DeepSeek has released V4.1 Flash with a claim that reads like a contradiction: the cheap model in the family now outbenches the expensive one on code and maths. The weights are public under an MIT licence, the context window holds more than a million tokens, and independent testing places it sixth out of 113 models.
For Australian teams setting AI budgets for the rest of 2026, the benchmark table is not the interesting part. The price of the tokens is, and so is the licence attached to them.
What DeepSeek actually shipped
The published model card puts the backbone at 552 billion parameters, but only 8 billion of those activate per token while the model reads a prompt and 16 billion while it writes a reply. That is a mixture of experts design pushed further than most: large enough to hold broad knowledge, small enough per step to serve cheaply.
The context window runs to 1,048,576 tokens with a maximum output of 384,000. The licence is MIT, which means the weights can be downloaded, modified and run commercially without asking DeepSeek for permission or paying it anything.
That combination is what makes this release worth reading twice. Open weights at this scale usually arrive with a restrictive licence or a benchmark sheet that quietly avoids the hard tests. This one does neither.
Cheaper and better, but not everywhere
On DeepSeek’s own numbers, V4.1 Flash scores 90.9 on GPQA Diamond, 90.6 on Terminal Bench 2.1, 74.2 on DeepSWE v1.1 and 93.0 on GSM8K, with a Codeforces rating of 3471. Those are flagship figures from a model sold as the budget tier of its family.
The caveat sits in the same tables. On SimpleQA Verified, which tests factual recall rather than reasoning, V4.1 Flash scores 42.3 against 55.2 for DeepSeek V4 Pro. The cheap model reasons like the expensive one and remembers less.
That distinction decides where it belongs. Work that reasons from material you supply, such as code review or document extraction, plays to its strengths. Work that leans on the model already knowing things does not, and the gap is wide enough to notice in production.
The memory trick behind the price
The savings come from compression rather than a smaller brain. DeepSeek’s technical report describes cutting the persistent KV cache to roughly one eighth of the previous Flash model, and the global cache to 890 bytes per token, about a quarter of its predecessor. Against DeepSeek V1 the reduction is 437 fold.
Cache size is what makes long context expensive to serve. Every token in a conversation has to be held in memory, so a million token window normally means a server that can hold very few users at once. Squeeze the cache and the same hardware serves far more of them.
That is the whole commercial story in one line. The million token window is not a research demo here because the memory bill finally fits inside a price a provider can publish.
Why Australian businesses should pay attention
DeepSeek bills in US dollars, so what an Australian team actually pays moves with the exchange rate. At the Reserve Bank rate for 10 September the peak prices work out to about $0.42 per million input tokens and $1.66 per million output tokens. Off peak DeepSeek charges half of that, which takes a million tokens in and out from roughly $2 down to roughly $1.
The timing is where it gets local. DeepSeek sets peak hours at 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. In Melbourne that lands at 11am to 2pm and 4pm to 8pm, so both expensive windows sit inside the Australian working day while a European team spends most of its day on the cheap rate.
That is schedulable. The cheapest hours here are overnight and the gap between 2pm and 4pm, which is where batch work belongs: overnight document extraction, summarisation, anything that can wait. Interactive work in business hours pays double, and no amount of prompt tuning changes that.
The licence is the other half of the story. A model you can host yourself is a different compliance conversation from one you call across the Pacific, particularly for records covered by the Privacy Act 1988 where the live question is where the data physically sits. Running 552 billion parameters is not a laptop job, but for document automation on sensitive files the option now exists at all.

How to price it before committing
The useful test is not a benchmark, it is a bill. Take a week of real prompts, run them through the model, and count the output tokens rather than trusting the advertised rate. A model that answers in 900 words where a rival answers in 300 costs three times as much at the same headline price.
Verbosity is the number to watch, not the headline price. Artificial Analysis rates the model at 40 on its intelligence index, sixth of 113, and describes it as very verbose against its peers, having produced 250 million output tokens across the evaluation run. At peak Australian rates that output alone bills about $416, and about $208 off peak.
The second test is recall. Ask it the things your staff ask it, the ones that depend on knowing a rule or a name rather than reasoning from a document. That is where the SimpleQA Verified gap shows up, and it shows up as confident wrong answers rather than refusals.
What open weights change from here
Three open weight releases have landed in quick succession and the pattern is consistent: the frontier labs set the ceiling, and open models arrive months later at a fraction of the price. Kimi K3 made that point about hardware. V4.1 Flash makes it about cost.
For teams choosing a model this quarter the honest answer is that it depends on the job. Reasoning and code work now have a genuinely cheap option, while anything resting on recall still points at the larger models, and our breakdown of Claude vs ChatGPT remains the better starting point for general business workloads.
What has changed is that the cheap option is no longer obviously the worse one. For anyone selling AI by the seat rather than by the token, that is a harder question to answer this quarter than it was last.
