Which AI Chat Model Should You Use?
FrameTide runs fifteen chat models from eleven labs, spanning a 100x cost range. Here is how to route a task to the right one instead of always reaching for the most expensive.
The most expensive model in FrameTide costs 100 times what the cheapest one does. Not 100% more — 100 times.
That is the number worth internalising, because the usual instinct is to pick the strongest model and stop thinking about it. On a single question that instinct costs you nothing. Across a working week it is the difference between a plan that lasts and one that does not.
The lineup, by what it costs you
Credits per message in FrameTide, cheapest first:
| Credits | Model | Lab | Reach for it when |
|---|---|---|---|
| 1 | DeepSeek V4 Flash | DeepSeek | anything routine |
| 5 | Gemini 3.6 Flash | fast work with long inputs | |
| 8 | DeepSeek V4 Pro | DeepSeek | hard reasoning on a budget |
| 8 | Qwen 3.7 Plus | Alibaba | multilingual work, code |
| 8 | Kimi K3 | Moonshot | agentic and long-context jobs |
| 8 | GLM 5.2 | Z.ai | bilingual reasoning, tool use |
| 8 | MiniMax M3 | MiniMax | long documents and mixed media |
| 8 | MiMo V2.5 | Xiaomi | fast reasoning, tool calls |
| 8 | Tencent HY 3 | Tencent | general reasoning |
| 20 | GPT-5.6 Luna | OpenAI | everyday work in OpenAI's voice |
| 30 | Gemini 3.1 Pro | deep reasoning over huge context | |
| 40 | GPT-5.6 Sol | OpenAI | the hardest OpenAI-shaped problems |
| 40 | Grok 4.5 | xAI | long agentic coding runs |
| 50 | Claude Sonnet 5 | Anthropic | writing and code you will ship |
| 100 | Claude Opus 5 | Anthropic | when being wrong is expensive |
Two things jump out. There is a large, flat middle tier at 8 credits containing most of the open-weight field — and the gap between that tier and the frontier models is 5x to 12x.
The only question that matters
Not "which model is best" but "what does being wrong cost here?"
That single question routes almost everything:
- Wrong is free. Brainstorming, rephrasing, summarising something you will read anyway, drafting text you were going to rewrite. Use DeepSeek V4 Flash at 1 credit. You will notice the quality difference less often than you expect, and when you do, rerunning on a stronger model still costs less than starting there.
- Wrong is annoying. Code you will review, research you will verify, a first draft with a real audience. The 8-credit tier is built for this. DeepSeek V4 Pro and Kimi K3 both post frontier-adjacent coding scores at a fraction of frontier prices.
- Wrong is expensive. Something going to a client, a migration you will not read line by line, a decision you will act on without checking. This is what Claude Opus 5 and GPT-5.6 Sol are for. Paying 100 credits to avoid one bad merge is a good trade; paying it to rename some variables is not.
What each lab is actually for
Anthropic — Sonnet 5 (50), Opus 5 (100). The strongest general writing in the lineup, and the pair most people reach for on code they intend to ship. Sonnet 5 carries a 1M-token context and posts 72.7% on SWE-bench Verified; Opus 5 sits above it with better self-verification and a more recent knowledge cutoff. Expensive, and usually worth it at the end of a task rather than the start.
OpenAI — Luna (20), Sol (40). GPT-5.6 arrived as distinct tiers rather than one model with an effort dial. Luna is the fast, cheap end; Sol is the flagship, leading Terminal-Bench 2.1 among the models here. Pick Luna for everyday work and Sol when a problem has genuinely beaten a cheaper model.
Google — Gemini 3.6 Flash (5), 3.1 Pro (30). Both carry 1M-token context. Flash is the value pick of the whole lineup for anything involving a lot of input text: at 5 credits it reads more for less than anything except DeepSeek Flash. Pro is for reasoning that has to hold across that entire context.
xAI — Grok 4.5 (40). Built for coding and agentic work, trained on real editor sessions, designed for long runs across many tool calls. Its context dropped to 500K from the previous generation's 1M, which matters if you were feeding it whole repositories.
DeepSeek — V4 Flash (1), V4 Pro (8). The price-performance story of the year. Both carry 1M-token context, and V4 Pro posts one of the highest open-weight SWE-bench Verified scores available. Flash at 1 credit is the correct default for most people.
The open-weight field — Kimi K3, GLM 5.2, Qwen 3.7 Plus, MiniMax M3, MiMo V2.5, Tencent HY 3 (8 each). All the same price here, so choose on character rather than cost: K3 for agentic pipelines and frontend code, GLM 5.2 for bilingual reasoning and tool use, Qwen for multilingual breadth, MiniMax M3 for long documents and mixed media, MiMo V2.5 for fast omnimodal input, HY 3 for efficient general reasoning.
A routing habit that saves real money
- Default to 1 credit. Start every conversation on DeepSeek V4 Flash. Most turns never need more.
- Escalate on evidence, not anxiety. Move up only after a cheaper model has actually failed on this task. "It might not be good enough" is not evidence.
- Escalate one step. Flash to the 8-credit tier before you jump to 100. The middle tier resolves more than people expect.
- Spend at the end, not the start. Explore cheaply, then run the final pass on the strongest model. The reverse — frontier model for exploration, cheap model for delivery — is the expensive mistake.
- Switch inside the conversation. The model picker sits in the composer; changing it does not lose your thread. Escalating a single hard turn costs one message at the higher rate, not a whole conversation.
What credits do not tell you
Credit cost tracks provider price, and provider price tracks capability loosely at best. Three things it does not capture:
- Context. Several 8-credit models carry 1M-token windows. If your bottleneck is how much you can feed the model rather than how hard it thinks, cheap and long beats expensive and short.
- Language. For Chinese work, GLM 5.2 and Qwen are frequently stronger than their price suggests, and the frontier premium buys less than it does in English.
- Shape of the task. Grok 4.5 is built for long agentic runs; MiniMax M3 for mixed media; K3 for frontend code. A specialist at 8 credits can beat a generalist at 100 on its home ground.
The short version
Route by consequence, not by reputation. Draft at 1 credit, work at 8, and spend 50 or 100 only where a mistake would actually cost you something. Most teams have the ratio backwards and pay for it every day.
Sources
Start creating in FrameTide
Open the Image workspace, describe what you want, and compare models on your own prompts.