Most people are overpaying for the wrong AI model tier.

The AI model landscape splits into five distinct categories. Most people only ever use one of them — the expensive one — and then form an opinion about cost based on running every task through it.

“Flagship” is a price tier. It is not a promise. Several mid-tier and open-source models beat flagships on specific jobs, once you know which job. The skill is not picking the best model; it is knowing which category a task belongs to before you send it.

Here is the full map — category by category, with what each one is actually for.

The short version

• Flagship = the hard problems. Mid-tier = the actual work. Light = the volume.
• Specialized = the one job. Local = your data, your rules.
• Roughly 80% of daily work belongs in the mid-tier. Most overpaying happens because it is run on flagships instead.

1. Flagship models

Reserve these for the hardest reasoning. Top of the benchmark, top of the invoice — and the two facts are related in a way that is easy to forget when you are choosing a default.

• Claude Fable 5 is the smartest, but slow and expensive. Claude Opus 5 is the more honest alternative for most flagship-class work.
• Gemini 3.1 Pro owns multimodal and video.
• GPT-5.6 Soul bundles coding, browsing and image generation in one place.
• Grok 4.5 is cost-effective with real-time data.
• Kimi K3 is open source — near-flagship capability, and self-hostable.

The test for whether a task belongs here is not how important it feels. It is whether the reasoning is genuinely hard: multi-step, ambiguous, expensive to get wrong. Drafting an email is important and is not hard reasoning.

2. Mid-tier workhorses

Run roughly 80 percent of your daily work here. This tier has the best balance of speed, cost and capability, and it is where most people should live by default.

• Claude Sonnet 5 handles casual chat and coding.
• GPT-5.6 Terra is the cheaper flagship spin-off.
• Gemini 3.6 Flash is surprisingly strong at multimodal for the price.
• Open source is closing fast here: GLM 5.2, MiniMax M3, and DeepSeek V4 Pro for dirt-cheap math and reasoning.

This is the tier where the overpaying happens. If your default is a flagship, every routine draft, rewrite and code change is billed at reasoning rates for work that did not require reasoning. Moving the default down one tier and promoting individual tasks upward when they need it is almost always cheaper and rarely worse.

3. Light models

Deploy these for automations, sub-agents and bulk processing. Tiny, fast, nearly free — and completely adequate for work where the answer is largely mechanical.

• Haiku 4.5 is built for sub-agent work.
• GPT-5.6 Luna handles high-volume chat.
• Gemini 3.5 Flash-Lite is the fastest measured model available today.
• DeepSeek V4 Flash and MiMo V2.5 dominate high-volume coding jobs.

The category that benefits most from this tier is agent architectures. A sub-agent doing a file rename, a classification or a lookup inherits whatever model the parent was using unless told otherwise — which quietly bills chores at heavy-reasoning rates. Pointing that work at a light model is usually the single biggest cost reduction available.

4. Specialized models

Pick the job first, then pick the model built only for that job. General models will do most of these adequately; specialists do them properly.

• Code: Kimi K2.7 Code, Qwen 3 Coder.
• Image: GPT Image 2, Nano Banana, Meta Muse.
• Video: Veo 3.1, Runway Gen-4.5.
• Audio: Suno and Udio for music, ElevenLabs for voice.
• Enterprise retrieval: NVIDIA Nemotron 3 and Cohere for search pipelines.

The ordering matters. People tend to pick a model and then ask what it can do. In this category the productive direction is the reverse: define the single job, then find the model whose entire existence is that job.