I Made Claude Fable 5.1 & GPT-6 Astra Automate The Same Accounting Work

I'm about to give Claude Fable 5.1 and ChatGPT's brand new GPT-6 Astra the exact same accounting work with the same files and the same instructions running side by side. Why? To answer one of the questions accounting firms ask me the most. Should we use ChatGPT or Claude? Because for a while, the answer has been simple for us, Claude. But there's been a lot of talk around ChatGPT's new model and how it may be a real competitor for Claude. So I want to put it to the test. In this video, we're going to take some of the accounting tasks we typically automate, give them to both models, and see how they do side by side. Side. That way, you know which one is the best for you to use in 2026. But first, we need to start why the decision between Claude and ChatGPT feels so difficult for firms to make. We even got it during one of our recent community Q&As. You see, up until recently, there was very little competition between the 2 tools. Claude has been shipping things much, much faster, things like skills before anyone else, and it got so good at spreadsheets and accounting work earlier than ChatGPT did. So you pick Claude, you build a few skills, you get your team using it. Amazing. But then OpenAI released GPT-4. GPT-6 Extra. And honestly, the gap looks a lot smaller now, which means instead of having just one default answer, it's gotten a lot trickier to answer what to use when. And every time you think you know and are ready to commit to a tool, a new model comes out, sending you straight back to square one comparing options. So you keep watching videos, LinkedIn posts saying how good a new model is, but that's not how you find which one is better. The only way to know for sure is to actually test it on the task your firm wants to automate. And that's exactly what we are about to do. All right, so here we are with Claude on Cowork with Fable 5.1 in High Effort, which is the middle of the 5 options. And we're gonna compare it with ChatGPT Work as the equivalent, both on their desktop applications, GPT-6 Extra and High as well here on the middle of 5 tiers. So this is as apples to apples as I can make it. We'll also provide both access to the exact same project. I'll make sure that we are working on my Computer, which is where I happen to have all the files, and that should allow me to add a specific folder that should be Demo Company Data right here. I'll make sure to do the exact same for Claude. So same files, same skills, same set of instructions and prompts. I'm going to make sure to run them at the same time so we can compare the speed, and then we'll score each one at the end. So let's start with test 1, which is the financial statement reviewer tool. You see, in most firms, every set of financials that goes out the door gets checked by a senior or a partner, right? But most juniors will miss things and nobody wants the clients to be the one who finds a mistake, which means your most expensive people are spending hours catching errors instead of doing the work only they can do. And you can't just hire your way out of it either, right? Like good accountants are hard to find and even new ones need months before you trust their work and you get them up to speed without a second review on top of it. So a review skill is meant to solve this. It runs your firm's checklist over every statement and compares it to the rules and requirements you laid out in the exact same way every time. So there's no wonder if someone overlooked something or missed an entry because their phone rang. And all that your seniors have to do is just review the issues that AI flagged at you. Now the question is, which of these models we're testing today is actually going to execute this better? Let's go ahead and find out. If we look at the folder structure we're working with, the skill financial statement review is available in this folder. And the files that we need are there. Technically, I give it access through the folder, but for the sake of comparison, I'll actually go ahead and drop both the skills and the files it needs directly into both tools. And we'll go ahead and hit enter as closely as I can to the same time. Let's see. So 1, 2, 3, go. They're both running. Let's see them side by side. This is asking, what would you like me to do with these files? I actually put a skill in there and it's just confirming, which I don't like and should be obvious. At least you can read through it, in my opinion. And I think that that's going to be what Claude does. But let's see. Okay, so we've got another question from ChatGPT. It's asking what firm and accounting software should I consider? Let's just say QuickBooks and Opsor Inc. I wonder if this is because, because it's further away, because I've noticed sometimes Claude does, it's able to work with your roadblocks and makes some assumptions and decisions for you. So we'll see here. Oh, there you go. It looks like we have it done already from Claude without it asking me for specifics and without asking me what to do. Worth knowing. Now we are also seeing the one from ChatGPT come up. So we will probably say it was about the same time that it took. But it took a little bit more back and forth from ChatGPT to