DeepSeek vs Opus: the numbers do not add up.

Tale of the tape

We believed him too. So we rented the box. Here is what happened.

The dream: run DeepSeek V4.1 Flash on our own rented GPUs. Point Claude Code at it. Run 64 to 128 robot coders at once. Save a fortune.

The model is huge: 763B parameters, including a 197B memory module, with a 1M-token context.

8x H200 boxes were sold out all morning, so we grabbed a 4x H200 box. Its real price (the kind they cannot take back) is $18.37 an hour. We paid the spare-capacity (spot) price, $9.19 an hour: about half, because the provider can take the box back.

Open source means the weights are free. Free like a free puppy.

The puppy comes with a ~511 GB download, a box to keep it in at $18.37 an hour ($9.19 on a spot discount), 5 starts before it stayed up, and somebody on call who speaks fluent GPU.

Round 1: DeepSeek, on the sticker. $0.60 per million written tokens is a tiny number.

Does not add up, part 1: the weights are free. The box to hold them is $18.37 an hour ($9.19 on spot).

Every start meant loading the model back in: about 10 to 15 minutes, every single time.

• Start 1: out of GPU memory. A speed trick called "adaptive verification" ate 41 GB per card.
• Start 2: froze while recording its GPU shortcuts. We sent deep-research robots. They traced it to that same adaptive-verification setting.

The owner, mid-freeze: "dont restart anything let it go"

• Start 3: crashed on a memory setting that stops the cards from talking to each other.
• Start 4: it worked! Then we threw 256 robots with long answers at it. Crash. We found the box's limit.
• Start 5: 128 robots max, 90% memory. Stable. Finally.

Bonus pain: a progress bar that said "48/48 files" and meant nothing. And our own script matched itself and killed the wrong thing.

We did. Five times.

Self-hosting is not a download. It is a job. You need a GPU whisperer on call.

Round 2: Opus 5.5. It comes with our Claude Code subscriptions: no box to rent, no 511 GB download, no restarts.

Does not add up, part 2: 5 starts, 10 to 15 minutes a start, before it stayed up under load.

Once it ran, we hit it with one-minute full-load tests, one kind of work at a time. Zero errors.

All 4 GPUs at 100%, about 600 W each.

Those numbers are real, and each one was measured alone. Real coding work does not come one kind at a time.

Round 3: DeepSeek. Raw muscle.

Does not add up, part 3: 5,281 written tokens a second sounds huge. Now watch what real work does to it.

• New input (blue): text the model has never seen.
• Cache (green): re-reading the conversation it already saw.
• Output (orange): what it writes.

Coding robots re-read the old conversation over and over. In our real Claude Code September, 96% of all input was re-reading old conversation.

For every 1 token Claude writes, it reads 41.6 new tokens and 1,042 old ones.

One average call re-reads about 196,000 tokens. A small novel, every call.

Every token costs box time.New: 60 microseconds.Re-read: 1.9 microseconds.Written: 189 microseconds.

Run our real mix through that, and 4x H200 top out at about 213 written tokens a second. About 20 billion tokens a day, in total.

We checked the formula against the live test. It matched within 3%.

Our busiest day in September was 51 billion tokens. The box tops out at about 20 billion a day.

Each re-read is cheap, buddy. But we do 1,042 of them, plus 41.6 brand-new tokens at 60 microseconds each, for every 1 token written. That is about 4,700 microseconds of box time per written token: about 213 a second, for the whole box.

Four H200s, pinned at 100% and about 600 W each, top out around 213 written tokens a second on real coding work. For everyone on the box, combined.

Opus 5.5 wrote about 85 tokens a second for EACH call.

Round 4: Opus 5.5.

Does not add up, part 4: 4 H200s, about 213 written tokens a second, shared by everyone.

DeepSeek also sells the same model by the token, on its own cloud.

The box costs $441 a day at the real price ($18.37 x 24 = $440.88), busy or not. On the spot discount we got: $220 a day.

A full, 100%-busy day of box work is worth $184 to $223 at DeepSeek's own per-token prices.

At the spot price, running flat out all day, the box only ties DeepSeek's own cloud. At the real price it costs 2x to 2.4x as much ($440.88 / $223 = 1.98; $440.88 / $184 = 2.40).

The spot discount is the only reason the box even tied. And you cannot run a business on a box that can be taken back. Ours was, minutes after the last test.

And that is before the restarts, the waiting and the expert on call.

DeepSeek's cloud prices per 1 million tokens, off-peak, are $0.15 new, $0.003 cached and $0.60 written. They double at weekday peak.

We rented it, on a discount that can be taken back. At the real $18.37 an hour, the cloud does the same work for about half the money or less, and nobody has to babysit it.

Round 5: DeepSeek's own cloud. The box only ties it on a spot discount.

Does not add up, part 5: $441 a day at the real price, for $184 to $223 of work. Even on the $220 spot discount, at most a tie.

80x CHEAPER

That 80x compares DeepSeek's price per token to Claude's posted list price. Nobody on a Claude Code plan pays list price.

Claude Code plans are 27 to 35 times cheaper than Claude's posted pay-per-token price. That is the owner's figure.

Divide the 80x by that, and most of the gap is gone.

Our real September (Sep 1-27, one server): our Claude Code subscriptions cost about $5.5K. The same usage on DeepSeek's cloud: $3.5K off-peak, $7.0K at peak.

So DeepSeek's cheap cloud lands somewhere between a bit cheaper and more expensive. Calling that "80x cheaper" is not a saving. It is gaslighting.

Now count what actually ships. Our Claude subscriptions came to about $1 per merged code change in September: 5,610 changes merged, 5 reverted.

Even on DeepSeek's cheap cloud, our estimate is $1.15 to $4.90 per merged change, after all the extra checking it needs.

Getting it right the first time is the cheapest token there is.

Round 6: Opus 5.5, per shipped change.

Does not add up, part 6: "80x cheaper" per token turned into about the same or worse per change that actually ships.

It is $0.60 per MILLION written tokens. And who cares, when the first reply takes 12.8 seconds (56.6 on a bad one) and it writes 8 tokens a second on the shared box?

Opus answered faster while reading conversations 20x longer.

Brains: on public test scores for hard terminal coding, DeepSeek Flash gets about 31. Opus 5.5 gets about 66.

One of those is a passing grade.

Round 7: Opus 5.5. Faster and sharper.

Does not add up, part 7: 8 tokens a second per call. A cheap token you sit and wait for is not cheap.

The plot twist: the reviewers of our robot sandbox kept finding ways robot code could escape.

One example: a settings file dropped in a shared temp folder could make robot code run as the admin.

So the builder robots never got switched on.

Fair is fair: during the run, 48 to 64 read-only reviewer robots read 2,377 folders of our code and filed 32 bug reports.

Code shipped by DeepSeek: zero lines. Code changes merged on our Claude subscriptions in September: 5,610.

After about 3 hours (about $28 on the spot discount, about $55 at the real price), the spare-capacity box was taken back by the provider, minutes after the last test.

Even the GPUs left us.

The official verdict, from the owner: "it sucked ass"

We went back to Opus 5.5 on high.

Want the subscription side priced out? The AI Career Lab ran the numbers: Can a $10 AI stack beat a $200 plan?

(ZOMG Guy has left the chat.)

Round 8: Opus 5.5, by knockout.

Does not add up, part 8: 763B parameters. Zero lines shipped.