Free model lists on OpenRouter go stale fast — a route that worked
last month can quietly disappear or lose a capability. This guide focuses on
the best free OpenRouter models for programming as verified
against OpenRouter's live Models API in August, 2026, organized by what
you're actually trying to do: generate code, run a coding agent, debug with
deep reasoning, work across a huge context window, or return structured
JSON. You'll get exact provider/model:free IDs, current context
and capability data, account-wide rate limits, fallback pairings, and the
difference between a fixed :free route and the dynamic
openrouter/free router — so you can pick a model based on your
task, not on hype or a single benchmark.
Quick Answer: Best Free OpenRouter Models for Programming
There isn't one universal "best" free OpenRouter model for coding — the right pick depends on the task:
-
Dedicated code generation:
cohere/north-mini-code:free -
Coding agents / terminal tools:
poolside/laguna-s-2.1:free -
Lightweight agentic coding:
poolside/laguna-xs-2.1:free -
Long-context programming prompts:
nvidia/nemotron-3-ultra-550b-a55b:free -
Reasoning-heavy debugging + structured output:
nvidia/nemotron-3-super-120b-a12b:free -
Structured JSON / multimodal dev work:
google/gemma-4-26b-a4b-it:free -
General coding fallback:
openai/gpt-oss-20b:free
What Counts as a Free OpenRouter Model?
OpenRouter provides one OpenAI-compatible API in front of models
from many developers and hosting providers. A model ID is the exact route
you send in a request — for example
cohere/north-mini-code:free. The
:free variant
marks a zero-token-price variant of a specific model, and that free variant
can have different context limits, rate limits, and provider availability
than the paid version of the same model. openrouter/free is a
different thing entirely: it's a router, not a foundation model. It
dynamically selects from whatever free models are currently available and
can filter candidates by requirements like tool calling or structured output
support.
Because free routes change quickly, this article treats
GET https://openrouter.ai/api/v1/models as the publication
gate. A :free landing page by itself isn't proof — if a route
doesn't appear with zero prompt and completion pricing in the current Models
API response, it isn't treated as verified here. That gives four practical
categories:
-
Verified specific
:freeroute — the exact ID appears in the current Models API with zero pricing. -
Router-only free access — reachable
through
openrouter/free, but no fixed route is being recommended. -
Conflicting official status — a
:freelanding page exists, but the route wasn't confirmed as a static entry in current model metadata. Not placed in the verified shortlist. - Open-weight but not free — weights are public, but the OpenRouter route currently charges per token. Open weights do not equal free API access.
Best Free OpenRouter Models for Programming Compared
| Model | Exact ID | Best for | Context | Reasoning | Tools | Structured output | Main limitation |
|---|---|---|---|---|---|---|---|
| Cohere North Mini Code | cohere/north-mini-code:free |
Code generation, terminal tasks | 256K | Yes | Yes | Not listed on the free route | Tool support isn't the same as OpenRouter structured outputs |
| Poolside Laguna S 2.1 | poolside/laguna-s-2.1:free |
Coding agents, repo work | 262K | Yes | Yes | Not listed | Free-route availability less predictable than paid |
| Poolside Laguna XS 2.1 | poolside/laguna-xs-2.1:free |
Lightweight agentic coding fallback | 262K | Yes | Yes | Not listed | Smaller model; don't expect Laguna S 2.1 results |
| NVIDIA Nemotron 3 Ultra | nvidia/nemotron-3-ultra-550b-a55b:free |
Long-context prompts, orchestration | 1,000,000 | Yes (configurable effort) | Yes | Not listed on the free route | 1M context isn't a coding-quality score |
| NVIDIA Nemotron 3 Super | nvidia/nemotron-3-super-120b-a12b:free |
Reasoning-heavy debugging, JSON | 262K | Yes | Yes | Supported | Not a coding-specialist model |
| Google Gemma 4 26B A4B | google/gemma-4-26b-a4b-it:free |
Structured JSON, multimodal dev analysis | 262K (provider endpoint may be lower) | Yes | Yes | Supported | Endpoint context can trail model-level metadata |
| OpenAI gpt-oss-20b | openai/gpt-oss-20b:free |
General coding, structured API responses | 131K | Yes (configurable effort) | Yes | Supported | General-purpose, not a coding specialist |
How We Selected and Verified Models
We prioritized OpenRouter's current Models API, official model pages,
capability metadata, and official developer documentation. Because free
routes and providers can change quickly, availability was checked on August
10, 2026. A :free landing page alone was not treated as proof
when the route could not be confirmed in current model metadata. This is a
verification pass, not a hands-on benchmark study — we did not run
standardized tests across these models, and nothing below should be read as
first-party test results.
