Best Free OpenRouter Models for Coding: November 2026

Checked October 28, 2026: OpenRouter lists 16 free (:free) models. 15 accept tool calls, and 4 were built for coding by their makers: Recursive Flint, Cohere North Mini Code, and Poolside Laguna S 2.1 and Laguna XS 2.1. Both Laguna models go away on October 31, 2026, the free and the paid versions. That leaves two coding specialists for November.

Start with Recursive Flint or Cohere North Mini Code. Both accept tool calls, and neither free endpoint uses your prompts for training. Flint's free endpoint allows up to 235,929 output tokens. Neither model has a paid version on OpenRouter, so when its single provider is full, there's no paid copy of the same model to switch to.

We picked these on stated purpose and endpoint data, not on a benchmark of our own. For the full free catalog, including non-coding models, see our OpenRouter free models guide.

Every free model that accepts tool calls

As of 2026-10-28 from OpenRouter's models API, each model's /endpoints API, and the data policy on each model's OpenRouter page.

How we computed it: we kept IDs ending in :free with zero input and output prices, text output, and tools in supported_parameters. Context and max output are the free endpoint's own limits, which can be lower than the paid version's. Uptime is OpenRouter's uptime_last_1d for the free endpoint. "Trains on prompts" is the free endpoint's data_policy on OpenRouter's model page. "Built for" summarizes each maker's model card. Paid prices are the paid listing in the models API. Each free endpoint is served by exactly one provider.

Three things the table shows:

• Free and paid can differ. Laguna S 2.1's model card describes a 1,048,576-token context. Its free endpoint serves 262,144 tokens with 4-bit (fp4) weights, while Flint's free endpoint runs at full bf16 precision. Benchmark scores from a model card may not hold on a quantized free endpoint.
• 10 of 16 free endpoints may train on your prompts. That covers both Laguna models, both Inkling models and every free NVIDIA endpoint. For Inkling and NVIDIA, the policy also lets OpenRouter train on them. The paid versions we checked, such as Inkling on Together and Laguna on Poolside, don't train on prompts.
• One endpoint was effectively down on October 28. Nemotron 3 Nano Omni reported 0% uptime for the last 30 minutes and 8.3% for the last day.

What the makers claim

Each maker tested its model with its own harness, settings, and benchmark version. These scores are not comparable with one another, and we didn't reproduce any of them.

The model cards also give sizes. Flint is about 2.5B active out of 39.8B parameters, North Mini Code 3B of 30B, Laguna XS 2.1 3B of 33B, Laguna S 2.1 8B of 118B, and Inkling 41B of 975B. Flint, North Mini Code and both Laguna models have open weights under Apache 2.0 or OpenMDW-1.1, so you can run them yourself if the free endpoint disappears.

Which free model to try for which coding job

Skip LFM2.5-2.6B for coding: Liquid's model card says it is "not recommended for agentic coding." Skip Nemotron 3.5 Content Safety too: it's a moderation model and doesn't take tool calls.

Free limits are tight for coding agents

OpenRouter's limits documentation sets these limits for :free models:

A coding agent sends one request per turn: read a file, run a test, apply an edit. On the lower tier, one debugging session can use up all 50 requests. Buying $10 of credits raises the daily ceiling, but it doesn't add providers. Every free endpoint above has a single provider, so a 429 from that provider has no same-model fallback. OpenRouter lets you pass a list of fallback models. If you list a paid model there, check its price first.

Laguna is a concrete trap. After October 31, any agent config that still calls poolside/laguna-s-2.1:free or poolside/laguna-xs-2.1:free will fail, and the paid Laguna IDs go away the same day. Check your IDs against the live list before each run:

curl -s https://openrouter.ai/api/v1/models | jq -r '.data[]
  | select((.id | endswith(":free")) and (.supported_parameters | index("tools")))
  | [.id, .context_length, (.expiration_date // "-")] | @tsv'

The third column shows the date OpenRouter has set for removing a model, if it has one.

A 20-minute test before you commit

Give each candidate the same small, checkable task in your coding agent: a failing test in your own repository, with the instruction to fix the code, not the test. Record the model ID, the number of requests used, whether the test passes, and what you had to correct. Run it twice, because the same model can give different answers. A free model that needs three rounds of fixes can cost more of your time than a cheap paid model costs in tokens.

For paid coding picks and their prices, see the most-used models on OpenRouter. For how to set up a coding agent with tests and review, see our guide to agentic coding.

From a test to shipped work in Teamday

Max, Teamday's app builder, builds landing pages, internal dashboards, automations and small web apps from a brief. Teamday's model picker lists four OpenRouter models (Kimi K3, GLM 5.3, GLM 5.3 Flash, and DeepSeek V4 Pro), and their usage is billed to your own OpenRouter account. The free endpoints in this guide are not in the picker. See Max's ship log for examples of the work, and current pricing.

Give Max a real task →

Related guides

• OpenRouter free models: the full catalog
• OpenRouter top models: usage rankings and prices
• The complete guide to agentic coding
• Best AI models of 2026