# Do we really need Fable 5? > Source: > Published: 2026-07-08 03:02:11+00:00 To be honest, whether large language models reach the level of Fable 5 is not especially meaningful to most people doing actual work. It is not going to leave everyone dizzy and collapsing into their chairs as though they had just witnessed an atomic bomb detonation. Whether you can make use of Fable 5-level capabilities depends, in practice, on how you code. There are many degrees of agentic coding. First, there are the fundamentalist Vibe coders. These people unwaveringly follow the Andrej Karpathy path and remain committed to using AI throughout the entire programming process. At this point, they have essentially undergone mechanical ascension. The code they produce can be compared to humanity in the Warhammer 40K universe relative to real-world humans—in simple terms, they are no longer human. Generally speaking, this group has an extremely high ceiling and an equally low floor. Reaching the ceiling requires becoming the god of prompt engineering while simultaneously using harness prompts to condition the AI into a sex slave. The floor consists of people who cannot articulate what they want at all. The archetypal example is the client who asks for “colorful black,” after which the system produces nothing but indescribable, Cthulhu-like creatures. Next come the heavy users who treat AI as a cybernetic implant. The stronger people in this group do not let AI write everything. They first prepare the design documents and architecture documents themselves. Some go even further by implementing the critical sections personally, leaving placeholder functions everywhere else, and writing detailed comments beside them. The AI is only responsible for filling in the blanks. This means they can get their work done with relatively small, low-parameter LLMs, which none of the other groups can reliably do. The weaker members of this group may write worse code than the AI, but at least the AI provides a safety net. Their basic logic can still be made to run. Finally, there are light AI users. They occasionally use AI to build an MVP and minimally validate an idea, or ask it to fix a bug they cannot solve themselves. Most of their code is still written by hand, so whether AI exists makes relatively little difference to them. Some people lump all of these groups together under the label “Vibe Coding.” I do not think that is appropriate. A Guangdong man with twin-tails and Hatsune Miku may both have twin-tails, but you cannot classify both of them as virtual singers. So I will not go further into that discussion. So, which kinds of people are suited to which kinds of models? Let us divide model capabilities into six categories: coding ability, tool use, instruction following, long-context reasoning, the boundary of academic knowledge—in other words, world knowledge—and factual reliability. On top of those, I would add three secondary reference dimensions: multimodal capability, cost-effectiveness, and output speed. For fundamentalist Vibe coders, certain model capabilities must be extremely strong, because the model is the primary producer of the output. This requires strong coding ability and strong tool-use capability. At present, models in this category include the Claude 4.6 family—Opus and Sonnet—as well as GPT-5.5. For people who use AI as a cybernetic implant, the AI is purely a fill-in-the-blanks tool. The human is still primarily responsible for writing the code. Such users need a model with strong coding ability, followed by strong instruction adherence. Models that currently meet these requirements include DeepSeek V4 Pro and Flash, Qwen 3.7 Max, GLM 5.1 and above, as well as the two model families mentioned earlier. For this use case, I recommend Chinese models. When you are only using the model to fill in blanks, paying that much money to A\ and CloseAI is not especially sensible. There is another group that uses AI for refactoring. These people need exceptionally strong long-context reasoning and coding ability. Among Chinese models, only GLM 5.2 and Qwen 3.7 Max are particularly suitable for this kind of work. If you do not want your project to be “fixed” into complete ruin, you may simply have to grit your teeth and choose Claude or GPT. People who use AI to learn things need tool-use capability—because they need web search—along with world knowledge and factual reliability. The Gemini family and DeepSeek V4 Pro are the most suitable for them. Just remember: do not choose Gemini 3.5 Flash. Otherwise, you really will turn into a hissing kitten, glaring at the model and going “hsssss.” Some people even use AI to assist their thinking. I would not do that myself, but I respect, understand, and bless their choice. These users need models with strong long-context reasoning, extensive world knowledge, and high factual reliability. Uncle Liang’s DeepSeek V4 Pro and Google’s Gemini 3.1 Pro are the best choices for them. As for the absolute beasts who merely use AI as a wrench and could still tighten the nut by hand without it, model strength is largely irrelevant. They can apparently use something like MiniMax M2.5—what, that puny thing?!—to perform feasibility validation. Then there is multimodal capability. For coding, it should only be treated as a bonus. It is mainly during debugging that you will say, “Oh, this is good. This is really good,” with genuine appreciation. It should not be a deciding factor. For everyday use, however, multimodal capability is still quite helpful. Finally, I have to explain why I am not choosing MiniMax M3 or MiMo 2.5 Pro. Their output speed is just so damn slow!!!! Someone using DeepSeek V4 Flash will already have completed several rounds of debugging while you are still slowly polishing every detail—yes, I am specifically calling out M3. Some models even manage to produce bad work despite taking their sweet time—MiMo, I am looking at you. It is genuinely impossible to keep a straight face. There is, however, one extreme case when it comes to output speed: MiMo 2.5 Pro UltraSpeed. That is right. The fastest and the slowest are both made by Xiaomi. More than 1,000 tokens per second. Brute force works miracles. It does not matter if the code is bad. In the time it takes you to write one version, I can write one, then debug it several times over. That said, this thing is not currently part of the evaluation framework, because it is almost impossible to get access to. So, do you see the point? Very few people can meaningfully use the full capabilities of Fable 5. Unless you are Vibe Coding an entire project from scratch, Opus 4.6, GPT-5.5, or GLM 5.2 will be sufficient in the overwhelming majority of cases. The people selling anxiety about model capability are either shills for LLM providers and API resellers, or adorable little victims who have already fallen for the marketing. So why should we continue developing domestic Chinese large models? I do not want to imitate certain smartphone manufacturers or A:divide: by invoking some grand national narrative, so I will mention only one important reason: Exploring new and more efficient large-model architectures helps make affordable, high-performance AI available to everyone. Once stronger models exist, their more efficient architectures will gradually trickle down. Smaller models will gain higher intelligence, on-device AI will continue improving, and costs will keep falling. Eventually, everyone will be able to run a relatively small model with at least Opus 4.6-level capability on their own phone, computer, or workstation. That is how technological progress advances equality of access to AI. But Anthropic certainly does not see it that way. Your Uncle Dario still plans to spend the rest of his life piling parameters onto low-end models and squeezing out incremental upgrades like toothpaste. Even when Fable 5 arrives, it will merely be another instrument for extracting money. As for when that architecture will finally trickle down to Sonnet- and Haiku-class models, we may have to wait until the seas dry up and the rocks crumble. Want to read more perspectives? Visit [https://liandanxia.io](https://liandanxia.io)