M3 |
|---|---|---|---|
| Near-pure input (20M in / 0.5M out) | $9.13 | $8.80 | $13.20 |
| Retrieval-heavy (10M in / 1M out) | $5.22 | $5.60 | $8.40 |
| Balanced chat (1M in / 1M out) | $1.31 | $2.00 | $3.00 |
| Reasoning-heavy (1M in / 3M out) | $3.05 | $5.20 | $7.80 |
Qwen only wins the top row, a 40:1 input-to-output ratio almost no real agent hits. By the 10:1 retrieval row, DeepSeek is already ahead, because its slightly higher input is more than offset by output that costs half as much. Accent marks the cheapest in each row.
One coding month, three meters
Take a small team running a coding agent: roughly 200 million input tokens and 40 million output tokens a month. That is a 5:1 ratio, typical when an agent reads a lot of repo context and writes comparatively little. Priced flat, no cache:
| Model | Flat monthly cost | With 70% cached input |
|---|---|---|
| DeepSeek V4-Pro | $121.80 | $61.41 |
| Qwen3.7 Plus | $144.00 | $99.20 |
| MiniMax M3 | $216.00 | $148.80 |
| GPT-5.5 (for scale) | $2,200.00 | - |
Cached column assumes 140M of the 200M input arrives as cache hits (repeated system prompt and repo context), 60M fresh, 40M output. GPT-5.5 shown only for scale at $5/$30; cache not modeled.
Flat, DeepSeek runs about $22 a month under Qwen and $94 under MiniMax. Turn on caching and the gap stretches, because DeepSeek's cache read is the real outlier here: $0.003625 per million against Qwen's $0.08 and MiniMax's $0.12. That is 22x cheaper than Qwen and 33x cheaper than MiniMax for the exact tokens a coding loop reuses most. On a cache-heavy agent, DeepSeek's input line item nearly disappears.
The GPT-5.5 row is there to keep perspective. The most expensive of these three, MiniMax at $216, is still roughly a tenth of GPT-5.5's $2,200 on the same workload. The decision between these models is a rounding error next to the decision to leave the premium tier in the first place.
Two reasons to pay more anyway
Cost is not the only axis, and the two pricier models each buy you something DeepSeek will not.
Qwen3.7 Plus sees. It takes image and video input and posts 79.0 on ScreenSpot Pro, a GUI-grounding benchmark, which is the kind of thing you need for screenshot-driven agents and UI automation. DeepSeek V4-Pro is text-only. If your agent has to look at a screen, the $0.04 input premium and the higher output rate are the cost of admission, and there is no DeepSeek configuration that closes that gap.
MiniMax M3 ships open weights. MiniMax has said the weights are coming to Hugging Face and GitHub, so if your plan is to start on the hosted API and later self-host for data control or to escape per-token billing entirely, M3 is the only one of the three with that exit. It also accepts image and video. You pay for it on the meter: $2.40 output is the steepest here, and the priority tier runs $0.90/$3.60 if you need lower latency.
DeepSeek V4-Pro's counter is range. It carries 1M context with up to 384K output tokens, more headroom than most workloads will use, plus three reasoning modes (non-think, think-high, think-max) so you can dial cost against effort per request. For a text-only coding or extraction pipeline, that combination at $0.87 output is hard to argue with.
On quality, the benchmarks do not line up
This is the honest gap in any cheap-model comparison. Each vendor reports its own numbers on its own harness, so a clean head-to-head does not exist. Third-party runners like Artificial Analysis normalize part of it on a shared intelligence index, but they have not put all three through the same coding suites yet. What we have:
| Model | Reported coding / agentic scores |
|---|---|
| DeepSeek V4-Pro | SWE-bench Verified 80.6%, LiveCodeBench 93.5% (think-max, model card) |
| MiniMax M3 | SWE-Bench Pro 59.0, Terminal-Bench 2.1 66.0, BrowseComp 83.5 (vendor) |
| Qwen3.7 Plus | AA Intelligence Index 39, ScreenSpot Pro 79.0 (GUI grounding) |
SWE-Bench Pro and SWE-bench Verified are different test sets, so DeepSeek's 80.6 and MiniMax's 59.0 are not on the same scale. Treat the table as "what each vendor claims it is good at," not a leaderboard. DeepSeek leans hardest into raw coding accuracy, MiniMax into long-horizon agentic and browsing tasks, Qwen into anything that involves looking at a screen.
The practical move is the same one it always is at this price point: pick the cheapest model that clears your own eval, not the one with the prettiest self-reported chart. When inference costs a tenth of a premium model, you can afford to run two of them through your test suite before committing.
Picking one
Text-only coding or extraction at volume: DeepSeek V4-Pro. Cheapest output, near-free cache, the most output headroom, and reasoning modes to tune spend. This is the default unless you need something it cannot do.
Agents that read screens: Qwen3.7 Plus. The vision support and ScreenSpot grounding are the reason to accept its higher output rate, and it is still the second-cheapest meter of the three.
A self-hosting plan or multimodal plus open weights: MiniMax M3. You pay the most per token, but it is the only one with a path off the hosted meter once the weights drop.
Whatever you pick, the prices move constantly at this end of the market, and promo rates expire without much notice. Put your real token counts into a calculator before you commit, because the difference between a 3:1 and a 5:1 input-to-output ratio can change which of these three is cheapest for you.
Sources
- - DeepSeek API pricing (official): api-docs.deepseek.com
- - DeepSeek V4-Pro model card and benchmarks: huggingface.co/deepseek-ai/DeepSeek-V4-Pro
- - Qwen API pricing (Alibaba Cloud Model Studio): alibabacloud.com/help/en/model-studio
- - MiniMax M3 pricing and docs: platform.minimax.io
- - Artificial Analysis Intelligence Index: artificialanalysis.ai
- - TokenCost pricing page (all three models): tokencost.app/pricing