New to OpenRouter? Start with our OpenRouter API guide for beginners to create an API key, send your first request, and understand free routes.
Best Free OpenRouter Models by Programming Task
| Task | Primary | Backup | Why |
|---|---|---|---|
| Dedicated code generation | North Mini Code | Laguna XS 2.1 | Both are coding/agent oriented |
| Coding agents / terminal tools | Laguna S 2.1 | North Mini Code | Tool support plus software-engineering positioning |
| Lightweight agentic coding | Laguna XS 2.1 | North Mini Code | Smaller, faster coding-agent route |
| Long-context coding prompts | Nemotron 3 Ultra | Gemma 4 26B A4B | Ultra exposes 1M tokens on its free route |
| Structured JSON / schema | Nemotron 3 Super | Gemma 4 26B A4B | Both list structured outputs on the free route |
| Multimodal developer analysis | Gemma 4 26B A4B | — | Image/video input plus tools and structured output |
| General coding + API automation | gpt-oss-20b | Nemotron 3 Super | Tools, reasoning, response format, structured output |
| Dynamic experimentation | openrouter/free |
Specific verified route | Simplest free access when model identity doesn't matter |
Best Free OpenRouter Models for Programming
1. Cohere North Mini Code
cohere/north-mini-code:free
Current free status: Verified — zero prompt/completion pricing on the current Models API. Best for: dedicated code generation, terminal tasks, and everyday agentic programming. Cohere's first agentic coding model was built specifically for this kind of work, which is why it's a solid default when you just want a model that treats code as its primary job rather than a side skill.
structured_outputs parameter. Don't
assume schema-constrained JSON works here just because tool calling does.
Fallback: Laguna XS 2.1
2. Poolside Laguna S 2.1
poolside/laguna-s-2.1:free
Current free status: Verified. Best for: coding agents, repository-level work, and tool-based software-engineering tasks. Laguna S 2.1 is positioned specifically for agentic coding workflows, which makes it a natural pick when a task involves multiple tool calls in sequence rather than a single code snippet.
Fallback: North Mini Code
3. Poolside Laguna XS 2.1
poolside/laguna-xs-2.1:free
Current free status: Verified. Best for: lighter agentic coding and tool-based development, or as a fallback when Laguna S 2.1 is unavailable.
Fallback: North Mini Code
4. NVIDIA Nemotron 3 Ultra
nvidia/nemotron-3-ultra-550b-a55b:free
Current free status: Verified — the free route exposes the full 1M-token context window. Best for: long-context programming prompts, large log files, planning across many documents, and multi-step agent orchestration where you need to keep a huge amount of project context in view at once.
Fallback: Nemotron 3 Super
5. NVIDIA Nemotron 3 Super
nvidia/nemotron-3-super-120b-a12b:free
Current free status: Verified, including
structured_outputs and response_format support
on the free route. Best for: reasoning-heavy debugging,
structured automation, and agent workflows that need to return validated
JSON rather than free-form text.
If you want to use these models inside automated workflows, see our guide to using OpenRouter with n8n .
Fallback: Gemma 4 26B A4B
6. Google Gemma 4 26B A4B
google/gemma-4-26b-a4b-it:free
Current free status: Verified, with tool calling and structured output support confirmed on the free route. Best for: structured JSON workflows, function calling, and multimodal developer tasks — screenshot or UI analysis, API automation that needs schema-constrained responses, and similar work where image or video input matters.
