On October 15, OpenCode made DeepSeek V4.1 Flash free, and on October 16, GLM 5.3 Flash. Two powerful and inexpensive models can now be used in the OpenCode agent without paying for tokens. But "free" here doesn't mean "unlimited": truncated context, operation only within OpenCode, and unstable speed. Below is what exactly has changed, how the two models differ, and, most importantly, in which tasks the free versions are truly worth using, and where they will let you down.
⚡ In Brief
- ✅ What:
deepseek-v4.1-flash-free and glm-5.3-flash-free are free in OpenCode "for a limited time," with no end date announced.
- 📏 Main limitations: 200K context instead of 1M, GLM also has an output up to 32K; they only work within OpenCode.
- 🧭 When to use: learning, pet projects, prototypes, small tasks within 200K. When not to use: large repositories, work with deadlines, and any use in your own application.
- 🔒 Privacy: zero-retention and prohibition of training on your data are declared for both models.
- 🚀 To try: update OpenCode, execute
/models in a session and select the free version.
About the method. This is an analytical review based on OpenCode documentation and user messages as of October 18, 2026. There are no independent tests of the free versions in the article: speed data comes from other developers' messages, and I mark them accordingly.
📚 Article Contents
🆓 What happened: DeepSeek V4.1 Flash and GLM 5.3 Flash are free in OpenCode
OpenCode is an open-source programming agent that runs in the terminal. In addition to connecting your own keys, it has Zen — a set of models that OpenCode checks and provides with a pay-per-use model.
On October 15, 2026, a free version of DeepSeek V4.1 Flash was added to this set, and on October 16, GLM 5.3 Flash. Both are available as separate free models alongside paid versions. In the OpenCode documentation, both are marked as free "for a limited time."
For DeepSeek, this is the end of a series of promotions: according to reports (ToolNavs), in September OpenCode temporarily increased the Go plan quota from $15 to $60, on October 12 to $80, and then made the model completely free.
🧩 Which models exactly and what "free" means
| Model |
ID in OpenCode |
Free from |
| DeepSeek V4.1 Flash Free | opencode/deepseek-v4.1-flash-free | October 15, 2026 |
| GLM 5.3 Flash Free | opencode/glm-5.3-flash-free | October 16, 2026 |
"Free" means zero token cost: input, output, and cached tokens cost $0. This does not mean the absence of limits: the context is truncated, speed is not guaranteed, and access only works through OpenCode (more details below).
Privacy. According to the documentation, the providers of both models adhere to a zero-retention policy and do not use your data for training. This is worth emphasizing because among other free OpenCode models, there are exceptions: for some of them (e.g., Big Pickle, MiMo, free Nemotron models), providers may store data or use it to improve models. If you switch between free models, check this for each one separately.
Which DeepSeek version exactly. On October 17, DeepSeek released the V4.1-Flash-1017 update (which I wrote about in the article about DeepSeek's Search and Fetch API). Which version Zen is currently providing — September's or 1017 — is not specified in the OpenCode documentation.
📊 GLM 5.3 Flash Free vs DeepSeek V4.1 Flash Free: comparison
| Parameter |
GLM 5.3 Flash Free |
DeepSeek V4.1 Flash Free |
| Context in free version | 200K (1M in paid) | 200K (1M in paid) |
| Maximum output | 32K (131K in paid) | Free version limit not specified |
| Input data | Text and images; PDF attachments, according to reports, do not work | Text and images (native multimodality from V4.1) |
| Speed (according to user reports) | Slow and unstable responses | Up to 100 tokens/s, closer to 200K around 50; drops during the day |
| Privacy | Zero-retention, no training | Zero-retention, no training |
| Price in Zen after promotion (input / output per 1M) | $0.15 / $0.50 | $0.30 / $1.20 |
| Cached input in Zen per 1M | $0.03 | $0.006 |
Speed data is taken from user reports (ToolNavs, Apidog), not from independent measurements. The rest is from OpenCode documentation and model pages.
Which of the two to choose
- Need to generate large files or long responses — I would choose DeepSeek, as GLM's output is limited to 32K tokens. For tasks where the response may exceed this limit, this is a significant limitation.