get it working. I'll read through both and try to make a quality assessment of the things it caught in the first place. So doing my analysis here, there's a testing invoice that both models caught. There's July essentially unbilled or no expenses in July, and that's something that both caught as well. There's no depreciation on a truck expense. Both caught that as well. Opening balance, that's again something that both caught. I see miscellaneous expenses as something that both caught, but this is where effectively ChatGPT stopped. And there's a couple of other issues that Claude Vim brought up that are perhaps a little bit less potentially false positive, but it's like, hey, there's a negative income for a service. Is that a refund or what's the state of it? There's actually a refund account, so maybe that's where we should go. Undeposited funds. Inventory and a couple of additional things as well. So at least it sounds to me like almost a tie, but just because Claude did it without asking me for anything, so it's able to kind of make its own judgment calls, and because it caught a couple of additional things, I'll definitely give the point to Claude. Now, this reviewer is a skill I already use, and the same file runs on both tools, but it's far from the only one. We've actually put together 11 skills built specifically for accounting firms. So if you want to test or use them for yourself, feel free to grab them for free in the link below. That said, let's get back to test number 2. This time it's a financial model builder. See, a lot of firms are selling compliance, so the books, the statement, the tax returns, and that's being increasingly commoditized. So they ended up competing on price. This is the same thing every other firm is selling, but that leaves the highest margin of work, the advisory, talking with clients, projecting into the future, sitting on the table And the reason many firms don't offer advisory is because it sometimes involves a lot of things like revenue forecasting, which typically a senior has to, and disappearing to spreadsheet for a day to accomplish, but not with AI. So the idea here is just like a review, having a skill that builds financial models completely automatically. So all your team has to do is adjust the levers, know exactly kind of what variables to introduce and play with. And that's the next thing that we will be testing. I'm giving both models the same P&L and asking each for an editable 12-month Excel model that shows how this particular client can get to their goal of $100,000 in their first year. Let's go. Just like last time, we're gonna be starting fresh on both, but I have both the financial model skill, and in this case, I'll put just the profit and loss and balance sheet on both sides. And again, like I did earlier, I'll say absolutely nothing and just do the following. Let's give GPT the benefit of going first here. Boop, boop. Let's see, same thing again. What would you like me to do with this? Like, read the skill and execute. Same thing again. I'm being asked a question here from ChatGPT, actually from Claude as well. So let's see if it's the same question. ChatGPT asked the P&L covers from January to July with no forecast assumptions. What forecast period and assumptions should I use? I can create an illustrative August 2026 to July 2027. Yeah, both of them are effectively asking for the period. Whether it's next year, next 12 months. Let's just go with next 12 months for both of them. I'll ask another question. What assumptions should drive historical run rates? I'll say, use your judgment to hit the client goal, $100,000 in their first year. And it's possible I didn't give that to ChatGPT. So let's be fair. Let's make sure we give that to both of them. And let's do just projecting P&L only. Oh, we hit a roadblock and there's some usage limits I hit. Fortunately, I have some resets. So let's actually, as quickly as we can, free that up for us here so we can continue. But I think the speed comparison is going to be a little bit muddled unless there's a significant adjustment. Let's continue to wait and see. All right. So actually it's funny enough, our ChatGPT model is done at this point. Oh, there you go. And we just have Claude done as well. Hard to judge here as well. Literally within seconds of each other. I would suspect that if I hadn't hit the limits with ChatGPT, perhaps this would have done it like 10, 20 seconds more faster. So I'll just go ahead and read through both projections and see, do a little bit of a judgment call which one is better. I think for the sake of better comparison, I'd like that both companies right now do have a preview that you can see. I will, however, download the file to just go through both and assess which one of the two I would be most comfortable handing on to a client and using in a client conversation. So comparing this apples to apples here, effectively, this is ChatGPT, this is Claude. Both models look fairly similar. Both models have the actuals already included. So like the raw data underneath, although here I said it says Sheet 1, so Claude went ahead and just renamed it better, which I think is just a little bit of an interesting touch there. In terms of the variables, ChatGPT effectively simply