Shortly after Fable 5 returned from its ban, Grok 4.5 and GPT-5.6 finally arrived on the scene. These models come with a long list of upgrades, but they all focus on improving the AI coding experience. As a hobbyist vibe coder, I tried out all three, continuing development of an app that breaks down replays of Magic: The Gathering Arena games and my Warframe build calculator. My main takeaways are that Fable 5 disappoints because of its painful usage limits and other quirks, while Grok 4.5 impresses with its efficiency and speed. However, GPT-5.6 steals the show, thanks to its top-notch intelligence. Read on for my full impressions.
GPT-5.6 and Grok 4.5 Debuted Just as Fable Reemerged #
Fable 5 isn’t new anymore, but it only recently came back online after its ban. Luckily, Anthropic just extended the period in which you can use Fable 5 with a premium Claude subscription instead of needing to buy dedicated AI credits through July 19.
Meanwhile, Grok 4.5 and GPT-5.6 are genuinely new models. Grok is currently offering limited-time free access to Grok 4.5 in Cursor and Grok Build, but you can also use Grok 4.5 in Grok’s chatbot interface via its free plan. Furthermore, Grok offers a free trial of its premium SuperGrok plan that supports increased Grok 4.5 usage.
With GPT-5.6, you get limited access to GPT-5.6 Terra on the Go ($8 per month) plan. If you want to try out the flagship, most intelligent variant of GPT-5.6, Sol, or the efficient Luna variant, you need to pony up for a Plus ($20 per month) or Pro (starting at $100 per month) subscription. Otherwise, you will largely end up using GPT-5.5, not the new GPT-5.6.
You can find more details about each model via their official announcements, but in general, Fable 5, Grok 4.5, and GPT-5.6 all promise to be smarter and more capable than their predecessors, especially with coding tasks. Efficiency is another major focus of Grok 4.5 and GPT-5.6; they aim to let you do more with fewer tokens and without hitting usage caps as often.
Before I get into the testing details, keep two things in mind. Most importantly, I’m not a programmer—I tested these apps from the perspective of a technically literate layman who just wants to make cool things. Additionally, I spent only a few days with Grok 4.5 and GPT-5.6, so my experience is somewhat limited; the punishing usage restrictions of Fable 5 also created some difficulties.
Ease of Use and Casual Queries: Fable’s Censorship and Grok Build Disappoint #
Coding aside, GPT-5.6 has the best chatbot experience. With a bevy of different modes and variants to choose from that increase intelligence at the cost of speed, whatever question you have, GPT-5.6 can answer it. Whether I asked for a deep analysis of a complex topic or just quick-yet-informed answers to questions, the response quality of GPT-5.6, much like GPT-5.5, is consistently high. Fable 5 is also impressively smart, but it suffers from frustrating censorship.
Completely innocent prompts, such as what happened with Fable 5 and the US government or what’s new in chemistry, can trip up its safeguards, after which it downgrades me to Opus 4.8. What’s even more frustrating is Fable 5’s inconsistent application of censorship. For example, sometimes, “What’s new in chemistry?” is an acceptable prompt for Fable 5 to answer, and sometimes it's not. Even if the chemistry prompt is fine, “What are the latest developments in biotech?” is still somehow problematic. GPT-5.6 doesn’t have this problem. In fact, it hasn't once downgraded me to a different model or told me my query violates some sort of policy.
Grok 4.5 stands out for different reasons. Like other iterations of Grok, it can automatically search X to inform answers, and it has no problem parsing websites with explicit content, either. This is unique functionality among AI models, and Grok 4.5 continues to excel within this niche more so than Fable 5 or GPT-5.6. However, in general, don’t expect big upgrades in response quality from any of these models over previous generations.
When it comes to coding, Anthropic’s Claude Code and OpenAI’s ChatGPT (formerly known as Codex) apps are easy to use, have tons of features, and work well. Both apps integrate excellent browsers that allow me to view my web apps while I work on them. Meanwhile, Grok’s in-house option, Grok Build, is limited to just a terminal-based interface, which isn’t nearly as user-friendly. However, considering SpaceX’s recent acquisition of Cursor, that might soon become Grok’s default coding platform.
Intelligence and Speed: Grok 4.5 and GPT-5.6 Take the Lead #
Although you can use any of these models to answer questions or do research, they’re meant for tougher tasks, such as coding. In terms of benchmarks, Fable 5, Grok 4.5, and GPT-5.6 all place highly: GPT-5.6 is usually at the top, with Fable 5 right behind it and Grok 4.5 hovering around the still-impressive level of GPT-5.5. My experience generally tracks with these benchmarks, but Grok 4.5 and GPT-5.6 surprised me.
Grok 4.5 is remarkably fast given its competence. The speed of Gemini’s 3.5 Flash model is even more impressive, but its shoddy work was disappointing, constantly leaving me with broken apps and half-finished tasks. Grok 4.5 is what I wanted 3.5 Flash to be—a responsive AI model that can harness the power of agents to complete most tasks without issues in just a few minutes at most.
