Fable 5 Review: Parallel Agents, Usage Limits, and Real Build Tests

By Beau JohnsonΒ·Β·8 min read

Quick answer

I ran three Fable 5 agents in parallel at once. One redesigned my landing page in 16 minutes, one built a first-person RPG in 11, and one made a Remotion demo video in about 40. Fable 5 hallucinated less than Opus 4.8, but its usage limits make me unsure it's worth paying extra credits for.

This is my review of Anthropic's Claude Fable 5. It comes after a 6-hour session where I used it to refactor my whole codebase. I also give it three real build tasks side by side: a landing page redesign, a first-person RPG and a Remotion demo video. Then I explain whether I think it's worth paying extra usage credits for. I recorded this in July 2026. Models, limits and pricing change fast, so check the current docs before you decide anything based on this.

What I cover in this video

  • 0:00: Six hours of refactoring put me at 28% of my weekly limit. I also cover the new usage-credit rule.
  • 0:47: A recap of the livestream where 9–10 Fable 5 agents worked on my codebase in parallel
  • 2:19: Why I think Fable 5 hallucinates less than Opus 4.8
  • 3:04: Handing out the landing page and RPG game tasks to separate agents
  • 6:53: How Fable 5 ranks on the LM Arena web dev leaderboard
  • 10:43: Playing the RPG Fable 5 built in 11 minutes
  • 12:18: Reviewing the one-shot landing page redesign
  • 15:21: Watching the Remotion demo video and giving my overall verdict

Six hours with Fable 5: what it fixed

The day before this video I ran a livestream with 9 to 10 Fable 5 agents working on my product's codebase at the same time. They restructured code, fixed bugs and shipped a new loop feature. It was a little slow, but when I asked it to fix something, it fixed it.

I gave it one broad instruction: wherever Opus 4.8 made mistakes, go back and fix them. It went back and fixed problems with the MCP setup and with pushing to GitHub. What impressed me most is that it finishes the job without making things up. It does still hallucinate sometimes, but less than Opus 4.8 did. With Opus 4.8 I had to get GPT-5.5 on extra high to check a lot of its work before pushing to production.

The catch is usage. Anthropic's announcement says Fable 5 is included for up to 50% of your weekly usage limits through July 7. Once you pass that, you pay usage credits, and after July 7 all Fable 5 usage goes on credits. Six hours put me at 28% of my weekly limit. That probably leaves me a day or two before I'd have to pay extra.

Three agents, three tasks, running in parallel

To see what it can really do, I ran three named agents at once, all on extra high. I talked to them by voice inside the environment we're building, ShipSpace, which we're rebranding to Beacon (heybeacon.ai). If you're new to running several agents together, my AI agents page covers the basics.

Alder: landing page redesign (16 minutes)

Here's the instruction I gave Alder:

Take a look at our new landing page and implement any changes. I don't really like the hero section video anymore. I'm going to hand you a remotion video to have play automatically or save the spot for it below the hero section text where it says "Watch demo" and it's going to auto play, but keep the parallax effects. Just take a look at the landing page, make any fixes according to principles that you find in the open design GitHub that we just gave you.

It reworked the hero framing and turned the scroll into one sequence: hero, unfolding frame, morphing features, demo pitch, then pricing. It also committed the changes. For one shot with very little direction, the result was good, and the parallax effect worked. With Opus 4.8, it took me several rounds just to get the parallax working and for the model to understand what I meant.

Cedar: a first-person RPG (11 minutes)

I asked Cedar to build a creative first-person RPG using modern techniques and the best graphics available, then open it on an HTML development server. It made a game called Lumenvale, with HDR bloom, ACES tone mapping and sound generated with the Web Audio API instead of audio files. You can jump around, and it explains the controls. It's fun to play for something built in 11 minutes. But Sonnet 5 made a very similar game a couple of days earlier, so my verdict is good, not great. For more on building games and apps from prompts, see vibe coding.

Juniper: a Remotion demo video (about 40 minutes)

A Remotion video is a video made with code, so the agent can put it together itself. Juniper used the Remotion skill, took screenshots of our product and turned them into a structured demo for the landing page. It took 38–40 minutes and has no audio because there was no licensed music. When I watched it on localhost 3001, it even showed the workbench feature. It's one of the best Remotion videos I've seen. Afterward I told Alder to add it to the new landing page.

