A Free Claude Code Alternative: OpenCode With NVIDIA's Nemotron

We opened a free app called OpenCode, picked NVIDIA's newest model from its free list and asked it for a landing page for our agency. It read our website, wrote the page, then wrote an upgraded version when we asked. From the first message to the finished page took 8 minutes and 12 seconds. The cost panel in the corner said $0.00 spent.

The model was Nemotron 3.5 Super VL, which NVIDIA released on 15 October 2026. OpenCode added it to its free list the same week.

If you want a free alternative to Claude Code, this is the place to start. This guide covers what OpenCode is, what the free model can and cannot do, every step from download to a finished page, and the limits to know before you put real work into it.

What OpenCode is

OpenCode calls itself "the open source AI coding agent." You point it at a folder on your computer and tell it what you want. It reads the files, writes new ones and runs commands to get the job done, the same way Claude Code or OpenAI's Codex do.

• It is open source under the MIT licence, with about 222,000 GitHub stars and 950 contributors. The repo is anomalyco/opencode.
• There is a desktop app for macOS, Windows and Linux, a terminal version and extensions for code editors such as VS Code, Cursor and Zed.
• It connects to more than 75 model providers, including Claude, GPT and Gemini, and to models running on your own machine.
• It has two built-in agents you switch between with the Tab key: a build agent that can change files, and a read-only plan agent that only looks and suggests.
• It comes with a list of free models. OpenCode's docs say those free models only work inside OpenCode, not from other apps.

The app is free. You only pay if you connect a paid model.

What Nemotron 3.5 Super VL is

NVIDIA is best known for the chips AI runs on, and it also builds its own open models. From NVIDIA's model card on Hugging Face:

• 120 billion parameters in total, with only 12 billion active for each word it writes. It is a mixture of experts: each layer has 512 small expert networks, and it switches on only the few it needs. That is why a big model can still answer quickly.
• Up to 1 million tokens of context, so it can hold a lot of your project in view at once.
• It reads text, images, documents and video (MP4 up to two minutes) and writes text.
• It is released under the OpenMDW-1.1 licence, and NVIDIA says it is ready for commercial use.
• NVIDIA reports 51.78 on SWE-bench Pro and 71.10 on SWE-bench Verified, two tests of fixing real software issues.

NVIDIA's launch post calls it the fastest open agentic model in its class, with "up to 5x" faster throughput than others in that class on Artificial Analysis's measurement. NVIDIA names Atlassian, CrowdStrike, Perplexity and Salesforce among the companies building it into their software. All the scores above are NVIDIA's own tests.

Before you start

• A computer it supports. The download page has the desktop app for macOS (Apple Silicon and Intel), Windows (x64 and ARM64) and Linux (.deb and .rpm).
• A folder to work in. Make a new, empty one for your first project, for example landing-page-test. OpenCode can change any file in the folder you open.
• No paid plan. The free models cost nothing. You may still be asked to sign in, which is covered below.
• A data rule. Decide now what never goes into the free model: client files, customer lists, contracts, passwords.
• About 10 minutes for a first page. Our test took just over 8 minutes, and the model itself was working for about 2 of them.

Stage 1: Install OpenCode

The desktop app (easiest)

1. Go to opencode.ai/download and find the OpenCode Desktop section.
2. Click the download for your system. On a Mac, check whether you have Apple Silicon or Intel (Apple menu, then About This Mac) and pick the matching file.
3. Run the installer and open OpenCode.

The terminal version (optional)

If you already live in a terminal, the same page lists install commands. On 27 October 2026 the macOS and Linux command was:

Copycurl -fsSL https://opencode.ai/v2/install | bash

On Windows, OpenCode's docs say "For the best experience on Windows, we recommend using Windows Subsystem for Linux (WSL)." Scoop, Chocolatey and npm installs also exist. If you are not sure, use the desktop app; it does the same job.

Stage 2: Open a folder and start a session

1. In the desktop app, start a new session and pick the folder you made. In the terminal, move into the folder with cd and type opencode.
2. Optional for a real project: in the terminal version, type /init. OpenCode reads the project and writes an AGENTS.md file with its structure and patterns, so later sessions start with that context.

