Verified October 5, 2026: OpenRouter's public API lists 17 :free model IDs with text output and zero input and output token prices. Below are their exact IDs, context limits, and input types, followed by a practical test for choosing one.
For a first experiment, try Gemma 4 for drafting or image input, Poolside Laguna or Cohere North Mini Code for coding, and a larger-context option for long documents. These are starting points based on the models' stated capabilities, not winners from an independent benchmark. A free endpoint can be unavailable even while its model card remains listed.
Want to turn a brief into content? See Maya's existing sample draft, then Install Maya and give Maya your own source material. The example shows the kind of output to review; it is not proof of a run on one of these free models.
Live Free Catalog: October 5, 2026
Source: OpenRouter's public model API. We include only IDs ending in :free where both pricing.prompt and pricing.completion equal zero and the output includes text. This excludes openrouter/free, audio-generation models, and zero-priced previews without that suffix. It is a reproducible catalog count, not a count of every way to make a zero-cost request.
Context values below are exact token counts reported by the API for the free endpoint. They can differ from a paid version of the same model. Input types describe supported modalities; every entry below produces text.
| Free model ID | Context tokens | Input types |
|---|---|---|
apodex/apodex-1.1-mini:free | 262,144 | text |
cohere/north-mini-code:free | 256,000 | text |
dots-studio/dots-3-note-preview:free | 512,000 | text, image |
google/gemma-4-26b-a4b-it:free | 262,144 | image, text, video |
google/gemma-4-31b-it:free | 262,144 | image, text, video |
inclusionai/ling-3.0-flash-sante:free | 262,144 | text |
liquid/lfm-2.5-2.6b:free | 65,536 | text |
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free | 256,000 | text, audio, image, video |
nvidia/nemotron-3-super-120b-a12b:free | 262,144 | text |
nvidia/nemotron-3-ultra-550b-a55b:free | 1,000,000 | text |
nvidia/nemotron-3.5-content-safety:free | 128,000 | text, image |
nvidia/nemotron-3.5-lightning:free | 1,000,000 | text |
poolside/laguna-s-2.1:free | 262,144 | text |
poolside/laguna-xs-2.1:free | 262,144 | text |
qwen/qwen3.8-27b:free | 262,144 | text, image, video |
thinkingmachines/inkling-small:free | 1,048,576 | text, image, audio |
thinkingmachines/inkling:free | 1,048,576 | text, image, audio |
What changed since September 17: the catalog shrank from 20 to 17 qualifying endpoints. Five free variants were withdrawn: Ling 3.0 Flash Fin and Flash VL, Nex-N2.5 Mini and Pro, and GLM-5.2. Their paid versions remain listed at non-zero prices. Two free variants were added: Qwen3.8 27B, which accepts text, image, and video input, and Apodex 1.1 Mini, which its model card describes as a reasoning model for research and forecasting tasks. Do not assume an old :free ID still works.
A concrete trap: the API reports 262,144 tokens for qwen/qwen3.8-27b:free, while the paid qwen/qwen3.8-27b endpoint lists 1,000,000. Use the free endpoint's limit when sizing your request.
Check the live free-model collection before starting work. The free router can select from eligible free models; use a specific model ID when comparing outputs so a changing selection does not invalidate your test.
Best Starting Points by Job
| Your task | Candidates to test | What to check |
|---|---|---|
| Article outline or short draft | Gemma 4, Nemotron 3 Super | Source accuracy, tone, missing facts |
| Code explanation or test suggestions | Laguna S/XS 2.1, North Mini Code, Qwen3.8 27B | Run the suggested tests; inspect changes |
| Long-document summary | Inkling, Nemotron 3 Ultra, Nemotron 3.5 Lightning | Whether important facts survive across the whole input |
| Extract information from images | Gemma 4, Qwen3.8 27B | Compare extracted fields with the image |
| Research question over your own files | Apodex 1.1 Mini | Whether each conclusion traces to the inputs you supplied |
| Small text extraction task | LFM2.5-2.6B | Valid output schema and field accuracy |
| Content-safety classification | Nemotron 3.5 Content Safety | False positives and missed cases on your examples |
Specialist models are not interchangeable with a general assistant. A finance- or health-focused name does not establish accuracy for a consequential decision. Context capacity also does not establish retrieval quality: test whether the answer actually uses the relevant passage.
Free-Model Rate Limits
OpenRouter's limits documentation lists these tiers:
| Lifetime credits purchased | Requests per minute | Requests per day |
|---|---|---|
| Below $10 | 20 | 50 |
| At least $10 | 20 | 1,000 |
The documentation allows a small threshold tolerance for rounding and top-up fees. Your actual daily ceiling and remaining allowance appear in free_model_daily_requests from GET /api/v1/key. Buying credits raises the documented allowance; it does not guarantee provider capacity.
A 429 can mean a platform limit or an upstream provider limit. Respect retry guidance and stop a recurring job cleanly when no endpoint is available. A negative credit balance can also block free requests. Check account status before assuming the model is broken.
A Practical Drafting Test
Start with a small, reviewable deliverable: an outline and introduction based only on an approved product brief. Use the same prompt and source text for each candidate:
Using only the product brief below, prepare:
1. A proposed title and five-section article outline.
2. A 150-word introduction for the named audience.
3. A list of factual claims with their supporting brief excerpts.
4. Any missing information that prevents an accurate draft.
Do not invent customer results, pricing, quotes, or product features.
Mark unsupported claims as questions for the reviewer.
Audience: [your audience]
Product brief: [paste approved source material]
Record the exact model ID, date, number of attempts, elapsed time, and reviewer corrections. If you compare a paid model, record its actual usage charge too. A free draft that needs repeated repairs can take more human time than a paid draft that meets the brief immediately.
A useful acceptance check is simple: every factual claim traces to the brief, the outline addresses the audience's question, and missing information is called out. Review the final text before publishing.
From a Model Test to Work in Teamday
Maya, the content creator, is a place to start with that same brief. View an existing content sample before deciding whether the output fits your work.
- Install Maya from the agent page.
- Complete the required AI provider setup and choose a model supported by the selected harness.
- Supply your brief, audience, and acceptance criteria in the conversation.
- Review the draft and its factual claims before it ships.
The OpenRouter catalog and Teamday's model picker are different lists. Connecting OpenRouter does not mean every :free endpoint appears in the picker. This article does not promise automatic model comparison, free-to-paid escalation, or automatic publication.
Free model tokens also do not remove Teamday's platform subscription or other service costs. See current Teamday pricing and supported harnesses before choosing your setup.
Install Maya and prepare a draft from your brief →
Frequently Asked Questions
Do I need to buy credits to test free models?
The documented lower free-model tier applies below $10 in lifetime credit purchases. Start by checking your account's actual allowance; a higher quota requires qualifying credit purchases. Model availability and account restrictions still apply.
Is this a ranking of the best free models?
No. The catalog is verified against the API, and the shortlist is based on stated capabilities. We have not run a controlled benchmark across all 17 endpoints. Use the drafting test above or your own acceptance criteria to choose.
Can I use free models for recurring work?
Only if the job can tolerate interruptions and you have tested its output. Check data-handling requirements, actual limits, failure behavior, and the cost of any fallback. A paid endpoint can also fail; paying is not a substitute for validation.
Can Teamday automatically choose a free model and escalate when needed?
That is not a feature established by this guide. Choose from the supported models in your setup and review the result. Treat the OpenRouter list as a catalog for experimentation, not a promise about Teamday routing.