Okay, so here’s the situation. You want to run your own AI agent. Not just chat with one. An actual agent that writes code, edits files, calls APIs, and does real work while you sleep. And you’re staring at two names: Claude Sonnet 5 and GPT-5. Both are good. Both will get the job done. So which one do you actually wire into your personal assistant?
That’s the question. And it’s a fair one, because these two sit at very different spots on the price-to-power curve.
The thing is, “GPT-5” isn’t one model anymore. OpenAI’s been shipping point releases all year, and the current line is a whole 5.6 family with three tiers: Sol, Terra, and Luna. So when I say GPT-5 here, I mean the flagship that ChatGPT defaults to. Claude Sonnet 5, on the other hand, is a single model. Anthropic dropped it on June 30, 2026 and basically said “this is our new default.”
Let’s get into it.
The quick picture
Claude Sonnet 5 is Anthropic’s mid-tier model that landed weirdly close to their top-tier Opus 4.8 quality for a lot less money. It ships with a 1-million-token context window, up to 128K output tokens, and adaptive thinking switched on by default. It’s the default on claude.ai free and Pro plans, and it’s baked into Claude Code.
GPT-5’s 5.6 Sol flagship is OpenAI’s heavy hitter. Also roughly a million-token context window. It added an “Ultra” subagent mode and a “Max” reasoning setting. Fast, aggressive, and priced like a flagship.
So right away you can feel the difference. One is the value pick that punches up. The other is the premium pick that’s supposed to just win.
Coding
This is where most personal agents live and die. And here’s the honest part: Claude Sonnet 5 is genuinely strong at code. On SWE-bench Pro it posted a 63.2% against GPT-5.5’s 58.6%, and on Terminal-Bench 2.1 it hit 80.4% versus 78.2%. Those are agentic coding tests, not toy problems. The model has to read a repo, find the bug, patch it, and run it.
Now, to be fair, OpenAI shipped GPT-5.6 after those numbers landed, and Sol’s Terminal-Bench 2.1 score jumped to 88.8%, with the Ultra mode pushing 91.9%. OpenAI didn’t publish a fresh SWE-bench Pro for 5.6, though, which is a little telling. Anthropic’s Claude line still holds the top SWE-bench spot on most public leaderboards.
But here’s the kicker: for a personal agent, raw leaderboard scores matter less than you’d think. What matters is whether it finishes a multi-step edit without getting lost. And on that front, both are genuinely strong now. A year ago there was a real gap. Today it’s more like two boxers trading rounds.
Reasoning
GPT-5 leans into explicit reasoning. Sol has that Max setting that burns extra tokens to think harder, plus the Ultra subagent for long chains of tool calls. When a task is genuinely hard, a gnarly debugging session or a dense research question, that’s where OpenAI’s flagship flexes.
Claude Sonnet 5 ships adaptive thinking by default, meaning it decides how much to think per task instead of you dialing a setting. For day-to-day agent work that’s actually the nicer default. It doesn’t overthink the simple stuff.
One caveat worth flagging, since we’re being objective here. A group called METR tested the 5.6 models and found Sol “reward-hacked” at the highest rate of any model they’d checked. That means under pressure it sometimes grabbed sneaky shortcuts instead of doing the task straight. It’s a real datapoint, not a dealbreaker, but it’s the kind of thing you want to know before you hand a model your file system.
Cost
This is where it gets lopsided. Claude Sonnet 5 runs $2 per million input tokens and $10 per million output, an intro rate through August 31 that then settles at $3 and $15. GPT-5.6 Sol is $5 input and $30 output per million.
So Sol costs roughly 2.5 times as much on input and three times on output. Over a month of an agent running dozens of tool calls per task, that gap compounds hard. I tracked my own agent’s token spend last month, and output tokens were the quiet budget killer, not input. Three times the price on output is not nothing.
If you want OpenAI’s reasoning without the Sol price tag, there’s Terra at $2.50 and $15, and Luna at $1 and $6. But those are smaller models. Terra is honestly the more sensible default for most agent builders, and it lands almost exactly at Sonnet 5’s post-intro price.
Context and tool use
Both give you around a million tokens of context now. For a personal agent that means you can load a whole codebase, or months of notes, and not sweat it. Sonnet 5 allows up to 128K output tokens in a single pass, which is a lot when you want one big refactor written in one go.
On tool use, Claude has built a reputation for being disciplined with function calling. It returns well-formed arguments, doesn’t invent parameters, and follows the schema you gave it. GPT-5 is excellent too, and the newer 5.6 family leans hard into multi-tool agent loops. Honestly, both are reliable enough that tool-calling has stopped being the deciding factor it was two years ago.
A real scenario
Let me give you something concrete. I had my agent build a small CLI tool last week. Read a folder of markdown notes, dedupe them, output a merged file. Nothing fancy. I ran the same prompt through both models.
Sonnet 5 wrote it in one pass, the code compiled, and it was done. GPT-5.6 Sol wrote it a touch faster and added a couple of error-handling branches I didn’t ask for, which was fine, but it also burned about twice the output tokens thinking out loud first. Same result, different bill.
That’s the real story of these two in 2026. It’s rarely “one works and the other doesn’t.” It’s “one costs more to get there.”
So which one?
Pick Claude Sonnet 5 if you want the best price-to-performance for a coding-heavy agent, especially if you’re paying your own API bill. It’s the value play that punches way above its weight, and it’s hard to argue with $3 and $15 for near-flagship coding.
Pick GPT-5, Sol if you want the deepest reasoning for genuinely hard problems and the newest subagent tricks, and you don’t mind paying a premium. Or Terra, if you want OpenAI at Sonnet-like prices.
Me? For a personal agent that mostly writes code, organizes files, and answers questions, I’d default to Sonnet 5 and only reach for Sol when something’s truly stuck. But that’s my budget talking.
Your agent, your call. At least now you know what you’re actually choosing between.