OpenRouter is the fastest place to test many language models without wiring a separate API key for every provider. Its live API listed 14 zero-cost :free model IDs on August 3, 2026 — down from 15 a week earlier, after Poolside's free Laguna M.1 was delisted, and down from 20 in the weeks before that as the entire free Meta Llama and Qwen tiers went. The catalog changes constantly, so the durable question is not "which free model is #1 today?"
The durable question is: which free model can safely handle a narrow step in a real workflow, and when should the work escalate to a stronger model?
For Teamday, free models are not the product. They are one cost-control layer inside an agent execution platform. AI employees can use cheap and free models for extraction, routing, classification, and drafts, then hand important decisions to stronger models and reviewable workflows.
These models are already wired in. Teamday's AI employees use free and cheap OpenRouter models for the grunt work — extraction, routing, drafts — and escalate the decisions. Try it on your own backlog: 20 work runs, 120 computer minutes, and up to $5 of AI usage free for 7 days. No card, no provider setup; connect your OpenRouter key anytime for direct rates.
Put a free model to work →Live Free Catalog: August 3, 2026
This snapshot comes from OpenRouter's public model API and includes the :free text and multimodal model IDs whose prompt and completion prices are both zero. (It excludes openrouter/free, the router, and the zero-priced Lyria audio-generation previews, which are not language models.) Check the live free-model collection before relying on a specific endpoint.
| Free model ID | Context | Practical starting point |
|---|---|---|
nvidia/nemotron-3-ultra-550b-a55b:free | 1M | Long-context reasoning and orchestration tests |
inclusionai/ling-3.0-flash:free | 262K | Fast general instruction work |
nvidia/nemotron-3-super-120b-a12b:free | 262K | General and agent tests |
poolside/laguna-s-2.1:free | 262K | Mid-size coding-agent experiments |
poolside/laguna-xs-2.1:free | 262K | Faster coding-agent experiments |
google/gemma-4-26b-a4b-it:free | 262K | General multimodal (image + video) instruction work |
google/gemma-4-31b-it:free | 262K | General multimodal (image + video) instruction work |
cohere/north-mini-code:free | 256K | Code generation and tool-use tests |
nvidia/nemotron-3-nano-30b-a3b:free | 256K | Efficient general agent tasks |
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free | 256K | Multimodal (image/audio/video) reasoning tests |
openai/gpt-oss-20b:free | 131K | General and faster reasoning experiments |
nvidia/nemotron-3.5-content-safety:free | 128K | Content-safety classification (text + image) |
nvidia/nemotron-nano-12b-v2-vl:free | 128K | Vision-language extraction tests |
nvidia/nemotron-nano-9b-v2:free | 128K | Lightweight general tasks |
What changed since the last refresh. In the week to August 3, OpenRouter delisted one more free endpoint — poolside/laguna-m.1:free — leaving the two smaller Laguna tiers (S-2.1, XS-2.1) still free. That follows the sharper churn of the prior nine days, when the entire free Meta Llama tier (llama-3.2-3b, llama-3.3-70b), the free Qwen tier (qwen3-coder, qwen3-next-80b), hermes-3-llama-3.1-405b, tencent/hy3, and dolphin-mistral-24b all went, and inclusionai/ling-3.0-flash arrived. Losing a coding-agent endpoint week over week is the whole point: never hard-wire a recurring mission to one :free endpoint.
OpenRouter also exposes openrouter/free, a router that chooses among available free models while filtering for requested capabilities such as image input, tool calling, or structured output. It is convenient for experiments, but random model selection is the wrong default when production work needs repeatable behavior.
Best OpenRouter Free Models by Job
| Job | Best free-model profile | Safe Teamday route |
|---|---|---|
| Daily chat and drafting | Ling-3.0-flash, Gemma 4, Nemotron 3 Super, or GPT-OSS | First draft for Maya |
| Coding exploration | Poolside Laguna (S-2.1 / XS-2.1) or Cohere North Mini Code | Planning input for Max, not final merge authority |
| Research summaries | Long-context free models (Nemotron 3 Ultra at 1M) | Source extraction before a frontier review |
| Data extraction | Smaller fast structured-output models | JSON rows for James |
| SEO classification | Cheap classifier with strict labels | Query/page tagging for Sarah |
| Brainstorming | Creative general models | Idea pool for Nova's marketing missions |
Free is useful for learning the shape of work. Free is dangerous when it becomes a hidden dependency for recurring business execution. OpenRouter currently documents a 50-request daily limit for free accounts, or 1,000 free-model requests per day after purchasing at least $10 in credits, with a 20-request-per-minute limit. Provider capacity can impose tighter limits.
