Everybody’s running an AI agent now. Not a chatbot. An actual agent that reads your inbox, drafts your replies, manages your calendar, and quietly fixes your code while you’re asleep. And the second you sit down to build one, you hit the exact same wall. Claude or Gemini?
It’s a genuinely annoying decision. Both are excellent. Both are cheap enough to run all day. Both have million-token context windows now. And yet they feel completely different in your hands.
So I spent a few weeks running both through the same personal-agent workloads. Email triage. Calendar wrangling. A coding sidekick. Some light research. Nothing exotic. Just the stuff an assistant actually does on a Tuesday.
Here’s what I found. And no, there isn’t a clean winner.
The two lineups, quickly
When people say “Claude” they usually mean one of three models. Opus 4.8 is the flagship — brilliant, careful, and the priciest at roughly $15 per million output tokens. Sonnet 5 is the workhorse, the default on Claude.ai, around $2 to $3 in and $10 to $15 out. Then there’s Haiku, the small fast one for bulk tasks.
“Gemini” is the same shape. 3.1 Pro is the brains, around $2 in and $12 out, though it’s still technically in preview. Flash is the cheap speed demon. Both families give you a 1M token context window these days, which is more than enough to hold a whole day of someone’s life in memory.
Thing is, the tiers map onto each other almost too neatly. Opus ≈ Pro. Sonnet ≈ mid. Haiku ≈ Flash. So the real question isn’t “which model.” It’s which house you want to live in.
Where Claude pulls ahead: coding and computer use
If your agent writes code — and let’s be honest, half of them do — Claude is still the safer bet. On SWE-bench Pro, the contamination-resistant coding-agent benchmark, Sonnet 5 posts a 63.2%. Gemini 3.1 Pro manages 54.2%. That’s a nine-point gap, and you feel it on real projects, not just on leaderboards.
Claude also owns the computer-use lane. Its OSWorld-Verified score sits at 81.2%, meaning it can actually operate a desktop — click, type, drag — with real reliability. Google doesn’t even publish a comparable number. For an agent that’s supposed to drive your browser and book your flights, that matters a lot.
But here’s the kicker. Claude’s strength is narrow. It’s a coding and tool-use machine. Ask it to reason through something abstract and it’s fine, not spectacular.
Where Gemini pulls ahead: reasoning and multimodal
Gemini’s the opposite shape. On pure reasoning it’s quietly monstrous. 94.3% on GPQA Diamond, a brutal graduate-level science benchmark. 77.1% on ARC-AGI-2. Anthropic simply doesn’t publish those numbers for Sonnet, which tells you something on its own.
And the multimodal thing? Real. Gemini 3.1 Pro ingests text, images, audio, video, and PDFs natively. Drop a thirty-minute voice memo into your agent’s context and it just gets it. Claude is text and images only. For a personal assistant that’s supposed to listen to your rambling voice notes or read a scanned document, that gap isn’t academic.
So Gemini’s the pick when your agent needs to reason hard and absorb messy real-world input. Claude’s the pick when your agent needs to do things.
The cost thing
Honestly? They’re close enough that cost shouldn’t decide this for you. Sonnet 5 runs $2 to $3 in and $10 to $15 out per million tokens. Gemini 3.1 Pro is $2 in and $12 out. At the mid tier it’s a wash.
The real cost lever is the cheap model. If your agent does a thousand tiny tasks a day — categorizing emails, summarizing threads, filing receipts — you want Flash or Haiku doing that grunt work, not the flagship. That’s the difference between a $5 month and a $50 month for an always-on agent.
Reliability, and the preview problem
This one’s underrated. Claude Sonnet 5 is generally available right now — it’s the boring default that’s been shipping for weeks. Gemini 3.1 Pro is still in preview. It launched back in February and hasn’t fully graduated. That means occasional rough edges, and availability that can wobble when you least want it to.
For a personal agent that runs while you sleep, “it’s in preview” is a real sentence to weigh. Not a dealbreaker. But not nothing either.
Ecosystem: the part nobody talks about
And honestly? It’s the part that actually decides it. If you live in Google’s world — Gmail, Calendar, Drive, Docs, the whole Workspace swamp — Gemini has a native edge. It plugs into that stuff without a dozen brittle API bridges. Google built it to live there.
Claude, meanwhile, lives in Claude Code and the API. It’s the developer’s assistant. If your agent is going to write code, run terminal commands, and control a computer, the Claude ecosystem is tighter and more mature.
A real week with both
I gave each of them my actual inbox. 214 unread. A pile of newsletters, two threads about a contractor, one angry client, and my kid’s school asking for a permission slip. Both drafted reasonable replies. But Gemini caught the subtext in the angry client’s email — the actual worry underneath the complaint — where Claude gave me a polite, slightly flat response. Claude, on the other hand, rewrote a broken Python script I’d been ignoring for a month and got it running on the second try.
Same week. Two different wins. That’s the whole article, really.
So which one?
If your agent codes, controls a computer, or needs to ship today with zero flakiness — pick Claude. Sonnet 5, specifically. It’s the mature, reliable choice, and I mean that as a compliment.
If your agent reasons, plans, listens to voice notes, and lives inside Google’s apps — pick Gemini. 3.1 Pro is doing more with raw reasoning and multimodality than anything else in its price bracket.
Me? I ended up running both. Gemini handles the inbox and the planning, Claude handles the code. Two agents, one life, and a slightly emptier brain for me. Not the tidy answer you came for. But it’s the honest one.