Which model should I actually use?
The labs, the benchmarks and their catches, and what models cost. Ends with how to choose.
The path.
- The Labs: Who's Who Snapshot as of July 19, 2026. Valuations and funding are dated and sourced.
- Frontier Model Comparison A dated snapshot of the current top-tier, closed ("frontier") models, meaning the API-only models from the major labs that set the capability bar.
- Open-Weight Models The models you can download, self-host, and fine-tune. As of July 2026. Two things to know: (1) "open weight" ≠ open source.
- Benchmarks & Leaderboards The single most useful mental model here is that benchmarks tell you what a model can do on someone else's task.
- Pricing & Cost Reference Pricing is a Frontier Model Comparison snapshot. This section is the durable part: how to reason about cost so you're not surprised by a bill.
- How to Choose a Model The honest answer is to shortlist from benchmarks and reputation, then run a small eval on your actual task.
Check yourself.
One question per step. Take it cold to find where to start, or after reading to see what stuck. Nobody's grading you.
Best so far:
Question 1 of 6
-
As of mid-2026, who leads the open-weight frontier?
The open-weight leaders are now DeepSeek, Alibaba's Qwen, Moonshot's Kimi, and Zhipu's GLM, while Meta has pivoted toward closed frontier models.
Covered in step 1 · The Labs: Who's Who -
How should you actually use a frontier model comparison table?
Prices and versions move weekly, so any table is a dated snapshot. It shows the shape of the field; current specs, speed, and prices come from live sources like Artificial Analysis and OpenRouter.
Covered in step 2 · Frontier Model Comparison -
You've picked an open-weight model to build a product on. What should you check first?
Some open-weight models ship under permissive Apache-2.0 or MIT licenses, but others carry custom use restrictions, so the license column matters before you build on one.
Covered in step 3 · Open-Weight Models -
A model tops SWE-bench and LMArena. What do those scores tell you?
Benchmarks tell you what a model can do on someone else's task; only an eval on your own workload tells you what it does on yours. Use leaderboards to shortlist, then test.
Covered in step 4 · Benchmarks & Leaderboards -
Two models list similar per-token prices. Why can real costs still differ sharply?
Different tokenizers can emit ~30% more tokens for the same text, and hybrid models bill hidden thinking tokens as output, so compare cost per task, not the headline per-token price.
Covered in step 5 · Pricing & Cost Reference -
Starting a new AI feature, which model choice does the handbook recommend by default?
Defaulting to a flagship is cheaper than debugging quality problems from an under-powered model, and behind a gateway you can step down a tier with a config change later.
Covered in step 6 · How to Choose a Model