DeepSeek V4 Pro vs Gemini 3.6 Flash: cheaper, but 3–6× slower to a finished answer

What we actually measured

Every Saturday our AI Arena rig — an NVIDIA RTX 5060 Ti 16 GB that runs local models through Ollama plus cloud API tests — pushes a batch of models through four real-world tasks: generating a PHP Drupal module, an HTML/JS animation, a Python "galaxy" scene, and an English article. This week's fresh run produced 16 results across eight model×test combinations, but unusually only cloud models landed in the dataset. Local Ollama models did not log fresh runs this cycle, so the head-to-head is between two cloud APIs: DeepSeek V4 Pro and Gemini 3.6 Flash. The new part is the timing — not just the rate card, but how long a developer actually waits for a finished answer.

Here are the raw numbers from our rig, exactly as recorded:

*The 4,096,000 tok/s figure is a measurement artifact, not a real speed — see below.

The 4,096,000 tok/s that isn't real

No model produces four million tokens a second. The figure comes from our runner logging a bad instantaneous sample rather than a steady-state rate, and we would rather flag our own tooling bug than print a number that would mislead you. The honest way to read DeepSeek V4 Pro is through wall-clock time and token count: it emitted 4,096 tokens per task, taking 66.2 s, 72.1 s, 87.7 s and 100.3 s respectively. Divide those and you get an effective throughput of roughly 61.9, 56.8, 46.7 and 40.8 tok/s.

Notice the direction of that decline. The longer and more open-ended the task, the slower DeepSeek V4 Pro gets. That is the fingerprint of a reasoning model: the "4096 tokens" it counts are not all visible text. A large share is hidden reasoning — the model thinking out loud to itself before it commits to an answer. You never see those tokens, but you pay for them and you wait for them.

Time to first token: why "Flash" sat silent for 28 seconds

Time-to-first-token (TTFT) measures how long you stare at a blinking cursor before the first token appears. The two models tell opposite stories here, and both deserve a skeptical look.

Gemini 3.6 Flash — despite the "Flash" name — logged a TTFT of 7.5 s on the Drupal task, 16.3 s on the animation, and a striking 28.5 s on the Python galaxy test. On that last test it produced only 164 tokens total, so 28.5 of its 29.7 seconds were spent before the first token. In practice this means the model spent most of its time planning the program before writing it, which is fine for a one-shot code generation but painful if you expect the kind of instant streaming the Flash brand implies.

DeepSeek V4 Pro logged a 0 ms TTFT, which looks great until you check the finish time: 66–100 seconds per task. What is happening is that DeepSeek's API streams its first reasoning token almost immediately, so our client records "first token" right away — but the token that arrives first is thinking, not the answer. The takeaway for anyone integrating these APIs: TTFT alone is a poor proxy for "how long until I get my result." Total time-to-answer matters more, and on that metric Gemini 3.6 Flash wins every single task this week, finishing 3.0× to 5.9× faster.

Local vs cloud: what it costs a European user

This week's run had no fresh local results, but the pricing picture is where the decision gets interesting. At current public rates, DeepSeek V4 Pro costs $0.66 per 1M input tokens (cache miss) and $1.98 per 1M output tokens off-peak, with peak rates double that and a cache-hit input price of just $0.022. Gemini 3.6 Flash is listed at $0.75 input / $3.75 output per 1M tokens as introductory pricing through December 31, 2026, rising to $1.50 / $7.50 from January 2027.

Converted at roughly $1 ≈ €0.90, that puts DeepSeek V4 Pro at about €0.60 input / €1.80 output per 1M tokens, versus Gemini 3.6 Flash at €0.68 / €3.40. DeepSeek is the cheaper token, and its cache-hit rate makes repeated, context-heavy calls dramatically cheaper still. Do the arithmetic on our own article test: roughly 4,096 total tokens from DeepSeek cost about €0.008, while Gemini's 2,330 tokens cost about €0.009. Both cost well under one euro cent per article — at these prices the real trade-off is not money, it is the 30 versus 100 seconds of latency, and that only flips in Gemini's favour if speed matters to you.

A local model on our RTX 5060 Ti 16 GB costs €0 per token forever after a one-time hardware purchase of a few hundred euros, but you are capped at what 16 GB of VRAM can hold and the throughput of a single consumer GPU. The cloud buys you frontier reasoning and no hardware to maintain, at the cost of a recurring per-token bill. There is also a compliance dimension: DeepSeek is a Chinese provider without EU data residency, so a European company handling personal data under GDPR must think twice before sending customer text to it, whereas Google lets you pin Gemini processing to an EU region. For European teams, that control is often worth more than the small token-price difference.

If you want to run this comparison yourself, our full history and the raw JSON are on the AI Arena comparison page.