DeepSeek V4.1 Pro vs Gemini 3.8 Flash: A Missing Flagship Against a Cheap Google Model

This is the least symmetrical matchup on the site, and the asymmetry is not the one the names suggest. Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 output, carries a 1M context with a 65K output ceiling, and is callable on the same key as everything else. DeepSeek V4.1 Pro has no card, no endpoint and no price. The DeepSeek flagship that exists — DeepSeek V4 Pro — costs $0.66 and $1.98 off-peak, roughly a tenth of Flash's input price and half its output price, with a 384K output ceiling instead of 65K.

Read that again, because it inverts the impression the word "flash" creates: Google's cheap tier is the more expensive one per token, and DeepSeek's flagship is the one with the larger generation ceiling.

The cards

• Gemini 3.8 Flash — $0.75 in, $3.75 out; cache reads at $0.075; 1M context; 65K maximum output; measured p50 first token 8.13 seconds.

• DeepSeek V4 Pro — $0.66 in, $1.98 out off-peak; $0.022 cached; doubles to $1.32 and $3.96 in two weekday windows; 1M context; 384K output; text-only; 6.22 seconds to first token.

• DeepSeek V4.1 Pro — nothing to compare. Named once on 10 September and never priced.

At off-peak, DeepSeek is 12% cheaper on input and 47% cheaper on output. In a peak window it is 76% more expensive on input and 5.6% more expensive on output. That swing — from clearly cheaper to clearly dearer within the same day — is the most important thing on this page, and it is invisible on either vendor's marketing copy.

The 65K ceiling decides more workloads than the price

Sixty-five thousand output tokens is a long answer and a short document. A single-call generation of a 40,000-word report does not fit; a code file of any ambition approaches the limit before it is finished. DeepSeek's 384K ceiling is nearly six times larger, and for anything in that class the price comparison stops mattering because one of the two cannot do the job in one request.

Where the ceiling does not bind — chat, classification, summarisation, most agent steps — the token price is what you pay, and there DeepSeek wins off-peak and loses in peak. Scheduling, not model choice, decides the outcome.

Latency flips the ranking again

Gemini 3.8 Flash measures 8.13 seconds to first token; DeepSeek V4 Pro measures 6.22. Neither is fast, and both are slow enough that an interactive product feels them. Against a 65K ceiling and an 8-second first token, the case for Flash is cache economics — $0.075 per million cached reads is the cheapest published cache price on any card in this comparison — rather than sticker price.

What to run

Gemini 3.8 Flash and DeepSeek V4 Pro are both on OrcaRouter at their vendors' list prices, passed through with no markup, on one OpenAI-compatible endpoint with automatic failover — so this comparison is a routing rule, not a migration. Send the long-generation work to the model with the 384K ceiling, the cache-heavy work to the one with the $0.075 cached rate, and the rest to whichever is cheaper at the hour it runs.

DeepSeek V4.1 Pro stays out of that routing table until DeepSeek publishes an identifier. Anything else is a plan built on a release note.