DeepSeek released V4.1 Flash on 10 September 2026, covered in an AI Builder Club community post by Kai Wang. The headline in most of the coverage was the benchmark: 74.2 on DeepSWE v1.1, against Claude Opus 5 at 74.0. That is a 0.2 point difference, and it is being read by some buyers as a changing of the guard.
It is not. A 0.2 point gap on a single eval, published by the vendor whose model came out on top, is not a result. It is a tie with a press release attached. Worth saying plainly before an Australian business reorganises a stack around it.
What the numbers actually are
Every figure below is self-reported by DeepSeek on its own DeepSWE v1.1 evaluation. We have not independently verified them, and neither, as far as we can tell, has anyone else at the time of writing.
| Model | Reported score | What it supports saying |
|---|---|---|
| DeepSeek V4.1 Flash | 74.2 | Front of the pack on this eval, by a margin smaller than most measurement noise |
| Claude Opus 5 | 74.0 | Statistically indistinguishable from the above on this eval |
| DeepSeek V4 Pro (prior) | 62.7 | The genuinely large movement here, an 11.5 point jump on DeepSeek's own previous model |
Notice which of those three rows is the real story. The interesting number is not 74.2 against 74.0. It is 74.2 against DeepSeek's own 62.7. A vendor closing an 11.5 point gap on itself in one release says something about the rate of improvement in open models. The 0.2 says nothing at all.
Should a 0.2-point benchmark gap change which AI model we use?
No, and treating it as a procurement signal is the mistake worth avoiding. A difference that small on a single evaluation, published by the vendor that came first, sits inside the range you would expect from run-to-run variation, prompt formatting and scoring choices. It is not evidence that one model does your work better. The things that do change the answer for an Australian business are support arrangements, where data is processed, how the model behaves on your own tasks, and what the total cost looks like once integration and oversight are counted.
The architecture change is more interesting than the score
V4.1 Flash uses what DeepSeek calls a Causal-Encoder-Decoder design, a 552B parameter mixture-of-experts model that splits compute unevenly between reading and writing: roughly 8B active parameters for processing input and 16B for generating output. Model weights are open.
That split matters for one specific workload. Coding agents spend most of their time reading, not writing. They pull in files, read tool output, scan diffs, and then produce a comparatively short edit. A design that allocates compute differently to those two phases is a sensible response to how agentic work actually behaves. Whether it pays off in practice is a question for your own tasks, not for a leaderboard.
Available through the API as deepseek-flash.
Also runnable through the DeepSeek Harness, invoked with npx @deepseek-ai/dsh web.
Open weights, which means self-hosting is on the table if you have the infrastructure and the appetite to run it.
Published token pricing splits off-peak and peak, with off-peak defined as weekdays 01:00 to 04:00 and 06:00 to 10:00 UTC.
Published rates per million tokens run $0.003 to $0.006 cached input, $0.15 to $0.30 uncached input, and $0.60 to $1.20 output across those two windows.
Read the off-peak window carefully if you are in Sydney or Brisbane. Weekdays 01:00 to 04:00 UTC is mid-morning to early afternoon in eastern Australia, so part of the cheaper window does land inside the local business day. The 06:00 to 10:00 UTC window lands in the evening. Whether that is worth designing around depends on whether your workload can wait, and most interactive workloads cannot.
Why token price is the wrong headline for an Australian buyer
Published per-token rates are the most visible number and usually the least decisive one. In our client work the token bill is rarely the largest line in an AI program. The larger lines are the integration effort, the review process around anything the model produces, the time someone spends maintaining prompts and context, and the cost of an incident when output goes out unchecked.
A typical mid-market automation engagement we run in Australia lands between $40,000 and $90,000 across build, governance and handover, against model spend that often sits in the hundreds of dollars a month. Halving the token rate moves a small number. Choosing a model your team cannot get support on, or that processes data somewhere your risk register will not accept, moves a large one.
We have written about the pricing trap before in the real cost of cheap AI for Australian businesses, and about the data and quality side in Claude versus DeepSeek on cost, quality and data questions. If you want to put your own numbers against it rather than ours, our ROI calculator is the faster path.
What to do with this release
Treat it as useful information about where open models have got to, and as a prompt to run your own evaluation rather than to accept someone else's. The practical version of that is small: pick ten tasks your team actually does, run them through both models, and have the person who owns the output judge the results. That takes a few days and it tells you something a benchmark cannot.
What it will probably tell you is that on most business tasks the two are close enough that the decision gets made on everything except capability. Which is the same conclusion the 0.2 points were always going to support, just arrived at honestly. For more on how much weight a benchmark headline deserves, see what reasoning benchmarks do and do not tell Australian buyers.
