GPT-5.6 vs Grok 4.5 vs Claude Fable: Who Wins?
GPT-5.6, Grok 4.5, and Claude Fable were built head-to-head. See which model won, plus the new ChatGPT and Claude app rebuilds.
> **TL;DR:** In a head-to-head build test, Anthropic's new Claude Fable beat both OpenAI's GPT-5.6 and xAI's Grok 4.5, producing the most visually realistic game world, the only Minecraft clone with working audio, and a photorealistic ad film. GPT-5.6 turned in a solid, functional build across a three-tier model family with an adjustable effort dial, while Grok 4.5 trailed with laggy, unfinished results. The launches landed alongside major app rebuilds at both OpenAI and Anthropic, adding new integrations, automation, and instant deployment tools.
Key Takeaways
- Claude Fable won a head-to-head build test against GPT-5.6 and Grok 4.5, producing the only Minecraft clone with working sound and the most realistic game visuals. - GPT-5.6 launched as a three-tier family (top, mid, light) with an adjustable effort dial from light to ultra and configurable autonomy settings. - Grok 4.5 finished last in the comparison, shipping a laggy open-world game and an incomplete Minecraft clone missing core mechanics like water and mob death. - OpenAI rebuilt the ChatGPT app around Work and Codex modes and added a '/sites' command that deploys live web apps instantly. - Anthropic rebuilt the Claude app around a Chat/Cowork home screen plus a separate Code section, with Fable selectable at effort levels up to max.
Three New Flagship Models, One Identical Test
The clearest way to compare AI models is still to make them build the same thing, and that's exactly what happened this cycle: OpenAI's GPT-5.6, xAI's Grok 4.5, and Anthropic's Claude Fable were each tasked with building an open-world, GTA-style game and a Minecraft-style voxel game from scratch. The results were not close. GPT-5.6 delivered a genuinely playable build, Grok 4.5 struggled to finish either project cleanly, and Claude Fable pulled ahead on both visual quality and technical completeness. For anyone tracking the [LLM Launches & Updates](https://speka.info/llm-updates/) beat, this round says less about benchmark scores and more about which model can turn a single prompt into a working, polished product.
GPT-5.6: A Three-Tier Family With a New Effort Dial
OpenAI didn't ship one model — it shipped a family. GPT-5.6 arrives in three tiers (top, mid, and light), paired with an adjustable effort dial that runs from light all the way up to ultra, plus separate autonomy settings that control how independently the model operates on a task. That combination lets a developer trade speed for depth on a per-request basis rather than picking a single fixed model size.
In the build test, GPT-5.6 held up well. It produced a solid open-world game complete with a functioning five-star wanted system, mirroring the escalating-pursuit mechanic familiar from the genre it was modeled after. Its Minecraft clone also worked as intended, with weather effects and block mechanics functioning correctly. It wasn't the most visually striking of the three builds, but it was coherent and complete — a meaningfully different outcome than what Grok 4.5 produced.
Grok 4.5: xAI's Entry Trails the Pack
Grok 4.5 was tested under the same conditions and came up short on both builds. Its GTA-style game was laggy and unfinished, and its Minecraft clone was missing core systems entirely — no mobs dying, no sheep, no water. Across the comparison, Grok 4.5 generally trailed both GPT-5.6 and Claude Fable in build quality. That's a notable gap for a model positioned as a frontier release, and it suggests xAI's current model still has ground to make up on the kind of multi-step, self-contained coding tasks that increasingly define how these systems get evaluated.
Claude Fable: Anthropic's New Flagship Takes the Win
Claude Fable, Anthropic's new flagship model, was rated the overall winner of the comparison. Its GTA-style build was the most visually realistic of the three, featuring dynamic lighting and a working sunset effect that neither competitor matched. Its Minecraft clone stood out for a different reason: it was the only one of the three builds with working sound and audio, a detail that's easy to overlook in a screenshot but substantial in a real playtest.
Fable's strongest result, though, was a photorealistic iPhone-ad film built entirely in HTML and three.js, described as being on par with quality from Apple's own production team. That's a striking claim for a model output, and it points to where the real competitive edge is emerging: not just functional code, but code that renders like finished creative work. Teams thinking about parallel investments in audio and voice — a growing part of the same generative stack — may also want to see how [ElevenLabs' voice marketplace lets creators earn royalties](https://speka.info/blog/elevenlabs-voice-marketplace-how-creators-earn-royalties) on synthetic voice work, since audio quality is clearly becoming a differentiator model-to-model, not just an afterthought.