Fallback: gpt-oss-20b
7. OpenAI gpt-oss-20b
openai/gpt-oss-20b:free
Current free status: Verified, with tools, structured outputs, and response format all confirmed on the free route. Best for: general programming help, structured API responses, and debugging support when you want a general-purpose fallback with broad capability coverage.
Fallback: Nemotron 3 Super
qwen/qwen3-coder:free, openai/gpt-oss-120b:free,
and poolside/laguna-m.1:free all currently have free-facing
landing pages, but at verification time we could not confirm the exact
:free route as a static entry in the live Models API response.
We don't treat a route as verified until it appears there. If you use one of
these three, check the Models API yourself immediately beforehand rather
than relying on the landing page alone.
How to Choose the Right OpenRouter Model for Coding
Work through this in order rather than picking by name recognition:
- What's the programming task — generation, debugging, agentic work, or automation?
- Do you need a coding specialist, or is general reasoning enough?
- Does the task need deep, multi-step reasoning?
- Does the prompt or codebase require long context?
- Does the workflow call tools — terminal, file edits, search?
- Does the output need to be schema-constrained JSON?
-
Choose a verified specific
:freeroute that matches, then add a fallback in case it's rate-limited or unavailable.
A large context window is a capacity number, not a coding-quality score — a model with a 1M-token window isn't automatically better at writing correct code than one with 131K.
Should You Use openrouter/free for Programming?
| Factor | Specific model:free |
openrouter/free |
|---|---|---|
| Model identity | Fixed | Dynamic |
| Reproducibility | Higher | Lower |
| Best for | Repeatable coding workflows | Experiments, low-stakes prototypes |
| Main risk | Route can disappear | Model behavior can change request to request |
openrouter/free
helps when you want free access without picking a model — quick prototypes,
experimentation, or cases where you genuinely don't care which free model
answers. Avoid it for reproducible debugging sessions, fixed benchmark
comparisons, stable coding agents, or any workflow where consistent output
style matters. A specific verified route is the better choice whenever model
identity or tool behavior needs to stay predictable.
Context, Tool Calling, and Structured Output
Context determines how much of a file, log, or repository excerpt fits in one request. It's genuinely useful for large codebases, but a bigger number isn't a quality score — check it against what your task actually needs.
Tool calling lets a model return function/tool calls for actions like terminal execution, file edits, repository search, or MCP-style workflows. It's a prerequisite for most coding agents.
Structured output is a separate capability: OpenRouter's
structured_outputs parameter constrains a response to a JSON
schema so it's machine-readable and validated. A route can support tools
without supporting structured outputs, and vice versa — always check the
exact free route's supported_parameters rather than assuming it
inherits capabilities from the paid or base version of the same model.
Free OpenRouter Rate Limits and Availability
Current figures from OpenRouter's official limits documentation, checked August 10, 2026:
- 20 free-model requests per minute
- 50 free-model requests per day if your account has purchased less than $10 in lifetime credits
- 1,000 free-model requests per day once you've purchased at least $10 in credits
These are account-wide free-model limits, not a separate quota per model —
calling three different :free models still draws from the same
pool. Exact limits may vary by account, region, usage, and current product
changes, so treat these as the documented baseline rather than a guarantee.
A 429 can come from OpenRouter when you hit a platform limit, or from an
upstream provider that is rate-limiting or at capacity. Use exponential
backoff and keep a fallback model ready, because free capacity can change or
become temporarily unavailable.
Before You Send Proprietary Code to a Free Route
"Free" describes price, not privacy. Before sending proprietary source code, credentials, API keys, customer data, or private repositories to a free route, check the current provider data policy shown by OpenRouter — it can differ by provider:
- Poolside's free Laguna routes (S 2.1, XS 2.1, and the watchlisted M.1) state that inputs and outputs may be used to train and improve Poolside's models.
- NVIDIA's free Nemotron endpoints warn against uploading confidential or personal information and state that usage is logged for security and product-improvement purposes.
Don't generalize either policy to every provider on OpenRouter — data-use terms are set per provider, and the safest approach is to check the current listing for the exact route you're about to call.