When Free Runs Out: Cheap Paid Workhorses
The moment a mission needs more than 50–1,000 requests a day — or a :free endpoint gets delisted mid-run, as eight have in recent weeks — the honest next step is a paid model that still costs cents, not a frontier model. These are the cheapest reliable OpenRouter picks as of July 27, 2026, priced per million tokens (input / output):
| Model | $/1M in | $/1M out | Context | Good for |
|---|---|---|---|---|
inclusionai/ling-2.6-flash | $0.01 | $0.03 | 262K | High-volume extraction and routing |
mistralai/mistral-nemo | $0.02 | $0.03 | 131K | Cheap general drafting |
openai/gpt-oss-120b | $0.04 | $0.17 | 131K | Stronger open reasoning than the 20B free tier |
qwen/qwen3-30b-a3b-instruct-2507 | $0.05 | $0.19 | 262K | Long-context instruction work |
deepseek/deepseek-v3.2 | $0.27 | $0.40 | — | Near-frontier reasoning at a fraction of the price |
google/gemini-3-flash-preview | $0.50 | $3.00 | — | Fast multimodal with a reliability SLA |
For comparison, the frontier tier this list escalates to runs roughly 20–200× higher: anthropic/claude-haiku-4.5 at $1 / $5, openai/gpt-5.1 at $1.25 / $10, and anthropic/claude-sonnet-5 at $2 / $10 per million tokens — see the best AI models of 2026 for how those frontier tiers compare on quality. The routing discipline below is how you spend at the free or cheap tier for volume and reserve those frontier rates for the few steps that actually decide something.
What OpenRouter Free Models Are Good For
Extraction
Use free models to pull names, URLs, dates, product claims, bullets, and simple fields from clean source material. Keep the schema strict and validate the output.
Routing
Free models can tag an incoming task as sales, support, SEO, content, engineering, or finance so Teamday can route it to the right AI employee.
First drafts
Use free models for low-stakes drafts: outlines, title variants, FAQ candidates, social post angles, and quick summaries.
Test generation
For engineering work, free coding models are useful for suggesting tests, explaining code paths, and sketching implementation options. A real coding mission still needs a harness, files, test commands, and review — the open-source OpenCode harness pairs open models like Kimi with exactly that discipline.
What Not To Delegate To OpenRouter Free Models
Do not let free endpoints own:
- customer-facing policy decisions,
- production financial analysis,
- security-sensitive code changes,
- public claims without fact checking,
- long-running autonomous missions without fallback,
- final legal, compliance, or medical wording.
The issue is variance. Free endpoints can be rate-limited, rerouted, changed, or unavailable. Production work needs a known reliability envelope.
OpenRouter Free-Model Routing Ladder
| Step | Model tier | Output | Review gate |
|---|---|---|---|
| Classify | Free or cheap model | Task label, priority, destination AI employee | Schema validation |
| Extract | Free or cheap model | Facts, rows, links, candidate claims | Source check |
| Draft | Fast mid-tier model | Brief, memo, outline, first artifact | Stronger model review |
| Decide | Frontier model | Recommendation, risk call, final wording | Human review when public or high stakes |
| Execute | Harness plus tools | File, report, code, video, image, campaign asset | Workspace artifact and approval |
This is how Teamday turns free-model curiosity into a product story: the buyer gets lower cost without giving up review, files, tools, and accountability.
Teamday Examples
| Workflow | Free-model role | Stronger layer | Proof path |
|---|---|---|---|
| Weekly SEO report | Classify pages and extract query clusters | Sarah reviews actions and writes the plan | Weekly SEO report |
| Content refresh | Generate candidate titles and angles | Maya edits for positioning and sources | Content sample |
| Product analytics | Extract rows and anomalies | James writes the business readout | Traffic pulse |
| App build | Explain code and propose tests | Max uses a coding harness to patch and verify | Ship log |
The user who searches "best free OpenRouter models" is trying to reduce AI cost. Teamday should answer: use free models where they are reliable, then install an AI employee that knows when to escalate.
What Teamday actually costs. Starter is $99 a month for the platform — the AI employees, work runs, and computer minutes. AI usage is billed by your own provider: connect your OpenRouter (or Anthropic, OpenAI) key and pay direct rates, so the free models in this catalog stay free. The 7-day trial includes 20 work runs, 120 computer minutes, and up to $5 of AI usage on us, with no card required.
Start your free trial →Privacy And Availability Checklist
Before using a free model in recurring work, check:
- whether the provider can process your data under your requirements,
- whether rate limits support the mission cadence,
- whether the model supports structured output reliably,
- whether a fallback model exists,
- whether the output is stored as a reviewable artifact,
- whether the workflow stops cleanly when the model is unavailable.
Practical Test Prompt
Use this prompt before choosing a free model:
You are doing production business work. Read the brief and return valid JSON:
{
"summary": ["three factual bullets"],
"risks": ["top five risks"],
"missing_context": ["questions that block correctness"],
"draft_output": "the requested work"
}
If the brief lacks enough information, say so instead of inventing facts.
Run the same prompt across three free models and one paid model. Pick the cheapest model that returns useful, reviewable work without retrying.
Found your model? Don't leave it in a chat tab. Give it a role instead: Maya drafts your content, Sarah runs your SEO, Max ships your app changes — each one runs this same test across free and frontier models automatically, then leaves reviewable work in your workspace.
Hire your first AI employee →