What we shipped in the workspace

While the agents worked, I showed some features we've added. The fleet view shows what every agent is doing. Click an agent's name to get a summary of its work, then continue the task or ship the update. The workbench lets several agents work on one task in parallel and run it in a loop. A loop pop-up lets you control loops across agent swarms.

Beacon is meant to be a development environment you control by voice, with a wake word. You can run terminals, browser previews, MCP, knowledge, memory and the workbench just by talking. It supports multiple workspaces. Voice runs on GPT Realtime 2 with your own API keys, and voice-to-text runs locally on NVIDIA Parakeet. You can also switch between static or video backgrounds.

Is Fable 5 worth paying extra for?

On the LM Arena web dev leaderboard, Fable 5 is clearly in first place, with GLM 5.2 next. That's why I'm testing it while it's still included in my plan. But OpenAI is expected to release GPT-5.6 within the next week as three models. On Terminal Bench 2.1, the top GPT-5.6 model scored 91.9% and Fable 5 scored 84%. Fable 5 also loses to it on a lot of other benchmarks.

Most builders aren't loyal to one company. They want the best model at the best price. I'm on the $100 subscription plan, and I expect to get a lot more use out of GPT-5.6 there than out of Fable 5 on usage credits. If you're comparing the two ecosystems, see Claude Code vs Codex.

Key takeaways

  • Fable 5 fixed many of the bugs Opus 4.8 left behind and hallucinated less in my testing.
  • It redesigned my landing page in one shot with working parallax, which took Opus 4.8 several tries.
  • The RPG it built in 11 minutes was good, but not much better than what Sonnet 5 made.
  • The Remotion demo took about 40 minutes and was one of the best I've seen.
  • Usage limits are the real problem. Six hours used 28% of my weekly limit, and after the 50% cutoff you pay credits. I'm watching GPT-5.6 because of that.

FAQ

How much Fable 5 usage is included in a Claude subscription?

When I recorded this in July 2026, Anthropic's announcement said Fable 5 is included for up to 50% of your weekly usage limits through July 7. After the 50% mark you pay usage credits, and after July 7 all Fable 5 usage is on credits. Check Anthropic's current docs, because this may have changed.

Is Fable 5 better than Opus 4.8?

In my testing, yes. It fixed bugs Opus 4.8 had left in my codebase, hallucinated less, and got a parallax landing page working in one shot, which took Opus 4.8 several tries.

How does Fable 5 compare to GPT-5.6?

On Terminal Bench 2.1, the top GPT-5.6 model scored 91.9% and Fable 5 scored 84%. Fable 5 also loses to it on a lot of other benchmarks. I expect GPT-5.6 to be cheaper to run and included in the subscription.

How long did Fable 5 take to build each project?

Running on extra high, the first-person RPG took 11 minutes, the landing page redesign took 16 minutes, and the Remotion demo video took about 38–40 minutes.

Transcript

Read the full transcript

0:00Yesterday I used Fable 5 to refactor my entire codebase for my product and it took 6 hours. And you can see here my current weekly limit is at 28%. And if you haven't heard, they are now starting to charge you usage credits once you pass your 50% weekly limit. This goes until July 7th and then all of your Fable 5 usage will be on usage credits. This is crazy. Now, ChatGPT is going to be releasing their newest version within the next week and 5.6 scores significantly higher on Soul Ultra compared to Claude Fable 5. On Terminal Bench 2.1 it scored a 91.9% compared to Fable 5 scoring 84%. And in

0:47this video, I just want to share with you some of my thoughts using it for 6 hours yesterday. If you didn't catch the live stream, you should go back and watch it. And we were running 9-10 Fable 5 agents in parallel going at my codebase, restructuring things, fixing bugs, updating, implementing the loop feature into our product. And it's was amazing. It was a little slow. It did fix things when we asked it to get them fixed. I just don't know if it's going to be worth paying extra usage credits. Right now, I have it working on a remotion video for our landing page. A remotion video is basically a programmatic video that it edits itself that I can put on to my landing page demoing and showcasing our product. So

