How to Choose the Right AI Model in 2026: Pros, Cons and Real Examples
Claude, GPT, Gemini or open-weight? Pick a tier before a brand: current models and prices, honest pros and cons, what I actually used in five real projects (including one with no model at all), and a cost estimator.
“Which AI model is best?” sounds like a question with one answer. It isn't.
The model that writes the best legal memo is a terrible choice for tagging ten thousand support tickets an hour. It's slow, and it costs about 80 times more than a small model that would do the tagging just as well. The question that actually helps is: best at what, how fast, and for how much?
This post is how I think about that. It covers the four tiers of models, what's in each one as of November 2026, honest pros and cons, and the choices I made in things I've built, including one where the right answer was no model at all. There's a cost estimator halfway down so you can plug in your own numbers.
Pick a tier before you pick a brand
Two years ago, picking a model meant picking a company. Today Anthropic, OpenAI and Google each sell a full ladder, from tiny and cheap to huge and slow, and the top models trade places on benchmarks every few weeks. The gap between tiers is much bigger than the gap between brands in the same tier.
So decide the tier first. Then try two or three models from that tier on your own examples, and keep the one that does best on your data.
The four tiers, with pros and cons
Frontier
$4 to $8 in, $20 to $40 outThe smartest, slowest and most expensive.
Good
- Best at long, multi-step work and hard reasoning
- Fewer silly mistakes on tricky inputs
- Can work on its own for a long time
Not so good
- Slow enough that users notice
- Expensive at any real volume
- Overkill for most requests
Use it for: Research agents, big code changes, legal or financial analysis, anything where one wrong answer is costly.
Everyday
$1.50 to $4 in, $7.50 to $20 outWhere most products should start.
Good
- Strong enough for most real tasks
- Good with tools and agents
- A sensible balance of speed and cost
Not so good
- Still adds up at high volume
- Can stumble on the hardest edge cases
Use it for: Chat assistants, coding help, writing drafts, agents with a few tools, answering questions over your docs.
Small and fast
$0.10 to $0.25 in, $0.50 to $1.50 outCheap enough to call on every request.
Good
- Very fast, very cheap
- Great at jobs with a clear right answer
- Easy to run on every keystroke or message
Not so good
- Weaker at judgement and long reasoning
- Needs tighter prompts and checks
Use it for: Sorting tickets, tagging content, pulling fields out of text, routing requests to the right place.
Open-weight
No per-token bill, but you pay for the hardwareModels you download and run yourself.
Good
- Data never leaves your servers or device
- Works offline
- Full control, no surprise model changes
Not so good
- You run, scale and update it
- Big ones need serious GPUs
- Licences vary from model to model
Use it for: Private or regulated data, on-device apps, offline use, very high steady volume.
Prices are per million tokens, list price on the providers' own pages as of November 20, 2026. Gemini 4 Carbon is shown at its price from December 1.
What each vendor is good at
Within a tier the models are close, so the tie-breakers are usually practical: price, free tiers, voice and media support, and what your team already uses.
Anthropic (Claude)
Every current model has a 1M-token context window. Cached prompts are very cheap to reuse, and there's a clear ladder from Haiku to Fable. Strong at coding and long agent work.
Watch for: No image, voice or video generation. If you need those, you'll pair it with something else.
OpenAI (GPT)
A wide range from Luna at $0.10 to Astra, plus dedicated real-time voice models (GPT-Realtime and GPT-Live) for spoken conversations.
Watch for: Lots of model names and versions in use at once, so pin the exact version you test.
Google (Gemini)
Free tiers on some models, including Gemini 4 Hydrogen, which is great for prototypes. The broadest media range: speech-to-text, text-to-speech, image and video generation, and multimodal embeddings.
Watch for: Prices move. Gemini 4 Carbon doubles on December 1, 2026, when its introductory price ends.
Open-weight (Qwen, DeepSeek, Mistral, Llama)
Run them anywhere, including laptops and phones for the smaller sizes. Some come under very permissive licences like Apache 2.0 or MIT.
Watch for: The hosting, monitoring and safety work is now yours. Read the licence before you ship.
The cost difference is bigger than you think
Say your feature gets 10,000 requests a day, each with about 1,500 tokens in and 300 out. Over a month:
Same feature, same traffic, an 80× spread. If the small model passes your tests, that's about $85,000 a year saved compared with the frontier tier. Try your own numbers:
A token is roughly three quarters of an English word. A page of text is about 500 to 700 tokens.
Same traffic, same prompt: the most expensive option costs about 80× the cheapest.
List prices per million tokens as of November 20, 2026, before caching or batch discounts. Gemini 4 Carbon uses its price from December 1 ($4 in, $20 out). Prices change often, so check the provider before you budget.
Two things bring these numbers down a lot. Caching: if every request starts with the same long instructions, providers charge a fraction of the price to reuse them (on Claude, cache reads cost 2.5 to 10% of the normal input price). Batching: work that can wait a few hours, like nightly summaries, is half price on Claude's Batch API.
What I picked, and why
Theory is easy. Here are the actual choices from things I've built, including the ones I'd make differently.
