MiniMax M3.1 Flash vs DeepSeek V4.1 Flash vs Mercury 3

Between October 17 and 19, 2026, three cheap, fast models aimed at agent work arrived. DeepSeek shipped V4.1 Flash 1017, an update with MIT-licensed weights. MiniMax released M3.1 Flash, a next-generation Flash model that reads images and video. Inception launched Mercury 3, a diffusion language model that it says writes over 1,000 tokens a second. MiniMax and DeepSeek list the same price, $0.15 per million input tokens and $0.60 per million output, and Mercury 3 undercuts both while its launch discount lasts.

We covered the first two on their own: DeepSeek V4.1 Flash 1017 and its search API, and MiniMax M3.1 Flash's price and benchmarks. This post compares all three, with OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 as reference rows, because Inception measures Mercury 3 against them and they sit in the same price band.

Prices come from each vendor's own page, speed measurements from OpenRouter, and benchmarks from each vendor's launch materials, all read on October 19. Where a figure is a claim rather than a measurement, we say so.

All three are built for the high-volume steps of agent work, where a model is called many times per task and every cent and second compounds. They differ in what they read, whether you can download them, and how their prices move.

The weights row settles the question for anyone who must self-host: only DeepSeek's are available today, under one of the most permissive licences there is. The inputs row settles it for agents that read screenshots or recordings: MiniMax is the only one of the three that takes video, and Mercury 3 takes text only.

A single price per million tokens hides how each of these is actually billed. DeepSeek's price doubles between 01:00 and 04:00 and 06:00 and 10:00 UTC on weekdays. Inception's Mercury 3 price is 50% off at launch with no stated end. Claude Haiku 5.5 and GPT-6 Luna both charge more for long prompts. Only MiniMax is flat.

Two things stand out. The big-lab models are not more expensive at this tier: GPT-6 Luna and Claude Haiku 5.5 list lower input and output prices than MiniMax or DeepSeek off-peak, as long as prompts stay short. And DeepSeek's cached input is five times cheaper than MiniMax's, which matters more than the headline for agents that re-read the same context on every turn. Our Haiku 5.5 post explains its long-prompt tier.

To make the price states concrete, take one typical agent step: 80,000 input tokens, of which 60,000 are cached context and 20,000 are new, and 4,000 output tokens. The figures below are our arithmetic on each vendor's listed prices, per 1,000 such steps, and assume every model uses the same number of tokens. In practice they will not, which is the point of the trial in section 08.

Under identical token counts, the five cheapest options sit within a few dollars of each other per thousand steps, and the spread is driven by price states, not by headline rates. The failure case is choosing on this table. A model that needs two attempts doubles its row; one that writes twice the output adds its output cost again. MiniMax defaults to its maximum effort, for example, which will raise its output count unless you set it lower.

Speed is the selling point of all three, and the claims do not match what a third party measured on launch day. The table puts each vendor's claim next to OpenRouter's best-provider median throughput from its model pages on October 19.

Read this as a range, not a ranking. Throughput on a router depends on the provider behind it, load, output length and reasoning effort, and two of these routes had only days of data. What it does suggest is that the vendor figures are best-case numbers from the vendor's own serving setup. Measure latency on your own prompts, through the route you will use in production.

Mercury 3 works differently from the other two. Most language models write one token after another. A diffusion model drafts a whole block of text at once and refines it over a few passes, which lets it produce many tokens in parallel. Inception's launch post says Mercury 3 is more than 40 times larger than its first model, runs at 1,068 tokens a second on NVIDIA GPUs, and matches or beats GPT-6 Luna and Claude Haiku 5.5 while running four to eight times faster. Its benchmark chart is an image, so we cannot reproduce the scores.

The context window grew from 260,000 tokens in Mercury 2.5 to one million, with up to 131,072 output tokens and four effort levels, from instant to high. Inception's models page adds a deadline: Mercury 2 and 2.5 are no longer available to new users, and existing customers with usage before October 19 keep access until November 15, 2026; enterprise customers stay on their agreements. We covered the previous version in our Mercury 2.5 post.

Partner quotes are partner-run tests, not independent benchmarks, and Inception chose which to publish. They do point at where diffusion speed matters: voice agents, real-time search, support chat and browser agents, where a model is called on every step and users notice each second. The failure case is choosing Mercury for a batch job where nobody is waiting; there, speed buys nothing and the price after the discount is what counts.

Only two benchmarks carry a score in both MiniMax's and DeepSeek's materials (MiniMax names Automation-Bench but prints no score), and Inception's benchmark chart is an image. DeepSeek ran DeepSWE v1.1 in the Mini-SWE-V2 harness; MiniMax does not say which harness it used. Read the two columns as separate self-reports.

On both shared rows DeepSeek reports the higher figure. Beyond them, each vendor emphasises different strengths: MiniMax reports computer use (61.7 on OSWorld V2) and security (82% on CyberGym); DeepSeek reports terminal work (42.3 on Terminal-Bench 4.0) and business automation (60.5 on Automation-Bench). For a trial, pick the benchmark row closest to your own work and then ignore it in favour of your own tasks.

The routing below is ours, from the inputs, weights, pricing and vendor figures above. It decides which two to trial, not which one to deploy.

None of these vendors' numbers will tell you what a finished task costs in your system, because token counts, retries and reasoning length differ by model. A short trial does. Take twenty real tasks from your agent's logs. Run each model with the effort level set explicitly, on the route you would use in production. Record input, cached and output tokens, wall-clock time, retries and whether the task finished correctly. Divide spend by finished tasks.

Two settings decide most trials before they start. Set MiniMax's reasoning effort below its default of max for routine steps, or it will spend output tokens you did not plan for. And keep DeepSeek runs out of its peak windows if your production schedule will, or the comparison will mislead in the other direction. Our model price index keeps the per-token side current; the per-task side is yours to measure. If you want help running it, that is part of our AI transformation work.