Mercury 3 vs DeepSeek V4 Pro: One Billed by the Clock, One Billed by the Launch

Two models can both cost "about a dollar per million output tokens" and still bill you very differently. Mercury 3 charges $0.35 per million output tokens this week and $0.70 after the launch discount ends — a schedule with one unknown date. DeepSeek V4 Pro charges $1.98 per million output tokens off-peak and $3.96 during two windows every weekday, 01:00–04:00 and 06:00–10:00 UTC — a schedule with a clock attached. If your traffic is bursty and lands in those windows, DeepSeek V4 Pro costs up to 11 times what Mercury 3 costs at launch, and about 5.7 times at Mercury's list price. If your traffic runs overnight and on weekends, the gap narrows to 2.8 times. Neither number is the one in the headline comparison, and that is the point of putting these two side by side. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The models themselves are also further apart than the price suggests. Mercury 3 is Inception Labs' diffusion language model, released October 19, 2026: 1M context, 131,072 output tokens, text only, 1,068 tokens per second by Inception's own measurement, and no numeric benchmark scores published. DeepSeek V4 Pro is D​eepSeek's flagship mixture-of-experts model, released April 24, 2026 and refreshed as the 0813 build in August: 1.6 trillion total parameters with roughly 49 billion active, 1,048,576 context, a 384,000-token output ceiling, text-only input, and a substantial body of third-party measurement behind it. One is a speed play with an unproven ceiling. The other is a capacity play with a proven — and uneven — scorecard.

The rate cards, including the parts that are easy to miss

• Output — Mercury 3 $0.35 per million at launch, $0.70 at list. DeepSeek V4 Pro $1.98 off-peak, $3.96 at peak.

• Input — Mercury 3 $0.10 at launch, $0.20 at list. DeepSeek V4 Pro $0.66 off-peak, $1.32 at peak.

• Cached input — Mercury 3 $0.01 at launch, $0.02 at list, a flat 90% discount. DeepSeek V4 Pro $0.022 off-peak, $0.044 at peak, which is the same 1/30 ratio but applied to a higher base.

• When peak applies — D​eepSeek's peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else, including all weekend hours, is billed at the lower rate. Mercury 3's discount has no published end date.

• Output ceiling — DeepSeek V4 Pro 384,000 tokens, roughly three times Mercury 3's 131,072.

• Context — both nominally 1M tokens.

• Inputs — text only on both. Neither is a multimodal model, so this dimension is a tie rather than the usual differentiator.

• Throughput — Mercury 3 at 1,068 tok/s per Inception, unreproduced. DeepSeek V4 Pro at 119.1 tok/s output measured across our own playground traffic over seven days.

• Reasoning control — Mercury 3 offers instant, low, medium and high effort and defaults to low. DeepSeek V4 Pro runs in thinking mode by default with a non-thinking mode available, and the two modes are priced identically.

The peak-window rule deserves more attention than it usually gets, because it makes a cost model that assumes a single rate wrong by a factor of two for roughly seven hours of every weekday — and those seven hours cover the European morning and the start of the US working day. A workload that looks like $1.98 per million on a spreadsheet can be $2.97 per million blended without anyone changing a line of code.

What D​eepSeek's own numbers say about it

DeepSeek V4 Pro is measurable in a way Mercury 3 currently is not, and the measurements are not uniformly flattering. Artificial Analysis puts it at an Intelligence Index of 36.0 and an AA Coding score of 68.8. Its GPQA Diamond score is 92.8%, Humanity's Last Exam 41%, SciCode 51%, long-context recall 80.3%. On τ²-Bench, the agentic tool-use benchmark, it scores 96.2 — and our own catalogue card leads with that number, which tells you what the model is actually good at.

The number that should give a coding agent pause is Terminal-Bench 4.0: 14.1%. That is not a typo and it is not a different benchmark from the one Claude Opus 5.5 scores 59.6% on. Terminal-Bench 4.0 is a long-horizon terminal task suite, and a model that scores in the low teens on it is not a model you hand a multi-step engineering task and walk away from. Contrast that with the 96.2 on τ²-Bench: DeepSeek V4 Pro is excellent at structured tool calls against well-defined functions and weak at open-ended shell work. Those are different products, and the two scores together describe the boundary.

