Here’s the thing about running your own AI agent: the model you pick quietly decides everything downstream. How fast it answers. How much you pay at the end of the month. Whether it can actually read a screenshot or just pretend to. Get it wrong and you’re either overpaying for power you never touch, or banging your head against a tool that trips over the exact thing you hired it for.
So this week I sat down with two very different contenders. Google’s Gemini and DeepSeek’s V4. Both ship a one-million-token context window now. Both sit in the “frontier” bucket on most leaderboards. But they couldn’t be further apart in philosophy — and in price.
Who these two actually are
Gemini (I’m looking at Gemini 3.1 Pro, the February 2026 release) is Google’s flagship. It’s natively multimodal, which is a polite way of saying it can chew on text, images, audio, video, and code in a single request, no translation layer in between. That’s rare, and it matters more than people admit.
DeepSeek V4 is the open-weights upstart. Two variants. V4 Pro is the heavy lifter — 1.6 trillion parameters, 49 billion active. V4 Flash is the lightweight at 284 billion total, 13 billion active. Both MIT-licensed. You can self-host if you’ve got the GPUs and the stomach for it.
Cost: this is where it gets loud
Gemini 3.1 Pro runs $2 per million input tokens and $12 per million output. DeepSeek V4 Pro? $0.435 in, $0.87 out. V4 Flash is $0.14 and $0.28.
Let that sink in. DeepSeek’s flagship is roughly four and a half times cheaper on input and nearly fourteen times cheaper on output than Gemini. Flash is basically pocket change. For a personal agent that churns through a few million tokens a month, that’s the difference between a coffee and a dinner bill.
But here’s the kicker: DeepSeek’s pricing is about to move. On August 16, 2026, both V4 models shift to peak and off-peak tiers. Off-peak isn’t a discount over today — it’s half of a raised peak. So the “cheapest model” crown has a foot in the sand.
Gemini’s pricing, by contrast, has been flat since launch. Predictable. You can budget against it without checking a rate card every morning. Some people will happily pay a premium for that calm.
Coding
This is the one most agent builders actually care about. On SWE-bench Verified, Gemini 3.1 Pro lands at 80.6%. DeepSeek V4 Pro sits at 80.6% too — statistically tied with Claude Opus 4.7’s 80.8%. On LiveCodeBench, V4 Pro hits 93.5 and a Codeforces ELO of 3206, which actually edges out GPT-5.5’s 3168.
So on raw coding benchmarks they’re basically neck and neck. The difference is the bill you pay to get there. Same rough competence, wildly different price.
Where Gemini pulls ahead is coding with context. Dropping in an entire repo plus a screenshot of the broken UI and asking “what’s wrong.” DeepSeek handles text and code just fine, but it isn’t natively multimodal the way Gemini is. If your agent ever needs to look at a screen, Gemini wins that round by default.
Reasoning and hard problems
Gemini 3.1 Pro put up 77.1% on ARC-AGI-2, the benchmark built specifically to make memorization useless. That’s more than double Gemini 3 Pro’s 31.1% from three months earlier, and it beats Claude Opus 4.6’s 68.8% and GPT-5.2’s 52.9%. On GPQA Diamond, a graduate-level science gauntlet, it hit 94.3% — the highest reported yet.
DeepSeek V4’s reasoning is solid but harder to pin down. It offers three modes — non-think, think, and a max-effort setting. The open-weights crowd tends to lead with coding numbers, not ARC-AGI, so V4 Pro doesn’t get the same headline treatment. On pure novel-reasoning flex, Gemini’s the one with the receipts.
But. Reasoning benchmarks are a party trick if your agent mostly does triage, drafting, and file work. And for that, DeepSeek’s “think high” mode gets the job done at a fraction of the tokens Gemini burns in its own thinking pass.
Self-hosting and data privacy
Here’s where DeepSeek does something Gemini simply can’t. Open weights. If your agent handles stuff you’d rather not ship to a third-party API — bank statements, medical notes, the messy personal stuff — DeepSeek V4 lets you run it on your own hardware. V4 Flash fits on a single H100 if you’re careful, or four H200s comfortably. V4 Pro wants eight H100s at FP8, so that’s a bigger lift.
Gemini has no open-weights story. You use Google’s API or you don’t use it. That’s not a knock — plenty of people are fine with it — but if self-hosting is a hard requirement, the argument’s already over. DeepSeek wins that one walking away.
Speed and reliability
Speed is murkier than the spec sheets let on. Gemini has low, medium, and high thinking levels — crank it up and you trade latency for accuracy. DeepSeek V4 Pro uses speculative decoding to keep generation snappy, with published concurrency of 500 for Pro and 2,500 for Flash.
Reliability, though, is where I’ve learned to be picky. DeepSeek V4 has a known multi-turn quirk — a “reasoning_content 400” error that bites popular clients on first contact. There’s a fix, but you’ll bump into it. Gemini’s tool loop is smoother out of the box. Fewer surprises when your agent is three steps into a five-step task.
A real scenario
Say you’re running a personal agent that watches your inbox, drafts replies, and occasionally writes a small script. You’re not doing frontier research. You just want it to work and not cost much.
On that workload, DeepSeek V4 Flash is hard to argue with. Pennies per day, MIT license, runs on modest hardware if you self-host. The agent is fast, the bill is forgettable.
Now flip it. Your agent reads screenshots, summarizes video-call notes, and you lean on it for genuinely hard reasoning when you’re stuck. That’s Gemini’s home turf. You’ll pay more, but you’ll actually get multimodal input and top-shelf novel reasoning.
Verdict
If you’re price-sensitive and your agent is text-and-code — DeepSeek V4. Probably Flash for volume, Pro when the task actually matters.
If your agent touches images, audio, video, or needs the strongest novel reasoning and a dead-simple tool loop — Gemini 3.1 Pro.
Me? I keep both keys in my agent config. DeepSeek for the daily grind, Gemini for the moments a screenshot or a genuinely hard question shows up. Two models, one agent, most of the savings.
But ask me again in a month. DeepSeek’s prices move on August 16, and these two are clearly watching each other.