Here’s the thing nobody warns you about when you start building a personal AI agent. The model choice stopped being a technical decision a while ago. Now it’s a budget decision with a personality attached.
Two names keep coming up. Claude Sonnet 5, the polished mid-tier workhorse from Anthropic. And DeepSeek V4, the MIT-licensed budget king you can download and run in your basement if the mood strikes.
Both want to be the brain inside your assistant. They can’t both be right — except they kind of can, once you know what your agent actually does all day.
What you’re actually choosing between
Claude Sonnet 5 is Anthropic’s middle child. Not the big Opus, not the tiny Haiku. It’s the one they specifically tuned for agentic work, the one most builders quietly default to. Closed-source, API-only, priced at $2 per million input tokens and $10 per million output. That was “introductory pricing” until a few weeks ago. Now it’s just the price.
DeepSeek V4 is the opposite philosophy entirely. Open weights, MIT license. A 1.6-trillion-parameter mixture-of-experts model with only around 49 billion parameters active at any given moment. Self-host it, fine-tune it, or hit their API at $0.435 input and $0.87 output per million. There’s a Flash tier too, at $0.14 and $0.28, for when you’re genuinely counting every cent.
Both give you a 1-million-token context window. So throwing a whole codebase or six months of notes at either one is fine.
Coding is where they split hard
If your agent writes code, pay attention here. Sonnet 5 leads on basically every shared software-engineering benchmark. SWE-bench Verified? 85.2% to DeepSeek’s 80.6%. The harder SWE-bench Pro? 63.2% to 55.4%. That gap isn’t a rounding error — it’s roughly 145 extra GitHub issues solved out of the 1,865 in the Pro set.
But here’s the kicker. DeepSeek V4 Pro currently holds the #1 LiveCodeBench score in the world at 93.5%, with a Codeforces rating of 3206. That’s about a top-25 human competitive programmer. Sonnet doesn’t publish those numbers at all.
So the split looks like this. Multi-file debugging, refactoring, wiring up tool calls that have to work — Sonnet is the steadier hand. Algorithmic puzzles, competitive-style problems, clever one-shot solutions — DeepSeek punches way above its price tag.
Reasoning and tool use — the quiet differentiator
Here’s where I’d actually put my money, if my agent has to chain steps together. On Humanity’s Last Exam with tools, Sonnet 5 scores 57.4% against DeepSeek’s 48.2%. Without tools it’s 43.2% to 37.7%. BrowseComp, the agentic-search benchmark, is nearly a dead heat — 84.7% to 83.4%.
And honestly, tool-augmented reasoning is the whole job. A personal agent checks your calendar, reads an email, drafts a reply, then asks itself whether that reply is actually good before it sends. Sonnet’s edge there is real, and you’ll feel it across a week of small tasks more than on any single one.
DeepSeek isn’t bad at it. It just wants more steering. Slightly clearer instructions, a bit more hand-holding on the first step. The difference is invisible on one task and obvious over fifty.
The cost math, because your agent never sleeps
A personal agent is a background process. It summarizes your inbox at 2 a.m. while you’re asleep. It’s chewing tokens whether you’re watching or not. That’s where DeepSeek’s price gets genuinely hard to argue with.
Let’s do a real number. Say your agent burns 50 million input tokens and 10 million output tokens in a month. On Sonnet 5, that’s $100 in plus $100 out — $200 total. On DeepSeek V4 Pro, it’s $21.75 in plus $8.70 out — about $30. On the Flash tier it barely registers.
That’s the entire story in one paragraph, really. DeepSeek is around 4.6× cheaper on input and 11.5× cheaper on output. For a thing that runs 24/7, that compounding difference isn’t trivial.
Speed, reliability, and the boring stuff
On speed, they’re close enough that most people won’t care. Sonnet 5 feels snappy on first token and steady on long outputs. DeepSeek V4 can be a hair slower to start, especially in its Max reasoning mode, where it thinks longer before answering. Not a dealbreaker. Just something you notice when your agent is running twenty little queries back to back.
Reliability is where the price gap shows up differently. Sonnet almost never derails mid-task. DeepSeek occasionally does — a lost step here, a misread instruction there. It recovers fine if your agent has good error handling. But a personal agent that silently goes off the rails at 3 a.m. is the kind of thing you only discover over breakfast.
The stuff spec sheets don’t lead with
One detail that gets skipped a lot. Sonnet 5 can look at images. Screenshots, photos, a picture of a receipt your agent needs to log. DeepSeek V4 is text-only. If your assistant’s job involves reading your screen or parsing an image, that’s not a minor quirk — it’s the difference between “it works” and “you need a workaround.”
On the flip side, open weights mean control. Air-gapped, self-hosted, fine-tuned on your own writing. For an agent that holds your calendar, your email, your private notes, data sovereignty isn’t a paranoid concern. It’s a legitimate one. DeepSeek V4 Pro is still technically labeled “preview,” though — worth knowing if you’re building something you don’t want to rebuild in three months.
A real Tuesday with each one
Imagine a normal morning. The agent pulls nine overnight emails, drafts replies to three, flags one that’s actually urgent, and updates your task list. Sonnet does this cleanly, writes replies that sound like you, and almost never needs a second pass. DeepSeek does it too, maybe 30 seconds slower, and one of the three drafts needs a quick edit before it’s sendable.
Now imagine you told the agent to refactor a messy Python script you’ve been putting off. Sonnet opens it, makes the changes, checks its own work. DeepSeek makes the changes faster and cheaper, but you’ll probably read the diff a little more carefully before you commit.
Neither is broken. They just cost different things — money versus your attention.
So which one?
If your agent does real work — coding, planning, multi-step tasks with tools — and the bill doesn’t make you wince, Sonnet 5 is the more reliable brain. It gets things right the first time more often, and it can see the screen.
If you’re price-sensitive, or you want to self-host, or your agent mostly does high-volume text work where a five-point benchmark gap is irrelevant, DeepSeek V4 is hard to beat. Near-frontier quality at a fraction of the cost, with a license that leaves the door open.
I swap the model under my own agent more than I’d like to admit, and I keep landing on the same conclusion. There’s no single right answer. Just the answer that fits the agent you’re actually running, on the Tuesdays you actually have.
Pick that one. Ignore the leaderboard.