For example, I wanted an easy way to back up all of the data my Magic: The Gathering app collects, pack it into a zip file, and provide the option to import backups. I explained this to Grok, and it added the feature in about a minute. I didn’t even have to bother setting up a memory file for Grok to follow certain rules or practices. And when I asked Grok to always boot my app after updating it, it successfully built that workflow much faster than Fable 5 or GPT-5.6. However, GPT-5.6 Sol still shines brighter. Its depth and intelligence are extraordinary, especially when it comes to spinning up agents to divvy up complex workloads. For example, one of my apps parses the log data of Magic: The Gathering Arena and then converts that data into turn-by-turn summaries of games. A big part of development involves dialing in those summaries so they are comprehensive enough to account for everything relevant in a game without including thousands of lines of unnecessary data.
During this process, I wanted to compare one iteration of a summary with an older version. I asked for this functionality using a short, one-sentence-long prompt. Nonetheless, based on my prompt, GPT-5.6 spun up agents to each analyze the summary of a particular game. Then, it created a new judge agent to evaluate each summary and give an impartial grade according to a rubric it generated. This not only let me understand the strengths and weaknesses of summaries, as well as how they were improving, but it quantified them, too.
GPT-5.6 is also significantly better than GPT-5.5 at designing interfaces. Without considerable prompt engineering when using GPT-5.5, you could end up with an interface straight out of the Windows 95 era. Meanwhile, Fable 5 and all the recent versions of Opus are capable of creating competent interfaces with relatively little guidance. GPT-5.6 now has this same capability. For example, below is the interface I created with Fable 5/Opus 4.8 for my app (first slide) alongside the interface I made with GPT-5.6 (second slide).
Fable 5 is also deeply intelligent, of course. For example, when I worked on my Warframe build calculator app with it, it was able to catch a data extraction issue related to weapon stats that GPT-5.5 and Opus 4.8 both missed. Furthermore, when I gave Fable 5 the same prompt I gave GPT-5.6 above to compare replay summaries, it went through a similar agent-driven evaluation process. On average, Fable 5 completed tasks slower than GPT-5.6, but the biggest problem with Fable 5 comes down to one thing: usage.
Usage Restrictions: Fable 5’s Aggressive Limits Remain Deal-Breakers #
Efficiency and usage are arguably the most important elements of an AI model. Unfortunately, usage caps are what ultimately make Fable 5 hard to recommend. First, only half the weekly usage allotted to you by your premium Claude plan goes toward Fable 5. Second, that’s only possible until July 19, after which you have to buy separate AI credits to use Fable 5 at all. Most importantly, the usage you do get during this limited-time window just doesn’t go very far.
For testing, I used both Claude’s and OpenAI’s $100-per-month plans (ChatGPT Pro 5x and Claude Max 5x) with Fable 5 and GPT-5.6 at their second-highest intelligence settings. On average, Fable 5 took roughly twice as much weekly usage to complete the same tasks. I also hit Claude’s five-hour usage limit constantly with Fable 5, often doing so across just a few prompts. That didn't happen with GPT-5.6 nearly as often or quickly. GPT-5.6, especially on its top intelligence settings or when it uses agents, does blow through usage much faster than GPT-5.5, but you still get significantly more access to it than you do Fable 5. What’s more is that OpenAI is just more generous with usage than Anthropic in general. OpenAI not only doles out usage resets fairly often, but you can also bank usage resets to use at your discretion, which Anthropic doesn’t allow. GPT-5.6 is also happy to keep working and pull from your weekly usage limit if it hits a five-hour usage limit during a task. However, even that isn’t an issue at the time of writing because OpenAI has, at least temporarily, done away with five-hour usage limits entirely.
Grok is worth calling out for its generous usage policy, too. Even on the $30-per-month SuperGrok plan, I was able to use Grok 4.5 in Grok Build for several hours without running into a usage cap or using even half of my allotted weekly usage. Grok doesn't have hourly limits, either, which makes for a lot fewer headaches. Just keep in mind that usage limits change constantly. With new policies, promotions, and random usage resets, alongside the rapid pace of new model debuts, what might work one day may not work the next, no matter what model you’re using. And if you constantly swap between different intelligence levels on models, these value equations become even more complicated.
Which of These New AI Models Should You Actually Use? #
Unsurprisingly, given all of the above, I just don’t recommend Fable 5 for most people, especially for vibe coders. However, if you already have a premium Claude plan and aren’t regularly hitting all your usage allotment with Opus or Sonnet, there’s no reason not to try Fable 5 until July 19, when it switches over to an AI credit system. If you have the money and your particular project doesn’t brush up against Fable’s censorship, it’s not meaningfully less capable than GPT-5.6, so you can use it to great effect if you prefer Claude’s ecosystem.
GPT-5.6 is the best overall option right now, but access to Grok 4.5 with a $30 per month SuperGrok subscription might be all you need. And rather than paying for an expensive $100-per-month plan from any company, you could augment a SuperGrok subscription with a ChatGPT Plus subscription, which only costs $20 per month, for a total of $50 per month. This allows you to take advantage of GPT-5.6 Sol’s top-notch intelligence for high-leveling planning and complex work while relying on Grok 4.5 for most tasks.
Sometimes, Like With GPT-5.6, the Hype Is Real #
I’ve written about the importance of not treating a new AI model like a fresh season of a favorite show you're just going to watch automatically. However, GPT-5.6 and, to a lesser extent, Grok 4.5, are new AI models worth your attention. Fable 5’s censorship and usage limits just don’t make it a compelling option right now. Of course, the AI landscape is constantly changing, so I plan to keep testing my apps on Fable, as well as Claude’s larger suite of models, as they develop.