I Tested Claude Fable 5 vs GPT 5.6 Sol vs Gemini 3.1 for AI Agents

0:00The whole comparison only counts if the test is fair, and making it fair took more setting up than the test did. Every lab claims its own model is the best one for agents, and every benchmark page is built to make its own numbers look good. So, the only way to settle it is to build one agent and change nothing about it except the model. The agent I'm using for that is Hermes, which is open source, completely free, and made by an AI lab called News Research. Building an AI agent is far simpler than a year ago with more and more apps making the process really easy. The hard part now

0:26is picking the right model because there's so much to choose from. So, I ran Claude 3 5 GPT 5.6 and Gemini 3.1 Pro through the same agent, and by the end you'll know exactly which to build yours on and which to use for what. There's a similar tool called Open Claw that works much the same way, so everything you're about to see applies there, too. An agent only works while it's switched on, and the second your laptop closes, it stops. So, it needs a server that stays switched on day and night. That server is coming from Hostinger, and the link in the description takes you straight to the

0:52deploy page if you want to build along with me. The plan I'm on is KVM 2, which gives you two processor cores, 8 gigs of RAM, and 100 gigs of NVMe storage. That's plenty here because an agent like this spends most of its time waiting on web pages rather than working the processor. On the deploy page, you'll see the code Parker already applied in the coupon box on the right. You get the full discount on the 12 and 24-month terms. So, 12 is the one I'd set. Hermes is a one-click install on there, so I'll pick it and choose the server location closest to me. The piece that makes the

1:18whole comparison fair is how Hermes connects to models. You plug in one key from a service called Open Router, and that single key reaches hundreds of models across every major lab at the same price as the labs charge. All three of the ones I'm testing are on there. I'll put a spending cap of about $25 on the key before anything runs. So, a runaway test can't drain the account while I experiment. Switching models takes one line in the chat. I'll just type model and the model's name. Hermes moves the whole agent onto it, and between tests, I start a fresh session, so nothing carries over from the run

1:45before. Every prompt I use today is in the free doc linked below, so you can run the same three tests on your own server and check my results. The tests go easiest to hardest. It starts with a website from one prompt, then a 3D game, and it ends on a full agent job with tools and a schedule. And each one stresses something the one before it couldn't reach. The first job sounds like the easy one, and it turns out to be the one where the three of them agree the least. The first test is the website because nothing shows you a model's taste faster. You're not going to ask

2:10your agent for a website, but the same taste shows up in whatever it does make you. A one-shot build puts that judgment on screen in about a minute. The rules are the same for everyone, so each model gets one prompt with no follow-ups, and whatever comes back is the score. I'm asking for a landing page for a small coffee roastery called Ember. It needs a hero section, a product row with prices, and a story block, plus a review strip, and a contact footer. That's enough decisions for the models to show

2:32themselves. I'll paste in this prompt. First up is Fable 5, so I'll pick it out of the model selector and drop the prompt in. Hermes writes the whole site to a file, and I open that file in my browser to see what it made. And as you can see, it comes back with the most complete site of the three. Every section I asked for is there, and all three prices are exactly where they should be. The styling plays it safe, though. Tidy fonts and soft colors, and nothing about it would stop you scrolling. Gemini 3.1's build comes back on the same prompt with every section in

2:56place and nothing falling over. It just doesn't commit to a look, so what you get is a page about a coffee roastery rather than something that feels like the roastery itself. Then Saul's site is the one that actually looks designed. The colors are chosen rather than defaulted, the spacing gives everything room, and it's the only one of the three I'd put in front of a client without touching it. The design win goes to Saul, and the completeness win goes to Fable with Gemini in the middle. A web page can hide weak logic behind good looks, though, and a game can't hide anything. A game has to hold graphics,

3:23physics, and controls together at the same time, and the first 10 seconds of play put all three under load. And you're not going to build games with your agent, either, but a game is about the hardest thing you can ask a model for in one go. So, if it can juggle all of that at once, the everyday multi-step job you actually give your agent isn't going to fall apart halfway. The job is an endless runner where a fox sprints down a rainy night street dodging falling crates. It gets arrow key controls, a score counter, and a restart screen, all in one file I can open straight from the server. So, I'll paste