- Working with PDF documentation — I would consider DeepSeek or the paid version of GLM. According to available reports, GLM Free has issues processing PDF attachments, so it's worth testing this feature on your own documents before choosing.
- Planning to use the model after the promotion ends — I would not focus solely on the base token price. GLM is twice as cheap at the standard rate for input tokens, but cached input in DeepSeek is five times cheaper. If we only consider input tokens, DeepSeek becomes more profitable when over 86% of input tokens are taken from cache: 0.15 - 0.12x = 0.30 - 0.294x, where x ≈ 0.86. For long agent sessions where the same history is repeatedly passed, such a proportion is possible. At the same time, GLM is cheaper for output tokens regardless of the cache proportion.
⚠️ Free Access Limitations
Temporality
Both models are free "for a limited time" with no end date. The history here is telling: previous free GLM models in OpenCode (glm-4.7-free, glm-5-free) are already marked as deprecated. Expect the promotion to end and don't build anything on it that will break without notice.
200K Context Instead of 1M
Both paid versions support 1M tokens, free versions only 200K. For an agent, this is less than it sounds: the system prompt, tool descriptions, session history, and read file content all take up space simultaneously. On an average project, a few large files plus a long session will exhaust 200K faster than you expect, and the agent will start losing early context.
Inside OpenCode Only
Free access works only within OpenCode itself. According to Apidog, requests to the free model from other tools or custom scripts receive a 403 FreeTierError. This means it's a free agent for terminal work, not a free API for your application.
Speed and Load
Free access to a popular model is a shared resource with many users accessing it simultaneously. This explains reports of DeepSeek speed up to 100 tokens/s dropping to ~50 near the context limit, slowdowns during the day, and unstable GLM responses. Another possible consequence is errors like 429 (rate limit exceeded) during peak hours. I haven't found specific reports of 429 for these models, but for free tiers under load, this is a typical risk, so factor it into your expectations.
🧭 When to Use the Free Version and When Not To
All the limitations above boil down to three questions: will the task fit within 200K, is time critical for you, and do you need the model outside of OpenCode. Here's how it looks in specific scenarios:
| Scenario |
Free Version? |
Why |
| Learning, getting acquainted with agents, pet projects | ✅ Yes | Zero cost, and unstable speed is not critical |
| Prototype or MVP | ✅ Yes | Small code will fit into 200K; if the model is suitable, switching to the paid version is a single configuration change |
| Tests, documentation, refactoring a small module | ✅ Yes | Short tasks with limited context are the best use case for free versions |
| Comparing models before purchasing | ✅ Yes | You can evaluate quality for your tasks for free, but not speed: the paid version runs on different infrastructure |
| Generating large files | ⚠️ Caution | For GLM, the output limit is 32K; choose DeepSeek |
| Large repository, inter-module changes | ❌ No | Requires more than 200K context; the agent will lose early files and decisions |
| Working with a deadline | ❌ No | Speed is unstable, and the promotion can end without notice |
| Sensitive company code | ⚠️ Caution | Zero-retention is declared, but check if your security policy allows third-party hosting in the USA |
| Model in your application (Spring AI, backend, bot) | ❌ No | The free version only works in OpenCode; for code, you need an API (see next section) |
A practical rule of thumb: the free version is suitable when an error or delay costs you only a few minutes. If something depends on the result—money, deadlines, production—rely on the paid model.
☕ If You Need the Model in Your Own Application
For Java and Spring developers, this is perhaps the biggest disappointment of the promotion: you can't connect the free version to Spring AI because external requests receive a 403. However, the models themselves are available through regular APIs, and they cost next to nothing.
In my Spring AI-based Character AI platform, I work with the previous DeepSeek V4 Flash via OpenRouter. For V4.1 Flash, the configuration is the same, only the model name changes:
spring.ai.openai.base-url=https://openrouter.ai/api
spring.ai.openai.api-key=${OPENROUTER_API_KEY}
spring.ai.openai.chat.options.model=deepseek/deepseek-v4.1-flash
For GLM, it's enough to change the model to z-ai/glm-5.3-flash. On OpenRouter, the model is specified with the provider prefix, and it's better to use the specific name rather than an alias like "Latest": the alias will move to a new model on its own, and your application's behavior will change without your knowledge.
