Ranking Top AI Models (Tier List)
The tier list, updated for mid-August 2026
AI has changed a lot in just a few months. The team split up to test different models every day, and this tier list is the result. It is not about raw benchmark scores. It is about what actually makes sense to use when you are building real projects and watching costs.
The top pick is still Fable 5. For planning and talking through ideas, it is S+. It is hyper intelligent but lazy when it comes to coding. It routes most coding tasks to other models, so use it for planning your day, thinking through a task, or having a smart conversation. GPT-5.6 Solo is the other S tier model. It is the daily driver for heavy coding work and even app design. The team built dogdog with it. It is not the most brilliant model, but it is smart enough for most work and it has been getting frequent usage resets from OpenAI. That makes it a strong value choice right now.
The new challengers
Grok 4.6 dropped this week and it is no longer a joke. After Elon Musk bought Cursor, the post-training improved a lot. The team split on where to put it. One says S tier because it is extremely fast, especially for agents. The other says A tier because it drifts halfway through tasks and intelligence feels like GPT-5.6 Solo level, not better. The practical call is A+ for coding and S for agents when speed matters. If you have X Premium, Grok is included, and running Hermes on it is fast enough to make a real difference.
DeepSeek V4 Flash and V4 Pro are the cheap Chinese models the team was waiting for. A full benchmark run cost about $1, compared to $25 to $40 for other models. The castle it built was floating, which is expected from a lightweight model, but for $1, that is fine. V4 Pro likely lands at A tier and Flash at B. Kimi K3 is still smart and near S tier, but it is not cheap anymore. The team burned through the monthly usage cap by August 14 and got stuck. Until Kimi gets cheaper or moves to American servers, it sits between A and S. GLM 5.3 and Qwen 3.8 Max also land at low A tier. GLM seems smarter than Kimi but still expensive. Qwen Max is very good at website design because it is trained on that kind of work.
What to actually do
If you are new to AI, try Fable 5, Grok 4.6, or GPT-5.6 Solo. Do not be cheap with big decisions. The team mentioned a farmer who asked an older AI what to spray on crops and lost a field. Older models gaslight you more. For daily website and coding work, DeepSeek is becoming the default because it is cheap and you can top up credits whenever you need them. Do not buy yearly plans. The race moves too fast, and a model that is five times smarter can arrive in a month. Pay month to month, check resets, and match the model to the task.