Why the folder matters: the agent can read, create and change files inside it. Keep it to one project per folder, and never open a folder that holds private client data while you are on a free, logged model.

Stage 3: Pick the free Nemotron model

1. Click the model name in the app. In the terminal, type /models.
2. Search "Nemotron."
3. Pick Nemotron 3.5 Super VL, the one marked Free. NVIDIA's two other free models, Nemotron 3 Ultra and Nemotron 3.5 Lightning, are on the same list.
4. If you do not see it, the free list has changed. Update the app, check the list again, and pick another free model if it has gone.

The free list moves fast. On 16 October 2026, OpenCode listed 15 free models. On 27 October its docs listed ten: Big Pickle, GLM-5.3-Flash, DeepSeek V4.1 Flash, Beam, Flint 40B, MiMo-V2.6-Flash, Nemotron 3 Ultra, Nemotron 3.5 Lightning, Nemotron 3.5 Super VL and Jev 1.13.

Stage 4: Plan first, then build

1. Press Tab to switch to the plan agent. An indicator in the lower right corner shows which mode you are in.
2. Paste your request. In plan mode the agent cannot change files. It reads, thinks and tells you how it would do the job.
3. Read the plan. Correct anything wrong now, while it costs nothing to change.
4. Press Tab again to switch back to the build agent and tell it to go ahead.

This step is worth the extra minute on a free model. Fixing a plan is quicker than fixing a finished page, and the plan agent cannot touch your files while you read it.

A first request that gives the model enough to work with (this is our starter prompt):

CopyBuild a one-page landing page for my business in a single HTML file.
Business: [what you sell, who it is for, your city or online]
Include: a headline and subheadline, three benefits, three customer reviews,
a short FAQ and a contact form. Use a clean, modern look.
When you are done, tell me the file name so I can open it in my browser.

If you already have a website, give the agent the address instead of describing the business. Our test request named our own site, and the page it wrote pulled in our services and case studies without us typing them out.

Stage 5: Check the result and ask for changes

1. Open the file it made in your browser. Double-click it in the folder, or ask the agent where it is.
2. Look at it on a laptop screen and on your phone width (shrink the browser window).
3. Ask for specific changes: "make the headline shorter," "use our brand colors #123456 and #abcdef," "add a pricing section." Vague requests like "make it better" get vague results.
4. If a change goes wrong, type /undo in the terminal version. It reverts the latest changes and puts your message back so you can rewrite it. /redo brings the change back.

Set your expectations honestly. Our test page had a hero section, services, case-study cards and a contact form. It worked, but it looked generic, the kind of page people call AI slop. For a free first draft that is a fair trade. You do not get the polish of a top paid model without more rounds of feedback.

The steps in one block

Keep this somewhere handy, or send it to a teammate:

CopyFree AI coding app: https://opencode.ai

Step by step:
1. Download the desktop app from opencode.ai/download (or install it in your terminal, the command is on the same page)
2. Open it, start a new session and pick the folder you want to work in
3. Click the model name, search "Nemotron" and pick Nemotron 3.5 Super VL, the one marked Free
4. Type what you want it to build

Good to know:
- The free Nemotron model is only free for a limited time.
- OpenCode uses the free period to collect feedback and improve the model, so don't paste personal or client data.
- The benchmark scores are NVIDIA's own tests.

Prefer a one-click copy? The same steps are on our free prompt page.

What to know before you rely on it

Free is temporary

OpenCode's docs say "Nemotron 3.5 Super VL Free is available on OpenCode for a limited time." No end date is given, and no paid price for it is listed yet. Plan for it to change.

Your prompts are logged

The same page says trial use only, "do not submit personal or confidential data," and that use is logged for security purposes and to improve NVIDIA's products and services. It adds that the logged data is not linked to your identity. Use it for your own marketing pages and experiments, never for client files or customer lists.

You may be asked to sign in

OpenCode's FAQ says the free models work without an account. Its Zen docs describe signing in, adding billing details and credits, and copying an API key for its hosted paid models, which you connect with /connect. If the app asks you to sign in, that is why. You only need billing details for paid models.

