Inception Labs’ Mercury 3 and DeepSeek V4 Pro offer very different bargains for people paying to use AI models: Mercury is cheaper at its current launch rate, while DeepSeek’s price rises during set weekday hours and it can produce much longer outputs. That makes the real choice about workload and timing, not just the rate-card headline.
DeepSeek Watch analysis
What happened
A comparison by OrcaRouter says Mercury 3 costs $0.35 per million output tokens during its launch offer, rising to $0.70 at list price. The offer’s end date is not published. DeepSeek V4 Pro costs $1.98 off-peak and $3.96 during weekday peak windows, 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. The peak rate applies for about seven hours on a typical weekday; weekends are off-peak.
Both models are text-only and list a one-million-token context window. DeepSeek’s maximum output is 384,000 tokens, against 131,072 for Mercury 3. Inception says Mercury reaches 1,068 tokens per second, but the comparison notes that this speed has not been reproduced and that Mercury lacks a published numerical benchmark table.
Why it matters
A model’s per-token price is only useful if you know when it applies and whether the model can handle the job. OrcaRouter’s example of a document-processing workload puts DeepSeek’s daily bill at $69.30 if all traffic is off-peak, or $103.95 when half falls in peak windows. The same example costs $11.25 on Mercury at its launch rate. Those figures are the comparison’s illustration, not a universal forecast for every workload.
There is a capability trade-off behind the price gap. The comparison lists DeepSeek at 96.2 on τ²-Bench for structured tool use, but 14.1% on Terminal-Bench 4.0, a test of longer-horizon terminal tasks. It cites no equivalent benchmark table for Mercury. A model can be good at calling defined tools without being a safe bet for an open-ended coding job. Annoying, perhaps, but cheaper than learning that distinction from an invoice.
Our read
This is a useful reminder to price the task you actually run, including peak hours, output length and retries, rather than picking a winner from one attractive number. Mercury looks compelling for short, text-only work where speed and a low trial cost matter. DeepSeek has the longer output ceiling and more published measurements, but costs more and its results vary sharply by task. OrcaRouter operates a model-routing service, so treat its comparison as a provider’s analysis, not a neutral final verdict.
What to watch
- Whether Inception publishes more comparable benchmark results or an end date for Mercury’s launch pricing.
- Whether DeepSeek changes its weekday peak windows or rates.
- How both models perform on the same real workloads, including the cost of retries and human review.
Discussion spark: Would you choose the cheaper, less-tested model for routine jobs, or pay more for a model with a longer output limit and a clearer performance record?
Sources and evidence
- Mercury 3 vs DeepSeek V4 Pro: Clock vs Launch Pricing – OrcaRouter (19 October 2026, 18:16 UTC)
not affiliated with or endorsed by DeepSeek