When you use AI models a lot, cost matters as much as performance. If you have a coding agent find a bug, fix the code, and repeat all the way through testing, it goes well beyond a few chats with the model.
And if you've used the AI agents that are popular these days, you've probably experienced tokens melting away fast when the context gets a bit long or the task gets even slightly complicated.
But after DeepSeek released V4.1-Flash-1017 on October 17, MiniMax officially announced M3.1 Flash on October 19. What the two companies put forward in common is clear. They want to handle coding and complex tasks better while cutting usage fees and latency.
If you thought from the name 'Flash' that these were lightweight chatbots, you need to look at these models a bit differently. What the two models aim for isn't an AI that ends at answering questions, but an AI agent that finishes a job through multiple steps.
The Era of Cheap Models Taking On Coding
DeepSeek's new model is named DeepSeek-V4.1-Flash-1017. It's a version updated on October 17 from the existing V4.1-Flash.
DeepSeek said scores rose in evaluations of reasoning, coding, agent tasks, and image understanding, and that the time it takes to finish a task and the length of answers also fell.
It stressed that this was an update that raised real task performance through post-training adjustment rather than building a bigger model from scratch.
✦ DeepSeek's Official Benchmark Changes
DeepSeek-V4.1-Flash-1017, per the official announcement
The official materials say the score on DeepSWE v1.1, an evaluation of analyzing and modifying code repositories, rose from 74.2 to 77.0. The score on Terminal-Bench 4.0, which tests carrying out development work in a terminal, also climbed from 31.2 to 42.3.
This is a truly remarkable improvement. Even if these figures are DeepSeek's own published evaluation results and so can't be 100% objective, they show anyway that this update isn't a mere model name change.
✦ Public Comparison Tables by Model
Scores per the two companies' public comparison tables
Shall we put the benchmarks side by side with other models? In the two companies' public comparison tables, DeepSWE v1.1 was 77.0 for DeepSeek, 73.7 for M3.1 Flash, 73.0 for GPT-5.6 Sol, and 74.0 for Claude Opus 5.
Terminal-Bench 4.0 came out at 42.3 for DeepSeek, 39.9 for GPT-5.6 Sol, 38.1 for M3.1 Flash, and 51.8 for Claude Opus 5.
Opus models have gotten a lot cheaper recently, but considering they're still premium models, the price-to-performance of DeepSeek and M3.1 Flash is excellent.
What drew attention were the usage examples DeepSeek showed. The official announcement features projects like making a web racing game, recreating a "Portal"-style game in a game engine, and composing a 3D scene.
These tasks are different from suggesting a single line of code. You have to interpret the requirements, create multiple files, check the results of running them, and fix things again if problems arise.
The more an AI carries out this process in sequence, the less effort the user spends giving instructions at every step. If you've used a lot of AI agents, you know how important this is.
For example, if you use Haiku, the cheapest of the Claude models, agent tasks keep getting interrupted. Even if you tell it not to stop and to keep going, it doesn't listen and stops partway through. This is really annoying.
You thought the agent was still carrying out the work you assigned, but it turns out it only did the first part of your instructions and has been idle for hours, and you can't help getting furious. With something like Sonnet, this problem is mostly solved.
But it's that much more expensive. If a low-cost model solves this problem, it's a really big advantage for AI agent users.
How Low Are the Token Prices?
✦ DeepSeek and MiniMax API Prices
Per million tokens, in dollars
DeepSeek's API prices are $0.15 per million input tokens in off-peak hours and $0.30 in peak hours.
Output tokens generated by the model are $0.60 and $1.20 per million respectively, and cache-hit input, which reuses input already processed, is $0.003 or $0.006.
Peak hours are Monday through Friday, 01:00 to 04:00 and 06:00 to 10:00 UTC, and off-peak rates apply on Chinese holidays and weekends. Converted to Korean time, weekday peak hours would be 10 a.m. to 1 p.m. and 3 p.m. to 7 p.m.
So to cut API costs, it's best to avoid those hours.
Here, a token isn't exactly the same unit as a character count. Think of it as the small pieces the model splits text into when processing sentences.
Input means your question along with the code, documents, and so on you hand to the model, and output is the answer or code the model generates. If you attach a big file and keep having it generate long code, costs on both sides will grow fast.
MiniMax M3.1 Flash also launched at a similar price point to DeepSeek. The official price of M3.1 Flash, released on October 19, is $0.15 per million input tokens, $0.60 for output, and $0.015 for cache input.
The base input and output prices match DeepSeek's off-peak rates. So during DeepSeek's off-peak hours DeepSeek is a bit cheaper, and during DeepSeek's peak hours MiniMax is likely cheaper, since DeepSeek's cache-hit input rate is lower than MiniMax's.
But even during DeepSeek's peak hours, if the work repeatedly sends long, similar prompts, it would be classified as cache-hit input (input that reuses values already cached), so DeepSeek might be cheaper in some cases.
Anyway, with two excellent models engaged in such a fierce price competition... it makes a consumer's heart swell.