Two things worth knowing before switching:
- Each model on OpenRouter has about 30 providers with different prices and speeds: for DeepSeek V4.1 Flash, the price of different providers varies by about 10 times. If predictability is important to you, fix the provider in the routing settings.
- A new version is not always better for your task. DeepSeek benchmarks measure agent programming, not, for example, extracting facts into JSON or role-playing dialogue. Before replacing a model in production, run your key scenarios. I wrote about how models behave differently with tool calling in the article "Which Ollama Model to Choose for an Agent with Tool Calling".
💸 What Happens After the Promotion: How Much These Models Cost
| Where |
GLM 5.3 Flash (in / out per 1M) |
DeepSeek V4.1 Flash (in / out per 1M) |
Context |
| OpenCode Zen (paid version) | $0.15 / $0.50 | $0.30 / $1.20 | 1M |
| OpenRouter, cheapest provider | $0.048 / $0.20 | ≈ $0.05–0.07 / $0.14–0.15 | 1M |
DeepSeek API (deepseek-flash) | — | $0.30 / $1.20 peak, $0.15 / $0.60 off-peak | 1M |
Prices as of October 18, 2026. On OpenRouter, they depend on the provider, and for DeepSeek, the model page itself shows slightly different figures, so check the current ones before choosing.
An non-obvious conclusion: the same models on OpenRouter are several times cheaper than on Zen, if you take the cheapest provider. Zen, on the other hand, provides a proven OpenCode set and a single bill within the agent. If you work with OpenCode and want to stick with these models after the promotion, compare both options: OpenCode can also connect your own provider keys. The DeepSeek API has no free tier at all, only prepayment.
As a reminder, in agent work, the main part of the cost is input tokens, which are resent at each step. Therefore, caching is no less important than the base price: I provided a detailed calculation for DeepSeek in the article about Search and Fetch API.
🚀 How to start in a minute
- Update OpenCode to the latest version.
- In a session, run
/models and select DeepSeek V4.1 Flash Free or GLM-5.3-Flash Free.
- If the model is not in the list, refresh it:
opencode models opencode --refresh.
To make a free model the default for a project, add it to opencode.json:
{
"model": "opencode/deepseek-v4.1-flash-free"
}
For a one-time task without changing settings:
opencode run -m opencode/glm-5.3-flash-free "Write tests for UserService"
❓ Frequently Asked Questions (FAQ)
Can I use the free versions via API in my application?
No. Free versions only work within OpenCode; requests from other tools, according to reports, receive a 403 FreeTierError. For applications, use paid versions through OpenRouter, Zen, or DeepSeek API.
How long will the promotion last?
Unknown. OpenCode documentation only states "for a limited time." Previous free GLM models in OpenCode have already been retired, so expect temporary access.
Are my data used for training?
For these two models, according to OpenCode documentation, no: providers adhere to zero-retention. However, this does not apply to all free OpenCode models: for some of them, data may be stored or used for training, so check each one individually.
Which of the two models should I choose?
For long responses and working with PDFs, choose DeepSeek V4.1 Flash: GLM has an output limit of 32K, and PDF attachments, according to reports, do not work. If your tasks are short, try both on your code: it's free.
Do the free versions work in Claude Code, Cursor, or other agents?
No, free access is tied to OpenCode. In other tools, these models can only be used for a fee, via the provider's API or OpenRouter.
✅ Conclusions
- 🔹 DeepSeek V4.1 Flash and GLM 5.3 Flash are free in OpenCode from October 15th and 16th, "for a limited time."
- 🔹 Main limitations: 200K context instead of 1M, GLM output up to 32K, access only within OpenCode, and unstable speed.
- 🔹 Free versions are well-suited for learning, prototyping, and short tasks, but not for large repositories, deadline-driven work, and production.
- 🔹 The free version cannot be connected to your own application; the same models via OpenRouter cost from a few cents per million input tokens.
- 🔹 After the promotion, GLM is cheaper on the list, but in long agent sessions with a high cache ratio, DeepSeek may turn out to be cheaper on input.
📖 Read also
📎 Sources
Important: Promotion terms and prices are subject to change. Check the current OpenCode documentation and model pages before use. Material is current as of October 18, 2026.
