DeepSeek kept the price at $0.14 and added ten index points. Eleven models now tie on intelligence, and the bill for proving it runs from $72 to $1,061.

The rate card did not move

DeepSeek published the weights at 07:30 UTC on July 31 under MIT, ungated, and updated the API docs without renaming anything. The model ID is still deepseek-v4-flash. If you had traffic pointed at it last week, you are on the new build now whether or not you noticed.

DeepSeek's own changelog is unusually blunt about what it did: the architecture and size are unchanged, and the model was "only re-post-trained". No new pretraining run, no new parameter budget, no new price. The gains it claims are concentrated in agentic work, and it added native Responses API support plus Codex compatibility, which is the part that actually decides whether you can drop it into an existing harness this afternoon.

One number worth pinning down because the coverage disagrees with itself. The safetensors index on Hugging Face totals 304,180,418,494 parameters. Artificial Analysis publishes 284B total and 13B active, and DeepSeek states neither in its own README. The likeliest reconciliation is that the checkpoint carries auxiliary modules that are not served at inference, which would make both numbers true of different things. If you need one you can point at, the safetensors total is the only figure with a primary source under it.

Eleven models, three points apart, sixty-four times the price

This is the part that should change how you shop. Take every live model configuration Artificial Analysis currently scores between 48.95 and 51.60 and put the rate cards next to each other. The intelligence column barely moves. The price column moves by a factor of sixty-four.

Two deprecated configurations also land in this band, GPT-5.4 at xhigh and GPT-5.5 at medium, and we have left them out because you should not be buying either today. The reasoning effort labels matter as much as the model names here. Luna appears twice because turning its effort down from max to xhigh costs two index points and saves 44% of the money. Opus 5 only reaches this band at low effort, which is worth remembering when someone quotes its full-effort score in the same breath as its price.

A 2.65-point spread is well inside the range a single benchmark revision could scramble. Treat the ordering inside this table as noise and the price column as signal.

Where the cheap number actually comes from

Artificial Analysis published a specific claim alongside the score: even after OpenAI cut Luna by 80% on July 30, DeepSeek V4 Flash 0731 comes in around 60% below Luna on cost per task. The figure we can check independently agrees: running the whole index costs $72.02 against Luna's $190.87, which is 62% lower. The per-task numbers behind AA's own claim are published rounded to the cent, so treat "about 60%" as the honest resolution rather than anything finer.

The mechanism is the cache discount, and it is the most aggressive in the industry. DeepSeek bills a cache hit at $0.0028 against $0.14 for a miss. That is 98% off, where OpenAI, Google and Anthropic all sit at 90%. On a conversation-heavy or agentic workload where the same system prompt and history get resent thousands of times, that eight-point difference compounds into most of the gap.

Here is the trap. The blended price shown on Artificial Analysis model pages weights cached input, fresh input and output at 7:2:1, which is a heavily cache-optimistic assumption. DeepSeek reads $0.06 there. On a conventional 3:1 input-to-output blend with no caching at all, the same model is $0.175, close to three times as much. Both are correct. We have seen the two quoted in a single paragraph more than once this week. If you are comparing against a number you found elsewhere, check which blend it used before you conclude anything.

DeepSeek pays for it in tokens

Every cost table that assumes two models emit the same number of output tokens for the same job is lying to you, and it lies hardest about exactly this model. Artificial Analysis needed about 206 million output tokens to push its suite through DeepSeek. Claude Opus 5 at low effort needed roughly 12 million. GPT-5.6 Sol at low needed under 7 million. Same evaluations, same questions, and DeepSeek wrote seventeen times what Opus did and better than thirty times what Sol did.

So run the comparison both ways. First the naive version, a flat 200M input and 40M output per month with no caching, which is how most people build a budget:

Now the honest version. Artificial Analysis bills what each model actually emitted to complete identical work, so verbosity is priced rather than assumed away. On that basis DeepSeek costs $72.02 and Opus 5 costs $556.06. The advantage is 7.7x, not 51x. Against Gemini 3.6 Flash you keep about a tenth of the bill rather than a fifteenth, and against Luna at max effort the saving thins to roughly 62%.

DeepSeek is still the cheapest model in the band by either measure, so this is not a debunking. But 7.7x is a different decision from 51x, and the second number is the one that will end up in somebody's slide deck. Verbosity also costs you wall-clock time and context headroom, neither of which appears on an invoice. Artificial Analysis ranks this model 36th of 101 open-weights models on verbosity, against a 100M-token median for that same cohort.

The minus sixteen needs explaining

Two figures on the Artificial Analysis page look alarming and are going to be misread. The Omniscience Index is -16 and the hallucination rate is 84%. Neither means what a quick glance suggests.

It is worth reading how Artificial Analysis defines these two before quoting either. Hallucination rate is the proportion of incorrect answers among all non-correct responses. Questions the model got right are excluded from the denominator entirely. Published accuracy is 37% and the index is correct minus incorrect, which implies roughly 53% answered wrongly and about 10% declined. That reconstruction reproduces both printed figures exactly: 53 divided by 63 is 84%, and 37 minus 53 is -16. What is not true is that 84% of its answers are wrong. Negative scores here are the norm rather than the exception, and only a handful of models sit above zero at all.