1:33it's doing that right now. I have it in this terminal here, Juniper. And basically it is taking the remotion skill, taking screenshots of our product, and it is applying the prompt to that to make a structured video that is going to be put on our landing page. I'm excited to see how it does. I've also had it kind of refactor everything in our code base. I said basically wherever Opus 4.8 made mistakes, go back and fix them. It fixed things with the MCP, it fixed things with pushing to GitHub. So, it did go back and correct a lot of issues. The thing that I'm most impressed with though is that when you tell it to do something, it actually gets the job done without hallucinating. Not saying it doesn't hallucinate, I

2:19just think that Opus 4.8 had a tend it tended to hallucinate more. I would have to have GPT 5.5 on extra high checked a lot of the work to make sure it was correct before we pushed to production. So, I'm going to be testing and pushing the limits as far as like what Fable 5 can do. Kind of like a demo video of just like what are its capabilities. So, it's working on the remotion video here. I'm going to have it make some changes to our landing page as well just to see the difference. And am I going to also have it create as a game just to be able to see how it is for vibe coding a game. I mean, you've seen games done by Sonnet 5, Opus 4.8, they're very lackluster. Let's push that to the limits as well.

3:04So, I'm going to go and tell Alder, "Can you tell Alder to take a look at our new landing page and to implement any changes? I don't really like the hero section video anymore. I'm going to hand you a remotion video to have play automatically or save the spot for it below the hero section text where it says "Watch demo" and it's going to auto play, but keep the parallax effects. Just take a look at the landing page, make any fixes according to principles that you find in the open design GitHub that we just gave you." >> I'll pass those landing page instructions to Alder. Sent to Alder. >> And can you tell Cedar to spin up a

3:52first-person game that is more of like an RPG. Be very creative, use modern techniques, use the best graphics available, spin it up, and then open it up in an HTML development server so I can play it once it's done. >> All right, I'll pass that game brief to Seeder and get it started. Send to Seeder. >> All right, so it's going to work on extra high on both of these terminals here. And Juniper here in the middle is working on the remotion video. We'll check back in a few minutes to see how it does. So while we're waiting for that to happen, we implemented a loop feature and a fleet feature where we could see what our agents are doing at any given time. And we can click on the agent's name and it will give us a summary on

4:38what the agent has done, as well as the ability to continue the task or ship the update. So that's something that I've just built for myself, but it can really help builders when they're building and doing a lot of testing and they don't quite remember what was happening inside of a particular terminal. Another thing that we shipped is a workbench feature. This workbench feature allows agents to work in parallel upon a task and actually run them in a loop. So if you click here on agent annotation MCP tools, you can see that I have the ability to run it in a loop by clicking this button. I've also shipped a feature just in this recent update here where we can actually have a loop modal pop-up and control all of our loops by agent

5:23swarms and how they actually go about running this process. I also want to mention real quick, we opened up a community that is 100% completely free, over 130 people in it, and that you can join alongside other builders building cool projects, sharing what they're doing with Fable 5. There is a paid upgrade within the group where you have access to courses and four live calls every single week. Right now we're into the local LLM machine learning and AI context engineering, but we also have courses on Hermes, Open Claw, Claude Code, and Codex, and actually building games using Codex within a loop. So, there's a lot going on here. I'll put the link to it down below. 100% free to join, and you can join a community of like-minded people. You can find the

6:08link in the description. I also want to show you how the Ship Space website looks right now. We are in the middle of a re- brand. It is called heybeacon.ai. Beacon meaning to sudden. We are building an agentic development environment that is completely controlled by voice, which is going to have a wake word, and you're going to be able to control your terminals, browser previews, MCP, knowledge, memory, and the workbench all by voice, and be able to interact with your workspace. So, this has been a complete refactor as of yesterday, and I think it's for a good cause. Because sometimes you get set on a name, and then you start iterating and iterating, and you change. So, we are now rebranding, and I want to see how