Mercury 3 has no equivalent table. Its only structured quality claim is an agent run on 50 live Mind2Web navigation tasks where it completed 50%, which Inception says is "significantly more than GPT-6.1-Sol" and which used the same token volume and number of steps. So the comparison a buyer actually wants — how do these two behave on the same agent task — has not been run by anyone in public. What can be said is that they are built for different failure modes: D​eepSeek has deep tool-call competence and a long output budget; Mercury has speed and a much thinner evidence base.

Cost, on a real clock

Take a document-processing workload: 60 million input tokens and 15 million output tokens per day, which is a plausible pipeline for contract or filing review at a mid-sized firm. On Mercury 3 at the launch rate, that is $6.00 of input and $5.25 of output, or $11.25 a day. On DeepSeek V4 Pro at the off-peak rate it is $39.60 and $29.70, or $69.30 a day.

Now move half of that traffic into D​eepSeek's peak windows. Thirty million input tokens at off-peak cost $19.80 and thirty million at peak cost $39.60; the output splits into $14.85 and $29.70. The daily total becomes $103.95. Same tokens, same model, and the invoice went from $69.30 to $103.95. Mercury 3's bill does not move with the clock at all.

Where D​eepSeek earns that back is capacity. The 384,000-token output ceiling means a single call can produce a structured extraction over a very long document without chunking — which for a legal or financial pipeline is not a convenience but a correctness property, because chunking is where cross-references get lost. Mercury 3's 131,072-token ceiling will cover most such jobs and will not cover all of them. If your task routinely produces outputs past 130K tokens, the price comparison is moot: the cheaper model cannot do the job.

The open-weights difference nobody mentions in a price table

D​eepSeek publishes weights. DeepSeek-V4-Pro-0813 sits on Hugging Face under the deepseek-ai organisation, dated August 13, 2026, alongside the V4.1-Flash line that has kept shipping through October. That means you can self-host, fine-tune, quantise or run the model in an air-gapped environment, and no vendor can withdraw it from you. Inception has not published Mercury 3 weights, and the announcement makes no commitment to do so. If your compliance posture requires the model to live inside your perimeter, the comparison has exactly one viable answer.

The trade is that self-hosting a 1.6-trillion-parameter mixture-of-experts model is not a weekend project, and the hosted price is low enough that most teams should not try. The relevant point is optionality, not economics.

Putting both behind one key

DeepSeek V4 Pro is on our catalogue as deepseek/deepseek-v4-pro, billed at DeepSeek's own rate with 0% markup — including the peak multiplier, which passes through unchanged rather than being smoothed into a flat number. That matters here more than usual: a provider whose price moves with the hour is exactly the case where a router should show you the real schedule rather than an average. Mercury 3 is not on our catalogue; our model API has no Inception models at all, so the route to it is Inception's own platform.

The pattern that makes sense for a team evaluating both is a task-shape split rather than a price split. Send long-output extraction and structured tool calling to DeepSeek V4 Pro, where the 384K ceiling and the τ²-Bench score are doing real work; send the latency-critical, short-output tier to Mercury 3 through Inception's key. One OpenAI-compatible endpoint covers the D​eepSeek side and more than 200 other models, with automatic failover if a provider degrades, and the routing DSL can hold the task-shape rule so it is configured once rather than re-implemented in every service. You keep the cheap fast path where speed is the product and the proven capacity where correctness is.

Which one to reach for

Choose DeepSeek V4 Pro when the job is a long structured output, a tool-calling workflow against defined functions, or anything that has to run inside your own infrastructure. Accept that you are paying between 2.8 and 5.7 times Mercury 3's launch price, that the bill changes with the hour, and that Terminal-Bench 4.0 says this is not your autonomous coding agent.

Choose Mercury 3 when the job is short, text-only, latency-critical and cheap to retry — voice turns, hints, search summaries, support drafts — and when you are willing to run the evaluation yourself, because nobody has published one. The launch discount makes the trial nearly free; the absence of an end date makes it a bad foundation for a five-year cost model.

The one thing not to do is pick between them on the headline token price. D​eepSeek's clock and Mercury's launch window are both temporary in their own way, and the number that should drive the decision — cost per completed task, including retries and human review — is the number neither vendor publishes.