What the pair actually tells you is that this model almost never abstains. Ten percent is a low refusal rate, and a model that guesses instead of declining is a poor fit for factual lookup without retrieval grounding. For code, transformation and agent work, where a wrong answer fails loudly against a test or a tool call, it matters much less. The rest of the scorecard is strong: GPQA Diamond 91%, Terminal-Bench 2.1 at 79%, AA-LCR 66%, and a GDPval-AA v2 Elo of 1559.

Worth flagging that DeepSeek self-reports Terminal-Bench 2.1 at 82.7 while Artificial Analysis independently measured 79. A 3.7-point vendor-favourable gap is unremarkable as these things go, but the independent figure is the one we would plan against.

A surcharge with no start date

DeepSeek's pricing page carries a notice that its API "will soon adopt a peak/off-peak pricing policy" under which peak prices are double the regular ones and it applies to all billing items. The stated windows are 09:00 to 12:00 and 14:00 to 18:00 Beijing time, which is 01:00 to 04:00 and 06:00 to 10:00 UTC, seven hours a day. The page also says the effective date is subject to a future announcement.

So it is not being charged today, and every figure in this post reflects the rates actually in force. But it is written down, which is more warning than most providers give. If it switches on, DeepSeek becomes $0.28 and $0.56 for those seven hours, and the 3:1 blend moves from $0.175 to $0.35. Note what that does and does not change: at $0.35 it is still the cheapest model in the index-50 band, because the next cheapest is Luna at $0.45. Batch and asynchronous work can simply be scheduled around the windows. Synchronous user-facing traffic in European or Asian business hours cannot.

One other thing that disappeared quietly. The old 16:30 to 00:30 UTC off-peak discount belonged to V3 and R1 and is gone. There is no discount window on V4 any more, only a surcharge window waiting for a date.

Which jobs we would hand it

High-volume asynchronous work is the easy call. Batch classification, bulk extraction, repository-wide refactors, anything where you control the schedule and nobody is watching a cursor blink. DeepSeek wins that on price by a margin nothing else in the band approaches, the 98% cache discount rewards exactly the repeated-prefix shape those jobs have, and the MIT weights mean a pricing change is an inconvenience rather than a migration.

Interactive traffic is where we would still pay more. The verbosity that makes DeepSeek cheap per token makes it slow per answer, and Artificial Analysis has no first-party throughput measurement for this build at all, so you would be buying latency blind. Luna at xhigh effort is the natural hedge at $0.45 blended and 49.07 on the index, from a provider with published speed floors.

And if your workload is dominated by output tokens rather than input, run the arithmetic before switching. A terse model at a high rate can genuinely beat a verbose model at a low one, which is the whole reason the 51x and the 7.7x in this post disagree. Put your own token mix into the calculator rather than trusting anyone's blended average, including ours.

Sources

• DeepSeek: API pricing - $0.14 cache miss, $0.0028 cache hit and $0.28 output per 1M tokens, the 1M context and 384K max output, the 2,500 concurrency ceiling, and the verbatim notice that peak pricing at 2x "will soon" apply during 09:00-12:00 and 14:00-18:00 Beijing time with the effective date still to be announced
• DeepSeek: API updates - The July 31 entry confirming the architecture and size are unchanged and the model was only re-post-trained, plus native Responses API support and Codex adaptation
• Hugging Face: DeepSeek-V4-Flash-0731 - MIT licence and ungated weights, the 07:30 UTC July 31 publication time, the 304,180,418,494-parameter safetensors total that contradicts the widely repeated 284B, and DeepSeek's own benchmark table including Terminal-Bench 2.1 at 82.7
• Artificial Analysis: DeepSeek V4 Flash 0731 scores 50 - The independent Intelligence Index of 50, the claim that cost per task lands about 60% below GPT-5.6 Luna even after the July 30 cut, and the attribution of that gap to the 98% cache discount against an industry norm of 90%
• Artificial Analysis: model comparison data - Every index score, cost-to-run-index figure and output-token count in the tables above, including DeepSeek's ~206M tokens against Claude Opus 5 low at ~12M and GPT-5.6 Sol low at under 7M. Note the rank and verbosity percentiles quoted for DeepSeek are drawn from the 101-model open-weights cohort, while closed models are ranked against a larger field
• Artificial Analysis: Omniscience methodology - The definitions that make -16 and 84% reconcile: hallucination rate is incorrect answers as a share of non-correct responses, and the index is correct minus incorrect on a -100 to 100 scale
• OpenAI: API pricing - GPT-5.6 Luna at $0.20 / $0.02 cached / $1.20, Terra at $2.00 / $12.00 and Sol at $5.00 / $30.00, the rate cards behind four of the eleven configurations in the band
• OpenAI: API changelog - The July 30 entry recording that Luna fell 80% and Terra 20%, which is the cut DeepSeek's new score arrived one day after
• Google: Gemini API pricing - Gemini 3.6 Flash at $1.50 / $7.50 with $0.15 cached and Gemini 3.5 Flash at $1.50 / $9.00, both with no context-length tiering and a single output column that includes thinking tokens
• Anthropic: Pricing - Claude Opus 5 at $5 / $25 with $0.50 cache hits, the card behind the $556.06 cost-to-run figure and the 90% cache discount DeepSeek undercuts