6:53Fable 5 does with this rebrand. You obviously, you know, the space theme needs to be changed, and I want to just see how it does with design skills. So, this is LM Arena, and it is for design. If you actually go to code, and you click on web dev, it's going to give you a ranking. And this is going to paint a beautiful picture here. Claude Fable 5 absolutely destroys when it comes to web development. And you can just see how well it scores compared to other models. JLM 5.2 is actually the next one on this list, which is pretty impressive as far as web development. But you can see Fable 5 it clearly winning at a 16.53 Elo score, which is impressive. So, that's why we're putting it to the test

7:38while we still have access. This is important because we're not going to have access forever. And I just mentioned that because in this press release that came out a few days ago, they literally say Fable 5 will be included for up to 50% of your weekly usage limits through July 7th. That means you only have another 5 days to use this model. If you've already hit your 50% weekly usage limits, I'm sorry, you're going to be paying for extra usage when it comes to Fable 5. Like I said, I used it for 6 hours yesterday. I'm already at 28%. So, I probably have another day or two usage out of it before I actually have to pay extra to use this model. Kind of sucks. This is why Chat GPT 5.6 releases imminent. They are going to take over and where

8:24Anthropic has fumbled the bag and rug pulled its users, people don't care. They're not married to one or the other. They just want the best model at the best price. So, if OpenAI releases GPT 5.6, Soul, Luna, and Terra and it's on the subscription plan, are you even going to care about Fable 5 because Fable 5 actually loses in a lot of benchmark scores compared to Soul Ultra. So, that's why we're testing Fable 5, seeing if it's actually worth paying extra for, and OpenAI is clearly has an opening to take over this model race. We have entered a new era where you can literally just build things now, but we don't want to be spending thousands of dollars in usage credits or API calls. That's where I think this 5.6 release is

9:11actually more exciting than getting Fable 5 back. Fable 5 has been great. You could see it working on auto mode. It's been working for 8 minutes here, 12 minutes here, and 6 minutes here for Cedar, Juniper, and Alder, and it's been doing great, but I'm more excited about GPT 5.6 and what it offers because I am on the $100 subscription plan, and I think that we'll be able to get a lot more usage out of those frontier models versus Fable 5. One of the coolest features that we shipped within Backed In is the ability to have multiple workspaces and the ability to control your entire workspace with voice. I mean, we are living in a new day of building. No more staring at a boring terminal anymore. You need to be able to have voice locally, which we use GPT

9:56real time two, bring your own keys, and we also use a local Nvidia Parakeet that allows you to also have voice to text commands within Backed In. Also, you can change your appearance. So, if you just are not set and you're feeling a little lackluster on your inspiration, you can also change between different appearances for backgrounds that can be either static or a video. Whatever you want to do, it doesn't matter. We also push updates every single day as we make the app build better. Okay, so you could see this pane is blinking right here. Cedar has finished the game and it took 11 minutes. So, I'm going to check it out. First, I want to check in on the review. Hey Backed In, can you check in on Cedar and give me a summary of the

10:43game and just let me know how it all works and how to open it. >> All right, let me check Cedar's latest output and then I'll walk you through the game and how you open it. Cedar's building a creative first-person RPG with a glowing immersive look using an HDR bloom and AC is tone mapping shade plus synthesized audio via web audio. Since there are no audio files, the scene and HUD render and the start screen up >> Okay, can you tell Cedar to open the game up? I don't >> Let me pass that instruction along to Cedar now. Sent to Cedar. >> Okay, the game is up and running here. Let's see, Lumenvale, the four shrine beacons have gone dark and Shadow Wisp drink little light remains. Okay, interesting. Wow, incredible. I can

11:30jump. It is a little dark. Interesting, in 11 minutes. Beautiful. I mean, for a game built in 11 minutes, I mean, this is something you can have fun with. Um what else can I do? Interesting. So, yeah, is this going to replace GTA 6? Probably not, but still the fact that you can spin up a game in 11 minutes, and obviously you can take it a bit further and make changes. Not too impressed because we did the same thing with Sonic 5 a couple days ago, and it did a very similar game. It did come up with more instructions on how to control drink drought, release mouse, light burst. So, 11 minutes, not terrible. Good, not great, I will say. Okay. Now, you can see the agent Alder right here has finished, and it took 16 minutes on