✦ What MiniMax Announced
M3.1 Flash, per the company
What MiniMax emphasized is low cost and speed. The company said M3.1 Flash's online inference speed is nearly 3 times faster than the previous version, and that it cut the cost of agent tasks by 70%.
It also said output speed can exceed 200 tokens per second in everyday use, with a median of over 170 tokens per second.
✦ Major Competitors' API Prices
Per million tokens, in dollars, as of October 20, 2026
So how big is the price gap compared with the latest high-performance models from ChatGPT, Claude, and Gemini? Looking at the official API price lists as of October 20, 2026, OpenAI's flagship GPT-6 Astra is $10 per million input tokens and $50 for output.
Anthropic's Claude Fable 5.1 is the same at $10 input and $50 output. Google's Gemini 4 Argon is $2 input and $10 output as a launch promotion through November 30, but will rise to $4 and $20 respectively from December 1.
Even among models one tier down in price, good for coding agents, GPT-6.1 Sol and Claude Sonnet 5.5 are $2 input and $10 output.
Of course, what's meant here isn't the monthly subscription fees for the ChatGPT, Claude, and Gemini apps, but the pay-as-you-go rates you pay when calling the models directly through the API.
✦ Cost for the Same Usage
1M input + 200K output tokens, excluding cache and tool fees
The gap shows up better if you match token usage. Suppose several API calls add up to 1 million input tokens and 200,000 output tokens, excluding cache discounts, long-context rates, and separate tool-use fees.
DeepSeek's off-peak rate and MiniMax come to about $0.27. Even in DeepSeek's peak hours it's about $0.54.
For the same usage, GPT-6 Astra and Claude Fable 5.1 come to about $20, GPT-6.1 Sol and Claude Sonnet 5.5 about $4, and Gemini 4 Argon about $4 at the current promotional price.
In other words, DeepSeek's off-peak rate or MiniMax can process the same amount of tokens at about 74 times lower cost than the top models and about 15 times lower than the $4 models.
Even calculated at DeepSeek's peak-hour rate, the gaps are about 37 times and 7.4 times respectively.
Of course, the calculation above compares only token prices. Premium models usually handle the same task with fewer tokens, so in reality the gap could shrink. But even taking that into account, it's true that it's enormously cheap.
It also makes sense that AI infrastructure-related stocks fell after this new DeepSeek model launched.
The logic is that if consumers use models this cheap and this good, it gets that much harder for companies spending big money to expand AI infrastructure to grow sales and generate profit going forward.
Where Do the Two Models Differ?
✦ How They Are Released
Weights and access routes
The two models look quite similar on the price list, but how they're released differs. DeepSeek uploaded V4.1-Flash-1017's weights to a public repository, and it's also available through the official API.
MiniMax, on the other hand, began offering M3.1 Flash through the MiniMax API, M Plan, and MiniMax Code, and said it will release the model weights later.
If you're interested in downloading it right now and putting it on your own server, DeepSeek is the better fit.
Price competition is now changing the criteria for choosing a model itself.
Where it used to be natural to hand the hardest problems to the most expensive model, it's now possible to hand everyday coding and repetitive work to a cheap model and use a top-tier model only for the truly tricky problems.
Of course, since an agent doesn't finish in one answer but repeats tool calls and fixes, finishing the job quickly and cutting unnecessary retries could make a bigger difference than a slightly lower price per token.
So it's only natural that attention is turning to DeepSeek's and MiniMax's new models, which cut the per-token price while reducing unnecessary costs.
Will Cheap Models Change How We Use AI?
The range of work cheap LLMs can take on is widening. Where it used to be simple code edits or writing short functions, there's now a clear trend of trying to hand tasks that modify multiple files and even run tests to low-cost models.
The direction shown by DeepSeek V4.1-Flash-1017 and MiniMax M3.1 Flash is clear.
Rather than waiting for a single answer from expensive models like GPT-6 Astra or Claude Fable, it's more realistic to complete the result by running a moderately strong model cheaply many times.
Right now we're using premium models like ChatGPT, Claude, Grok, and Gemini freely under flat-rate plans, but many expect flat-rate plans to disappear in the future. Those companies are suffering enormous losses because of flat-rate plans.
(If you found out how much the tokens you freely used under a flat-rate plan would have cost if called through the API, you'd likely be shocked.)
Then pricing will gradually shift to API plans where you're charged by the tokens you use, and as anyone who has recently used Muse AI would have felt, even a seemingly huge amount like 100 million tokens can melt away in an instant depending on the case.
In this situation, the emergence of cheap Chinese models like DeepSeek and MiniMax is welcome news for consumers like us. Though it's probably not very pleasing to AI infrastructure companies. Ehehe♡
That's all for this post. See you next time.
- Ollama's New Q2 Quantization Support: Quantization Explained
- FLUX 3 Image Klein vs Qwen Image 2.1: Preserve or Blend?
- Nano Banana 2.1: 4K Images at $0.113 Each, Worth It?
🌐 Korean version: 딥시크 V4.1 플래시-1017에 이어 미니맥스 M3.1 플래시까지.. 중국산 가성비 AI 모델이 쏟아지네
No comments:
No comments yet. Want to leave the first one?
Post a Comment