DeepSeek's tokenizer counted a test image at roughly 500 prompt tokens; Google's counted the same image at roughly 1,150 [13], and neither number shows up on either company's price page. Run those through the input rates and one page costs about $0.00011 on DeepSeek against about $0.00086 on Gemini, near eight times cheaper before the model has said a word [2]. For extraction work where the answer is four fields and a flag, that ratio is the entire business case.
The output side runs the other way. On the invoice audit DeepSeek billed 3,467 completion tokens for an answer of a few paragraphs, which the reviewer at The New Stack reads as internal reasoning billed as output [12][14]. That is a lot of deliberation about a monitor arm. Price the call with image tokens standing in for input: DeepSeek lands near $0.0024, Gemini near $0.0044 [3]. So a sticker gap of 3.4x on input and 5.7x on output [1] compresses to about 1.8x on that call, and at the doubled weekday rate DeepSeek's version costs roughly 9 percent more than Gemini's [3].
The sticker ratio only reaches your invoice if two conditions hold. Completion length has to stay bounded, which means capping output tokens and accepting that a reasoning-heavy model will sometimes hit the cap mid-answer. And the traffic has to sit outside the peak window. Both prices here are gateway prices on OpenRouter, which DeepSeek's model only reached on August 27 [2][5].
Clock time tracks the tokens: 30.5 seconds against 7.9 on the same audit [12], about 3.9x [4]. The reviewer calls that the only anomaly across the three tests [12], so the other calls were closer. Google markets Gemini 3.7 Flash as its "most intelligent workhorse model yet" [7]; on the two tasks whose results are reported, the workhorse tied a budget model on accuracy and won on latency.
The accuracy read is thinner than the pricing read. On the dual-axis chart, where revenue ran 0 to 12 on the left and costs 0 to 16 on the right so the cost line never visually clears the bars, both models answered Q1, named Subscriptions, and estimated $36.1 million, matching the source data [9]. On the invoice DeepSeek went further and noticed that the printed subtotal did not reconcile with the printed line items [11]. That is one graded call per model per test, by one grader, with the images and prompts published for replication [17][16]. That is more replication material than most comparisons ship; it is still too few runs to support a latency budget.
The test set is also cleaner than production. A synthetic invoice with three planted arithmetic errors and a chart with a deliberately mismatched second axis are legible inputs [8][9]. In my experience vision models diverge on skewed phone photos and stamps printed over totals, and nothing here measures that.
Where that leaves the buy: if the pipeline is a queue that tolerates a 30-second tail and runs off-peak, the discount is real and the input tokenization makes it larger than advertised. If a human is waiting on the screen, or the batch runs inside the peak window, the cheaper model is not reliably the cheaper call.