12:18extra high to update my landing page. Hey, Backen, can you check to see what Alder did to the landing page and just give me a quick summary on what did. >> Let me take a look at Alder's recent output, and then I'll give you a quick summary. Alder reworked the landing page flow and visuals. The hero section's framing and composition were adjusted, and the scroll journey now transitions through the hero, an unfolding frame, morphing features, the demo pitch, and then pricing. The changes are committed, and it's ready for you to review visually with a simple swap planned later when the promotion film is ready. >> Sweet. Okay. So, let's take a look at this and just see how it did. You saw the before. Now, let's take a look at the after. Tell Alder, or if you could do this, Backen, can you just open up

13:04the newly made landing page in a browser preview, so I could see it within the workspace? >> I'll open a live preview in the workspace, so you can inspect the landing page there. Opening the landing page preview now. >> All right. Let's just see how it did. I don't know how it did. I just want to just take a look. This is just one shot with very minimal direction. So, hey, Backen, build it. Mission control for your AI coding agents. You your own cloud code, code X, and cursor working in parallel on one workspace summon by voice. Okay, wow, okay, good parallax effect. Awesome. You back end, they build. Awesome. Trying to really avoid the vibe coded look. There I mean, for just one shot, I think it did a really good job, but there is a few little things that I want to do to make it

13:50better and better than Opus 4.8 giving it a try. With Opus 4.8, it took multiple iterations to get the parallax effect even working and for it to even understand what I was saying. So, I give it a good try, much better result. Let's check in on Alder to see the demo video. Awesome. I love that. Juniper is still actually making the remotion video. It has been going for 34 minutes. Let's check back when that's done because I want to implement it into the landing page. Okay, 38 minutes later, it is finished. And I am going to just see how it did. Hey back end, can you talk or see with Juniper of the summary of the remotion video? Just give me a quick

14:36update back here. Tell me how it works and how I can view it or maybe pull it up in the browser number one. >> Let me check in with Juniper and then I'll walk you through a quick update sent to Juniper. Quick update. The remotion demo appears to be structured with a cohesive product and narrative including title and outro, but it's currently silent because there's no licensed audio. And one narrated voice command is a bit aspirational compared with the current interface. You can try viewing it in the browser pane at localhost on port 3000. If it doesn't load, Juniper will need to confirm the exact start command and URL for the local preview. >> Tell Juniper to open the back end demo

15:21on localhost so I can see it because I don't see >> All right, I'll nudge Juniper to bring up the local demo for you. Sent to Juniper. >> Hey Becca, can you have Juniper open the browser preview with localhost 3001 within this workspace or maybe you could just do that for me. >> Opening that preview for you now. >> All right. >> Opening the localhost 3001 preview. >> Let's check out this video. I have not seen it. Wow, it did the workbench. Awesome. So it did a good job with that. This is one of the best re-motion videos I've ever seen. Awesome. I mean, I'm pretty impressed. Yeah, it took 40 minutes to build, but that is pretty impressive. Wow, pretty

16:06cool. Pretty cool. Hey Becca, tell Alder that the re-motion video is ready. It's called Becca's demo. Can a apply it to the new landing page? >> Got it. I'll pass that along to Alder now. Sent to Alder. >> Okay, so basically, you know, Fable 5 is obviously the most capable model that we've had our hands on as builders, but just around the corner in the next week, we are waiting the GPT 5.6 release of Soul, Terra, and Luna. Those are the three different models that they're releasing within that version and I think it's going to massively outcompete Fable 5, not only because it's going to be cheaper to run and be included in your subscription, but because the benchmarks are already starting to come

16:52out that this is going to be the most capable model. Now we are judging these model releases whether they get regulated by the government or not. So if it's not being regulated by the government, it's probably not a model worth testing. Beautiful day to be alive, to be building your ideas. Drop in the comments down below, what are you building? How are you using Fable 5 or are you just waiting for GPT 5.6 to come out? If you want to be in part of the community, it is a free community of 130 builders. I'll put the link in the description down below. If you want to join the waitlist for Back End, check out the link down below in the comment section. We're going to be launching this on public beta within the next week. We'll see you in the next one. Have a blessed day.

Ready to start building with AI?

Join Shipping Skool free β€” 4 live build sessions a week, courses, and a community that ships.

Join Shipping Skool

Keep reading