relied on a percentage of revenue growth, OpEx rate change, expense rate change, and adjustments. Both of them interpreted the $100,000 as the goal for the next 12 months, which I think is fair because it's what I asked it to effectively map out for me. I do like that Claude really broke it down by the various accounts underneath revenue, where ChatGPT just showed revenue, for example. I think could provide some additional insight for the conversation with the client. And of course, let's actually go ahead and just test whether I was to change the goal. I see some variables changing as a result. If I change the percentage of COGS of revenue, I can see some of those changing the appropriate fields. So both of them at a quick glance seem to have been properly mapped to the right fields and all of that. Although I do like the Claude version, just a couple of additional details. I think this is pretty much a tie from my perspective. Now, if you don't want to spend your own time testing tools, building skills, figuring out which one fits each workflow, that's kind of what we do here at Opsur. So we're a done-for-you AI and automation agency for accounting firms specifically. And we find where your team's hours are going and build the skills, the automations, and the apps to take that work off of your plate. It's actually how we've helped many of our clients be able to take more clients and cut time on their month-end closes and their tax return work. So if you want to see what that could look like for your firm, Feel free to book a call using the link below. And that brings us to test number 3, which involves something that very few firms actually do— track their own numbers beyond their financials. And while a lot of firms will sell their own clients on knowing their KPIs, it's far too common that many of them are not doing this for themselves. Because if they have to see their capacity, their revenue by partner, and if they wanted to see any of those like more detailed metrics, the client churn, someone has to like stop client work and build an entire spreadsheet that nobody has time for, right? And usually what that means is that you don't realize you're over- capacity until busy season, it's already too late to hire, someone resigns, and you have to do all this mess to try to make up for that situation. The idea here is that with that KPI skill, we'll be able to keep track of our own numbers and make that a lot easier. It's going to pull straight out of your practice management systems, compare them to your targets, and turn them into a dashboard that you can check any week. Now, this is also the most complex task in this video. It requires quite a bit of data manipulation. So it's where I expect to potentially be the biggest difference between the 2 models. Let's test them. In this instance, what I have available is 3 exports effectively. An export from Ignition that assumes there's some like sales summary data, an export from clients from Carbon and organizations, and also an export from Work exports where we're going to have like the data of the budgets for the various work items, the data of who is assigned to them, who is the client manager, etc., etc. We also have a simple sheet that has here the team capacity of how many hours per week and the billing rate. And this should be enough for us to, alongside this skill, drop into both models and see how they perform. All right. So you already know the deal here. We're going to drop this on Claude. We're also going to drop the same amount of files here on ChatGPT. We're going to put them side to side, make sure they're both on the same level of effort. I'm gonna hit start task on both sides. Let's see. We see a pattern here. It's almost like ChatGPT always asking for confirmation. Execute skill. In this case, Claude also came up with a question around like what exactly you want them to do. I want you to build a dashboard. It's worth highlighting here that at least I can see ChatGPT giving me some additional feedback and inputs, right? It has told me, oh, there you go. Looks like ChatGPT is already done. While Claude looks like for the first test ever might be significantly slower this time around. It's suggesting it's rendering some screenshots. I believe this is something that ChatGPT also did just to render the file and make sure it's working properly. So while Claude finishes, I'll go ahead and open the dashboard and see what these metrics look like. So trailing 12 months revenue compared to prior year, lovely. Current AMRR, it's able to extract specifically which clients are recurring. Active clients, how many we have onboarded and offboarded over the last 12 months, how much work we have completed on time. This looks great. Revenue by service line, I like that. That's great. Clients by type, clients by billing type, and I can see here the number of active clients and where did we gain or lose clients. I like that chart. This is probably one of my favorite ones where now the team can see, oh, this is like a nice touch, either by type or by team member, either the next 6 months that I've planned or the trailing 12 months. How busy is my team and overloaded potentially and pulling up specifically from the budgets that they have assigned in our