1. Turning a search box into filters: small model
For smart search, a user types “2 bed apartment under $2000 that allows pets” and the model returns structured filters: bedrooms, max price, pet friendly. I used a small, cheap model (gpt-4o-mini at the time) with a strict schema. This is the textbook small-model job. There's a clear right answer, the schema catches bad output, and it runs on every search, so speed and price matter more than brilliance.
2. A voice app that works offline: open-weight, on the device
Kutti AI is a voice-first learning app for visually challenged children, and many of them may not have reliable internet. No cloud model, however good, helps if there's no connection. So speech recognition runs on the phone with Whisper through Whisper.rn. Here the choice wasn't about how smart the model was. It was about where it had to run.
3. A severe-weather assistant: the data mattered more than the model
Stormwise is built on Claude. When I evaluated it against real National Weather Service alerts (write-up on arXiv), the scary failure was false reassurance: without access to alerts, it told people conditions were calm during an active warning in 26 of 69 cases. Adding an instruction to admit it couldn't see alerts made it worse, 31 of 69. Giving it the real alerts with a plain prompt brought it to zero.
The lesson for choosing models: a bigger model would not have fixed that. If the model doesn't have the facts, it will guess, and a smarter model just guesses more convincingly. Fix the data first, then decide whether you need a bigger model.
4. A dinner recommender: the bill decided
My dinner agent was one model with three tools (location, weather, restaurant search) and a cap on how many steps it could take. It worked, and it never ran away. I still took it offline, because the API bills were more than a side project could justify. Every tool call is another model call, and they add up. If I rebuilt it, I'd use a small model for the tool calls and only bring in a bigger one to write the final suggestion. More on that in my multi-agent post.
5. A scam link checker: no model at all
The scam link checker on this site spots fake brand names, look-alike letters and risky domains. It would be easy to send each link to a model and ask “is this a scam?”. I didn't. It's plain code that runs in your browser, so it's instant, free, works offline, and never sends your messages anywhere. The rules are written down, so the same link always gets the same answer. When the job is a set of rules you can write down, write them down.
How I test before I commit
Benchmarks tell you which model is good in general. They don't tell you which one is good at your job. This takes an afternoon and saves a lot of guessing:
- Collect 30 to 50 real examples. Real user inputs, including the weird ones. Write down what a good answer looks like.
- Start with the cheapest model you think might work. Not the best one. You're looking for the floor.
- Run every candidate on the same set. Same prompt, same examples. Pin the exact model version.
- Read the failures yourself. A score hides whether the misses are harmless or dangerous. Automatic scorers can be wrong too.
- Move up a tier only for what fails. Often a small model handles most requests and you only route the hard ones up.
I wrote more about building eval sets in shipping AI features to production.
A quick way to choose
Can you write the rules down?
Plain code. No model.
Must data stay on the device or your servers?
Open-weight.
Is there a clear right answer, at high volume?
Small and fast.
Does it need judgement, tools or good writing?
Everyday tier.
Long, multi-step, and costly to get wrong?
Frontier.
Frequently Asked Questions
What is the best AI model right now?
There isn't one best model. Claude Fable 5.2, GPT-6.2 Astra and Gemini 4 Carbon are the strongest for hard reasoning in November 2026, but most products don't need them. Pick the tier that fits the job first, then test two or three models from that tier on your own examples.
Should I use a small model or a large one?
Use a small model for jobs with a clear right answer: sorting, tagging, pulling fields out of text, routing. Use a large model when the task needs judgement, several steps, or writing that has to be right the first time. Many good products use both: a small model for most requests and a large one for the hard few.
Is Claude, GPT or Gemini better?
Each vendor now sells a full range from cheap and fast to expensive and smart, and the top models are close on most tests. The differences that matter in practice are price, speed, free tiers, voice and media support, and how each one does on your own test set.
When should I use an open-weight model?
When data can't leave your servers or the device, when you need to work offline, or when you have very high, steady volume and the team to run GPUs. Check the licence for each model, because they range from fully permissive to custom terms.
How do I estimate what an AI model will cost?
Multiply requests per month by tokens per request, then by the price per million tokens, separately for input and output. Output is usually priced about five times higher than input, so long answers cost more than long prompts. Caching repeated prompts and batching non-urgent work can cut the bill a lot.
The short version
Don't start with “which model is best?”. Start with the job. Pick the cheapest tier that could plausibly do it, test a few models on your real examples, and only move up for the cases that fail. Fix missing data before you blame the model. And sometimes the best model is a few lines of plain code.
Continue Reading
Explore more articles on software engineering and technology
Model Context Protocol (MCP): How to Plug AI Agents into Your Product
What MCP is in plain terms, how hosts, clients and servers talk, how to design tools AI agents use well, and the security rules that matter, with an interactive walkthrough of real MCP messages and a hidden prompt-injection attack.
Shipping AI Features to Production: Evals, Guardrails, Latency and Cost
From impressive demo to reliable product: build an eval set, add guardrails and graceful fallbacks, keep latency and cost under control, and launch safely, with an interactive reliability simulator.
Framework-Agnostic Frontend Architecture: Migrate from Angular to React Without a Rewrite
Make the UI framework a swappable detail: MVC vs. React's 'just the view', a plain-TypeScript core, Web Components, an app shell that runs Angular and React side by side, and a safe step-by-step migration plan, with a live demo.