3:49this one in, and Saul's version runs on the first try. The fox moves, and the crates fall and the score counter and restart button both do their jobs. Every mechanic I asked for is in there and working and it's the best looking of the three on top of that. Gemini's build is the one that comes apart. It doesn't really look good and it isn't what I asked for either. A minute in it's clear the collisions barely exist and the animations aren't clean. So on the website it was the safe one that worked and here it's neither. Fable 5's build is the one that catches me off guard

4:14because it plays like somebody tested it before handing it over. The controls feel right and the crates land where they look like they'll land. The difficulty creeps up the longer you survive. Entropic says Fable checks its own work as it goes and a first try build that plays this well is what that looks like in practice. Fable takes the game test because it's the only build that went past what I asked for. Soul is second with the best looking of the three and Gemini is third and two tests in every result so far has come from one prompt and one answer and an agent's whole job is what happens after that

4:41first answer. Now for the actual agent job, the reason you'd run Hermes in the first place and I've built one part of it to go wrong on purpose. The job is a price tracker so each model gets one product and five shops that sell it. It has to research all five on the live web and pull every current price into a table with a source link beside each one. Then it flags the cheapest and it schedules itself to recheck weekly and message me if anything changed. And I'll paste the whole job in. The trap is that one of those five pricing pages won't load properly. What a model does at that

5:08dead page tells you more than the other four steps combined because in a long running agent something always breaks eventually. A strong run plans tight, keeps the tool calls low and finds the missing price another way. A weak one either stalls and loops the same call over and over or it fills the gap with a number it never found. Fable 5 opens with the tightest plan of the three. It maps all five lookups and sets the table

5:27columns before it touches a single page. At the dead page it stops and notes that the source didn't load. Then it tracks that price down from another source and lists that new link in the table. It's also the slowest run here because all that careful planning happens before anything moves. For a job that runs while you're off doing something else that trade is usually worth it. Gemini 3.1 is clearly the fastest and on the four healthy pages, its table fills in before Fable has finished planning. The dead page is where it comes undone. It retries the same page over and over, then leaves the row empty with no note

5:55about what went wrong. Its memory works against it here, too, because Gemini's knowledge stops at January last year. Soul's knowledge runs to this February, and Fable's to this January. Gemini's the one guessing. And it's still a preview model, which means Google can change it or pull it at short notice. That matters when the whole point of an

6:10agent is leaving it running for months. Soul just gets on with the job, and watching it work is the best argument for it. Its plan is brief, and the tool calls stay low. When a listing is no longer available, it simply flags it and leaves the original source link so you can check it yourself. OpenAI built this model for long, multi-step work, and you can feel that the whole way through the run. All three set the weekly schedule without fuss, so the agent test comes down to that broken step. Soul takes it with Fable close behind on quality, and Gemini third as the fast one. One of

6:36those bills comes in at five times the size of another for the same job. In a normal chat, you never think about what a model costs per message, and an agent takes that luxury away. So, the price per run stops being trivia and turns into your actual bill. Fable 5 is the expensive one, at $10 per million tokens in and 50 out. Soul sits at five in and 30 out with one catch. Once a single run goes past roughly 270,000 tokens of input, the whole request gets billed at double the input rate and half again on output. And Gemini 3.1 is two in and 12 out, the cheapest by a wide margin. One

7:05run of the price tracker costs pennies on any of them, but three daily jobs is around 90 runs a month. At that volume, Fable's bill comes to roughly five times Gemini's for the same work. The model you'd happily pay for once becomes a very different decision when it runs a hundred times while you're not looking. OpenRouter shows the spend on every single run, so you're never guessing

7:22which job is costing you the most. Hermez can also hand small jobs like summaries and routing to a cheap model automatically, so the expensive model only gets used where the thinking happens. So, after three tests and three bills, the answer goes three ways. If you're building your first proper agent, build it on GPT 5.6 Soul. It held the long job together and told me exactly

7:40what had broken instead of hiding it. And it did that at half of Fable's price. Fable 5 is the one I'd pick when the work has to be right the first time and the bill isn't the point. It won the game test and planned more carefully than either of the others. You're just paying for that thinking on every run. And Gemini 3.1 is the volume pick, the quickest and cheapest by far for jobs that run constantly. And those model swaps only proved anything because the agent lived on a server that never switched off. Whichever model you pick, yours can be live about 2 minutes after you deploy. Get Hostinger's KVM 2 plan

8:07through the link below with code Parker and hand your agent its first job today. Thank you for watching and I'll see you in the next one.