It is not the strongest free coder on the list

NVIDIA's own table on the model card compares Nemotron with other models. On SWE-bench Pro, Nemotron scores 51.78, DeepSeek V4 Flash 56.91 and GLM-5.3-Flash 59.70. GLM-5.3-Flash is on OpenCode's free list; the list carries the newer DeepSeek V4.1 Flash rather than the version NVIDIA tested.

The GLM, DeepSeek and MiMo models on the list come from Chinese labs. If your company keeps those out of business work, the three NVIDIA models are the obvious free picks.

Troubleshooting

• Nemotron is not in the model list. The free list rotates. Update OpenCode, then pick another model marked Free.
• The app asks for a sign-in or an API key. That is for OpenCode's hosted paid models. Billing details are only needed for paid models, so stay on a model marked Free.
• The page looks generic. Give it more to work from: your website address, the address of a page you like, your brand colors as hex codes and your real wording.
• It changed a file you did not want changed. Use /undo in the terminal version, and start the next request in plan mode.
• The terminal version misbehaves on Windows. Follow OpenCode's advice and run it in WSL, or use the desktop app.

OpenCode vs Claude Code

You can also use Claude inside OpenCode. Its hosted price list has Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, paid per use instead of a monthly plan. That suits light, occasional use better than a subscription.

Another free route is Codex, OpenAI's coding agent, which is included with limits in the free ChatGPT plan according to OpenAI's Codex pricing page. It is worth a try if you already use ChatGPT.

Our view: OpenCode with a free model is a sensible way to find out whether an AI coding agent fits how you work, without paying anything. If you start depending on it for client work, move to a paid model you trust and read its data terms. See Claude's plans if Claude Code is the next step.

Who this is for, and who should skip it

• It suits a founder or marketer who wants a quick landing page, a small internal tool or a prototype to show a developer.
• It suits anyone curious about Claude Code who does not want to pay for a plan yet.
• Skip the free model for anything with customer data, contracts or private numbers, because trial use is logged.
• Skip it if you need a polished result in one go. Expect to ask for changes.

If you want an AI workflow like this set up properly for your team, with the right model and data rules, our AI automation team can help.

Frequently asked questions

Is there a free alternative to Claude Code?

Yes. OpenCode is a free, open-source AI coding agent under the MIT licence that works like Claude Code, in a desktop app or your terminal. It comes with a list of free models, including NVIDIA's Nemotron 3.5 Super VL as of 27 October 2026, and can also connect to Claude, GPT and Gemini.

Is OpenCode really free?

The app is free and open source. Its free models cost nothing to use, but they rotate, and OpenCode says Nemotron 3.5 Super VL Free is only available for a limited time. You pay only if you connect a paid model.

Is my data private when I use the free Nemotron model?

Treat it as not private. OpenCode's docs say the free version is trial use only, that use is logged for security and to improve NVIDIA's products and services, and that you should not submit personal or confidential data.

Do I need an account to use OpenCode?

OpenCode's FAQ says its free models work without an account. Its Zen docs describe signing in, adding billing details and copying an API key for hosted paid models, so you may be asked to sign in for those.

How good is Nemotron 3.5 Super VL at coding?

NVIDIA reports 51.78 on SWE-bench Pro and 71.10 on SWE-bench Verified. In the same table on its model card, GLM-5.3-Flash scores 59.70 and DeepSeek V4 Flash 56.91 on SWE-bench Pro. GLM-5.3-Flash is also free in OpenCode, so by NVIDIA's own numbers Nemotron is not the top coder on the free list.

How much does Claude Code cost?

Claude Code is included in paid Claude plans. Pro costs $17 a month billed yearly or $20 billed monthly, and Max starts at $100 a month. It is not on the free plan.

Sources

• OpenCode homepage and FAQ
• OpenCode docs: install, plan and build modes, undo
• OpenCode Zen docs: free models and privacy notes
• OpenCode download page
• OpenCode on GitHub (anomalyco/opencode)
• NVIDIA Nemotron 3.5 Super VL model card (Hugging Face)
• NVIDIA Blog: Nemotron 3.5 Super VL launch (15 Oct 2026)
• Claude plans and pricing