various reports. I see it's highlighting when they are almost at capacity, when they are closer to capacity, which makes sense that some of the tax people a little bit busier on the season, of course. And guess what? Funny enough, this looks pretty good. I mean, some of this was built into the skills. So ChatGPT has delivered and Claude is still running. We'll wait and see how long it takes. All right. So now after 5 minutes compared with closer to 2.5 or 3, Claude seems to be now wrapping up the work. And as you can see here, effectively a lot of the same metrics and exact same charts, because again, we're using the skills with both of them and that allows them to produce a lot more of a consistent output. So this is not going to look materially different, although here in terms of speed is the one area that in this instance ChatGPT won. So point for ChatGPT in this case. Now, every test so far has been work that lives within files you can upload, within exports you can download and drop into Claude or ChatGPT. But this next run is different, about a use case that has been promised for a long time. But very tricky to actually crack. And it's all about computer use, which basically means AI using your computer the same way you would, opening up an app, clicking, typing, and moving between screens on its own. For an accounting firm, that matters a lot more than it sounds because a lot of the software you rely on every day, tax software, sometimes some accounting software, depending on the clients you have, may only have a desktop app with no browser version at all, which means most AI and automation hasn't been able to reach those specific applications Since there's no web page to work inside. So the results of this round could matter more than anything else we've tested thus far. And it's also the first test because OpenAI is promoting computer use as GPT-6 Astra's main strength, while Anthropic is saying the newest model, Stable Diffusion 5.1, got better at it too. So we have both sides walking with a real shot. Now, I know what you might be thinking. Wait, it's gonna be controlling my computer? What are some of the security considerations for that? What is my IT team gonna say? But as I pointed in all my previous videos, both Anthropic and ChatGPT give you the option not to let this use your data for training. And you can actually toggle some of these features off if you don't want your team to have access to them. You also have controls over specifically which apps you allow them to access. So it's not that different than, you know, hiring a new team member and giving them a device and specific applications. that they can install. With that out of the way, let's see how this test goes because that's exactly why this runs on demo data and not a real client file. And that's how I tell any firm to start, you know, just testing it with some test files, watching every step, really supervising it before it touches live client work. But let me show you how both handle it. All right, so here's a desktop task that I'm going to put in front of both Claude and ChatGPT. I use this app to record videos. I'm not the one finally doing all the edits, but I do have to provide our YouTube team both a version that I have to remove the borders and then put it just like this, which is just my camera. Then I have to remove this and go ahead and export this file. So it's a relatively quick task for me, but, you know, someone would have to actually be able to open this up, tap here. You know, sometimes even I lose track here. Like, I don't know what I would have to do next here to to keep moving forward. I'm stuck. So I think it might be a fair task to put Claude on. There you go. Now it was removed, but it's effectively clicking this, removing the border, then being able to put it like this, clicking on export, naming it properly, and then allow it to effectively be saved. Now I will be opening up in this instance, we don't have our Claude computer use enabled for our team. So I'm actually go ahead and just move to a pro version. Just worth knowing, this is something you can enable. We just happen to have disabled for security purposes at this point in time. Similar concept, we're going to go for Fable 5.1. I'm going to ask of it the following. Hey, I am using Wondershare DemoCreator as a desktop app to provide the final export to my YouTube team. I want you to go ahead and open a current project that I'm working on right now, but it has to be just— I need 2 exports, one that is just me. And my face. Now the other one is just the screen. I want to test you by going through the first step of having the— creating the export for the camera only. So you'll have to make sure you navigate the software through my computer to be able to create that camera-only version that, as you'll see right now, the camera is only a little bubble on one corner of the Go ahead. Let me make sure I copy this for us to use it later and we'll let it run. Now, as you can see here, it's requesting app access specifically to one of the apps I have. That's part of its security protocol. So I'm going to allow specifically that. I see that it's listing windows. It's capturing the window. There you go. So it's opening things up. Actually, I would