When Free Models Are Not Enough
Free routes are best framed as tools for experimentation, learning, prototypes, and low-volume development — not guaranteed production infrastructure. Consider a paid route once any of these apply: your application needs high or predictable request volume beyond the daily free cap, you need consistent latency and a stable model version rather than one that can rotate out of the free catalog, you're running a business-critical coding agent that can't tolerate a route disappearing, or you're working with private or proprietary data where a provider's free-tier data-use terms are a dealbreaker. This isn't a case for abandoning free models entirely — it's a case for knowing when the zero-cost tier stops matching the job.
Frequently Asked Questions
What is the best free OpenRouter model for programming?
There's no single universal winner — it depends on the task. For dedicated
code generation, cohere/north-mini-code:free is a strong
starting point; for coding agents,
poolside/laguna-s-2.1:free; for long-context work,
nvidia/nemotron-3-ultra-550b-a55b:free. See the task table
above for the full breakdown.
What does :free mean on OpenRouter?
The :free suffix marks a specific zero-token-price variant of
a model — for example openai/gpt-oss-20b:free. It has $0
prompt and completion pricing, but it can carry different context limits,
rate limits, and provider availability than the paid version of the same
model.
What is openrouter/free?
openrouter/free is a dynamic router, not a specific model. It
automatically selects from currently available free models and can filter
candidates by requirements like tool calling or structured output, which
makes it convenient for experimentation but less predictable than a fixed
route.
Should I use openrouter/free or a specific free model?
Use a specific verified :free route when model identity, tool
behavior, or output style needs to stay consistent — coding agents and
reproducible debugging fall here. Use openrouter/free for
quick experiments or prototypes where you don't care which free model
responds.
How many free OpenRouter requests can I make?
20 requests per minute on any account, plus 50 requests per day if you haven't purchased at least $10 in lifetime credits, or 1,000 requests per day once you have. These are account-wide limits shared across all free models, not a separate quota per model.
Can a free OpenRouter model disappear?
Yes. Free routes can lose capacity, change providers, or drop out of the free catalog without much notice. That's why this guide treats the live Models API as the source of truth and recommends keeping a fallback model configured.
Which free OpenRouter models support tool calling?
All seven models in the core shortlist — North Mini Code, Laguna S 2.1, Laguna XS 2.1, Nemotron 3 Ultra, Nemotron 3 Super, Gemma 4 26B A4B, and gpt-oss-20b — support tool calling on their free routes.
Which free routes support structured output?
Nemotron 3 Super, Gemma 4 26B A4B, and gpt-oss-20b list OpenRouter's structured-outputs parameter on their free routes. North Mini Code, Laguna S 2.1, Laguna XS 2.1, and Nemotron 3 Ultra support tools but do not list structured outputs on the free route.
Are free OpenRouter models suitable for production?
They're best suited to experimentation, prototyping, and low-volume development. Free routes carry account-wide rate limits, and upstream free capacity can change or become temporarily unavailable. That makes them a weaker fit for business-critical or high-volume production workloads that need predictable availability and throughput.
Is it safe to send proprietary code to a free OpenRouter model?
Check the specific provider's data-use policy first. Poolside's free Laguna routes may use inputs and outputs to improve their models, and NVIDIA's free Nemotron endpoints log usage and warn against uploading confidential information. "Free" describes price, not privacy — the policy varies by provider.
Final Recommendations
- Dedicated coding: North Mini Code
- Coding agents: Laguna S 2.1
- Lightweight agent fallback: Laguna XS 2.1
- Long-context reasoning: Nemotron 3 Ultra
- Reasoning + structured output: Nemotron 3 Super
- Structured / multimodal workflow: Gemma 4 26B A4B
- General coding fallback: gpt-oss-20b
-
Automatic dynamic free routing:
openrouter/free
Free OpenRouter routes move fast — models get added, renamed, or
dropped from the free catalog with little warning. Before wiring any ID from
this article into a long-lived application, recheck it against
https://openrouter.ai/api/v1/models.