You Don't Have to Pay to Start, Because Open Models Are Better Than You Think

I was in the middle of real work when the session stopped. I had hit the usage limit on Claude, with a pile of code on the screen that I had not fully understood yet. People joke that this is a sign to touch grass. In the middle of a build, it does not feel like that. It feels like being cut off. How do I finish this? What was it thinking?

What got me through that moment was an open source tool and a free model. This is about that, and about why I think many students feel they cannot compete unless they pay. I do not think that is true. Free is enough to start, and it goes further than most people expect. It does not do everything, and I will say where it stops.

My first encounter with an open model was not in a tool like this. I tried to run models locally, on my own MacBook. I installed a small Qwen model in LM Studio and an MLX version of Gemma 12B in Ollama, along with a few others. The 12B refers to the model size, in billions of parameters. Larger models generally need more memory, but size alone does not tell you how well one will handle your task.

It was manageable, but the limits were real. The models could do some tasks okay, but they were not smart enough for much more. Sometimes they were very slow, because they were competing with everything else running on the laptop. That is why free models that run on a provider's servers interested me. Then I learned about OpenCode.

OpenCode is a free, open source coding agent that runs in your terminal. It reads your files, runs commands and edits code until the job is done. The program that runs the model in that loop and gives it tools is called a harness.

I am not promoting it. It interests me because, although I do hit rate limits, the allowance is generous enough for long coding sessions and repeated changes across my codebases. OpenCode Console, which I knew as Zen, lists models its team has tested for coding agents. Some are free for a limited time; that does not promise unlimited access forever.

Many people only ever use GPT or Claude, so they never find out how capable open models have become. I wrote about why I still believe in AI in What I Believe About AI. This piece is the practical side of that belief.

The first free model I tried in OpenCode was Big Pickle, a "stealth model" whose makers have not said what it really is. On my own repository it was good at following instructions. For interface work, how things look and feel, it was lacking. It still carried on from where I left off. Free and open are different things. Open weights can be downloaded, but that alone does not make a model fully open source. Big Pickle is free to use; I cannot call its undisclosed model open. And an open model can be at the frontier too.

DeepSeek V4.1 Flash became free in OpenCode this month, and the latest version, 1017, is available there too. I use it mostly on the back end. It is also good on the front end because it studies the repository and follows the conventions already in it. That matters more than people think.

Before any of this I had GitHub Copilot Pro for free through my school email. GitHub Copilot started for me as autocomplete, and later got a chat sidebar with some rough edges. Looking back, it is fascinating how far it has come.

Picture a council. The head of the council is Claude Opus 5.5, the strongest model I have. The hands and fingers are cheaper open source models that are capable enough for the small jobs. Using the head for everything is overkill, and it is expensive.

A big part of what the hands do is tool calling. The model does not run anything itself. It writes a structured request, like "search for this" or "call this service with these values", and the harness runs it and hands the result back. That is a narrow job, and cheaper models handle it well.

I have built this into my own work with Hermes Agent, mostly through NYO, a service that gives one key to many models. NYO has published case studies on three of my projects, so I am not a neutral reviewer. What I can say is that it is quick, and the free credits go a long way. Planning is another place to save. Open models are good at thinking things through, so I plan with one and keep my paid model's limits for the moments I need them.

Launch announcements are partly marketing. Your own tests are the evidence. That is the real benefit of open models: you can try them, switch between them, and find out for yourself.

I do not want to say a free model is as good as the best paid ones. That is not true everywhere. Here is DeepSeek V4.1 Flash 1017 against two flagships, using DeepSeek's own reported results. These test setups are not measurements of the free OpenCode endpoint.

My reading is simple. It is comparable on some tests and behind on others. These results give me a reason to try it, not a promise about every job. What convinced me was seeing it work on my own project.

Some people assume a better model gives a better result every time. It does not. Quality depends on the job. In 2023, a Stanford team published FrugalGPT, which asks cheaper models first and moves up only when the answer is not good enough. In their experiments it matched GPT-4 with up to 98% lower cost. A year later a Berkeley team published RouteLLM, which decides before answering whether a question needs the strong model, and reports cutting costs by over two times in certain cases without hurting quality.