like to put this on another screen potentially so I can keep seeing. What it is thinking about. It's realizing here this needs full control over my screen. So I'll bring this back here and click on allow. Okay. So hands off. Let's see. This is so scary. I don't know if this is scary or exciting, but it feels emotional. It is slow though, right? Like let's just time it. This would be probably a 2-minute task for me. So we'll wait and see how long in this instance Claude specifically takes. Well, I'll feel free to interrupt it because it's been another minute or so and I don't see it doing anything unless it's doing it and I'm not seeing it. Let's see. Perhaps that's it, what it is, because I do see that it seems to have— oh, I can't move this over. Well, I see it's on the border. I don't know if I wasn't seeing that part of the screen, but it looks like it's moving forward. So I'll just like keep waiting here. This is something I like to do when I have a long task is just to prompt it to give me a little bit more feedback as it goes so that I know where it's at. It seems like it's stuck. At this point. So this might be a good, perhaps another point for ChatGPT. We'll see. Okay, so this is suggesting that Mac hasn't answered any of the last 4 calls. So I haven't seen the project and nothing in DemoCreator has been touched. I'm retrying now. Supposedly a prompt came up in my screen, but I see nothing. So I think it's failing. So I'll probably stop it at this point. We might come back to it in a second, but let's actually jump into ChatGPT. What suggests is the better model at this as of right now. And we'll have the opportunity to see if that's actually the case. So remember, we have our prompt here already loaded. So I'm going to go ahead and upload that, make sure it's on my computer. And let's see, actually, just real quick, I will make sure to not be in this section because I think there'll be a little bit of cheating. I wanted to realize that it can click on that in the first place. And let's even minimize it because this is the same state that it was in before Claude started. Same thing here. I have to provide some level of approval. There you go. It opened it up. I see that it puts it right here. This is a much better experience actually. So within the chat itself, I'm able to see from what I'm understanding here what it is doing. There you go. So actually this is at least from a user experience standpoint, even though it's taken over my screen, it's actually kind of like having a little bubble here And there you go. It looks like it clicked on importing. It clicked on save. I'm saving a separate project copy. Oh, you want to save? Oh, I'm doing nothing here. Go ahead and do what you want to do. I'm curious though, like, can I use my other monitor at the same time to keep moving forward? I'm afraid to test this right now, not to mess up. I just let it keep running. Rarely do I need to actually save the project itself. I just like export the final file, but I'll let it do what it wants to do. Oh, and there you go. I'm out of Codex usage again. Let's just— this is something I've noticed too. It does seem like both models are consuming a ton of credits lately. So we're out of that period of being subsidized by these companies and they're taking their sweet tokens and dollars to get stuff done. And this is a heavy task to use a computer because it's like taking a ton of screenshots, things like that. But I'm curious to see its current performance. It looks like it Remove the screen. It's— I think it's a fair step, although it would have to restore it later for the second step. All right, so as we can see, it now puts it bigger. It would just have to remove the border, and at this point it would be just one click away from exporting and having the test. There you go, it was able to remove the border. It clicked on export. Nice. Okay, I'll stop it here, but I can see that effectively, let's say, 1 minute plus 42 seconds 2 minutes in, it's getting the work done as expected and the same way I would have done. So I feel like here we really have a winner as well with ChatGPT. My summary is this. I do like that Claude doesn't need to be as prompted and doesn't come back to you with as many questions. And I like that when it comes with questions, it asks you a couple of questions at a time with easy options and recommendations. So it just makes it a lot easier for you to unblock Both models perform fairly well on most of the skills we provided. Claude seems to still struggle with computer use at this stage. We might test it more and that might vary in some future tests that we do, but ChatGPT seems to be the superior option there. And at least with that specific test and dashboard building that we saw, in this particular instance, ChatGPT-6 Astra was able to do it in half the time. And that's exactly how Claude and ChatGPT newest models stack up on the accounting work your firm does. But here's the thing, picking the tool is only the first decision because no one saves time in your team until you know which skills you need and how to use them. So if you want to see some breakdowns of real skill use cases for accounting firms in 2026, watch this video next. I'll see you there, and until then, stay efficient.