Not every model race is about chasing the biggest flagship, either — smaller, efficiency-focused releases like [Bonsai 27B, PrismML's phone-ready LLM](https://speka.info/blog/bonsai-27b-prismmls-phone-ready-llm-explained), are targeting a different part of the market entirely, optimizing for on-device performance rather than head-to-head build contests.
Beyond the Models: Both OpenAI and Anthropic Rebuilt Their Apps
The model race wasn't the only thing that changed. Both OpenAI and Anthropic used this cycle to restructure their flagship consumer apps, and the changes point toward the same idea from two different directions: separate "just chat with me" experiences from "go build or manage something for me" experiences.
ChatGPT Splits Into Work and Codex Modes
OpenAI rebuilt the ChatGPT app around two distinct modes. "Work" is built for productivity tasks — documents, planning, research — while "Codex" is built for building apps, tools, and features, and for fixing bugs. Codex mode is explicitly positioned as usable by non-developers, not just engineers, which signals OpenAI's intent to push app-building further into the mainstream ChatGPT user base rather than keeping it siloed as a developer-only tool. That positioning echoes a broader trend toward accessible coding agents, similar in spirit to tools like [Juggler, JUCE's new GUI coding agent](https://speka.info/blog/juggler-juce-creators-new-gui-coding-agent), which is aimed at lowering the barrier to building real software.
Real Integrations and Automated Workflows
ChatGPT also picked up direct connections to a wider set of everyday tools: Chrome, GitHub, Notion, Google Calendar, ClickUp, Asana, Dropbox, Canva, and Figma. In practice, that means ChatGPT can take real actions in those services — the example given is creating a calendar event directly from a plain-language request, rather than just describing how to do it.
On top of that, OpenAI added a Projects feature that groups related work together in one place, and a Schedule feature that lets users set up recurring automations — daily briefings, weekly reviews — that run on their own. A side-chat feature lets users consult ChatGPT on an unrelated question without derailing whatever task is already in progress.
Codex also gained a deployment shortcut: typing "/sites" takes a built web page, app, or tool and deploys it to a live, shareable link instantly, with no manual hosting setup required. That's a meaningful reduction in friction between "the AI built something" and "other people can actually use it."
Claude Splits Into Cowork and Code
Anthropic made a parallel move. The rebuilt Claude app now opens to a home screen that combines standard Chat with a new "Cowork" mode — a no-code, productivity-focused experience conceptually similar to ChatGPT's Work mode — alongside a separate, dedicated Claude Code section for development work. Users can select their model, including the new Fable model, and choose an effort level running up to "max."
The overall shape of both rebuilds is the same: general-purpose chat, a no-code productivity layer, and a dedicated coding/building layer, now packaged as distinct modes rather than one undifferentiated chat window. For a broader view of how these launches fit into the wider model landscape, the [LLM Launches & Updates hub](https://speka.info/llm-updates/) tracks each new release and app change as it lands.
Frequently Asked Questions
Which model won the GPT-5.6 vs Grok 4.5 vs Claude Fable comparison?
Claude Fable was rated the overall winner, producing the most visually realistic open-world game and the only Minecraft clone with working audio, along with a photorealistic iPhone-ad film built in HTML and three.js.
What is GPT-5.6's effort dial?
It's an adjustable setting that runs from light to ultra, letting users control how much computation and depth GPT-5.6 applies to a task, alongside separate autonomy settings and three model tiers (top, mid, light).
Why did Grok 4.5 perform worse in the test?
Grok 4.5 produced a laggy, unfinished GTA-style game and a Minecraft clone missing core mechanics like water, sheep, and mob death, trailing both GPT-5.6 and Claude Fable in overall build quality.
What's the difference between ChatGPT's Work and Codex modes?
Work mode is built for productivity tasks like documents, planning, and research, while Codex mode is for building apps, tools, and features or fixing bugs, and is designed to be usable by non-developers.
What is Claude Cowork?
Cowork is a no-code, productivity-focused mode in Anthropic's rebuilt Claude app, sitting on the same home screen as standard Chat, alongside a separate dedicated Claude Code section for development work.
What does ChatGPT's '/sites' command do?
Typing '/sites' in Codex mode deploys a built web page, app, or tool to a live, shareable link instantly, without requiring any manual hosting setup.