Many questions do not need the strongest model, so paying for it every time buys nothing on those. These studies used models and tasks of their time, and "up to" means the best case. They do not say open models beat the best paid ones. They say matching the model to the job is a skill, and you can only learn it if you have tried more than one.

I go back to a paid model, Claude or GPT, when the work is about:

• Interface design.
• Back end flow, how the pieces of a system connect.
• Robustness, so it keeps working when things go wrong.

That is my own benchmark, and it is the only one that counts for my work. Start free. When you hit the capability limit, you will know what you are paying for. If you only ever use one high-powered model, you never learn what the others can do.

Guardrails are for every model, not only free ones. In my resort agent, every change is previewed first and saved only after an explicit yes. I do not do that because open models are bad at following instructions. Even the most powerful models have flaws, so the rule is the same for all of them.

A subscription can give you more room to work. I save its usage allowance for the harder jobs and use free models where they are enough. Separate API calls can still cost money; a chat subscription does not generally cover another provider's API. Prices change, so check them. A few examples, as of October 2026:

• Claude Pro and ChatGPT Plus are about 20 dollars a month on monthly billing.
• Claude Max starts at 100 dollars a month. ChatGPT Pro has 100-, 200- and 500-dollar monthly tiers. Those prices are out of reach for many students here in the Philippines.
• OpenCode Go is 10 dollars a month; Go Plus is 40. Both offer access to several models with usage limits.

If your ChatGPT account has a trial offer, check its duration, payment requirements and renewal terms. Offers vary by account and campaign; a free month is not guaranteed. Cancel before a trial converts to a paid subscription if you do not want the charge. If you get one, use it hard. Run the same tasks on the paid model and on an open one, and treat the difference as your own benchmark.

When something is free, people ask what the catch is. Is it our data? It is a fair question. OpenCode's official terms differ by model. DeepSeek V4.1 Flash Free has a zero-retention policy and is not used for training. Big Pickle's collected data may be used to improve the model during its free period. That is a reason to check the particular model's terms, not assume every free offer works the same way. Keep passwords and anything private out of the files you hand over. My own rules and skills reduce mistakes, but they do not stop what a hosted model reads from reaching the provider.

Here is the irony. We hand over our emails and photos online every day without a second thought, and then we get defensive when an AI tool asks for the same. Researchers have a name for it: the privacy paradox, which Susan Barnes wrote about in 2006. We worry about privacy, and we share anyway. Be careful with your data. Just be careful everywhere, and not only here.

Our capstone project, Auxiliare, was a Laravel and TypeScript app with thousands of lines of code and a lot of interactions. The database and other services no longer exist, but I wanted people to still see what we built. This was a whole application to convert, not a single page.

I gave DeepSeek V4.1 Flash in OpenCode one prompt to turn it into a TypeScript demo. It kept the front end and its interactions, without a live back end. Every account and number in it is made up. It is a tour, not the product.

That left me in awe. How can a model available for free do this much from one prompt? I did not need to use my paid models for that conversion. This one was capable enough for the job, and seeing it handle a project that big made the point more clearly than a benchmark could.

I do not think this is where it stops. I expect more open models to arrive, and some of them may be better than today's flagships. I cannot prove that. It is my own expectation, and it rests on three things I can see: there is real demand for them, companies keep investing in building them, and the data and know-how behind them keep growing.

Some of that is already measurable. Mozilla's September 2026 report, The State of Open Source AI, says eight of the top ten models by token volume on OpenRouter in August were open-weight. It describes a roughly four-month gap on the capability measures it discusses, not a guarantee for every task. If that is true now, there is more to it than we realize yet. A student who learns to use open models today will be ready for whatever comes next.

Some students think they cannot compete because they cannot pay for the best tools. I understand that, but I do not believe it.

It is never about competition. The only competition there is, is of yourself.

— Marc

Pay to compete? Pay to win? There is no such thing. Even with the most advanced model in the world, if your mind is set against yourself, you will not get far.

So here is the one thing I would ask you to do. Stop reading. Take ten minutes and ask yourself:

1. What have I been doing, and why?
2. What in my own thinking needs to change?
3. How can I use these tools to bring an idea to life and help someone else?

The feeling of contributing to something good for others is a win. No competition, however big, can compensate for it.