DeepSeek V4 Flash
DeepSeek V4 Flash charges $0.14 per million input tokens and $0.28 per million output, which makes its output close to ninety times cheaper than Opus 5. On the WebDev Arena leaderboard it scores 1585, the lowest of the six candidates in our table.
- $0.28 per million output tokens
- 384k output ceiling
- 1M token context window
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
It is cheap, and here is how far cheap goes
Opus 5 charges $25 per million output tokens. DeepSeek V4 Flash charges $0.28. On input it is $5 against $0.14, about thirty-six times. This is not a discount, it is a different price class.
The same table holds a second number. On the web-dev leaderboard DeepSeek scores 1585 and Opus 5 scores 1692: a gap of 107 points, and the lowest figure in the table. The trade is plain and we are not hiding it. Less money, weaker answers on the hardest work.
A 384k output ceiling, a number few comparisons print
Most tables print the context window and skip the output ceiling. For a lot of jobs the output ceiling is what stops you, not the input window. DeepSeek V4 Flash returns up to 384,000 tokens in one answer; the Anthropic models stop at 128,000. Three times as much.
Where that matters is work whose input and output are both long: rewriting a long document, translating a large file in one pass, generating a dataset. With a 128k model you have to cut the text into pieces and stitch it back, and the stitching is exactly where quality breaks.
A rate fifty times cheaper that is not in our price column
DeepSeek publishes two separate input rates: $0.14 when the prompt is not cached and $0.0028 when it is. The difference is exactly fifty times.
We left the cached rate out of the price column. No real workload is one hundred percent cache, and a number that only holds in the best case becomes a false claim once it is sitting in a comparison table. If your system has a long fixed prefix that every request starts with, your real bill lands somewhere between the two, and working out where is your job rather than ours.
Twenty points out of a hundred, all from one column
DeepSeek totals 20 in the coding table and every point of it comes from the score-per-dollar column, where its 5660 is the highest figure in the table and takes the full twenty points of weight. It scores zero in the forty-point leaderboard column, because it is the floor of that column. Ten points for the general text board and five for Iran access are empty too.
So this model wins one column outright and loses the biggest one outright. Reading the ranking from the final number alone gives you the wrong picture; you have to look at where the points came from. That is why every row in the table shows its own per-column breakdown.
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- not checked
- Payment
- not checked
- Free tier
- no
How we checked: DeepSeek publishes no supported-countries list, so the access column stays unchecked and scores nothing.
What it is good at
- The cheapest model in the coding table by a distance: $0.14 in and $0.28 out
- A 384,000 token output ceiling, three times the Anthropic models
- A one million token context window, level with the dearest models in the table
- First in the score-per-dollar column at 5660, taking all twenty points of it
Where it falls short
- On the web-dev leaderboard its 1585 is the floor of our table, 107 points below the top.
- We have not read its row on the general text leaderboard, so ten points of the table stay empty.
- DeepSeek publishes no supported-countries list, so the Iran column is unchecked.
- It scores zero in the tooling column: no official agent tooling or editor extension is listed.
- The $0.0028 rate applies only to a cached prompt and is not a real number for an ordinary workload.
Our take
For high-volume repeated work whose output you review afterwards anyway, this is the cheapest sensible route on the table. For complex code that has to be right the first time, no discount closes a 107 point gap. We reach for it on bulk jobs, not for the core of a product.
Questions people actually ask
What does DeepSeek V4 Flash cost
Fourteen cents per million input tokens and twenty-eight cents per million output tokens. When the prompt is cached, input bills at $0.0028.
Is DeepSeek good for coding
For bulk work yes, for hard code no. It scores 1585 on WebDev Arena, the lowest of the six candidates in our table and 107 points under Opus 5.
What does a 384k output ceiling mean
It means a single answer can be that long. It is not the same as the context window: the window caps what you send, the output ceiling caps what comes back.
Does DeepSeek work from Iran
Unchecked. DeepSeek publishes no supported-countries list and we have no independent measurement from an Iranian connection.
Sources
- DeepSeek API pricing vendor source 12 August 2026