Free OpenRouter Models Benchmarked: November 2026 Results

Tested October 28 to November 6, 2026: we sent every free OpenRouter model 282 requests over 10 days. Only five of the 16 models answered every request, and all five were still free on November 6: Nemotron 3.5 Super VL, Nemotron 3 Ultra, Nemotron 3.5 Lightning, Cohere North Mini Code and Dots3-Note Preview. The best result was Nemotron 3.5 Super VL (nvidia/nemotron-3.5-super-vl:free): 20 of 20 answered, 20 of 20 passed, a median 106 output tokens per second, and a 1,000,000-token context. If you don't want your prompts used for training, Cohere North Mini Code answered 20 of 20 and passed 18.

Both Gemma 4 free models, the ones most people try first, answered 0 of 20 requests. Every miss was a 429 temporarily rate-limited upstream, while OpenRouter showed their endpoints at 100% uptime.

This is a test of free endpoints from a free-tier key at one time of day, not a quality leaderboard. For the full catalog with context windows and limits, see OpenRouter free models. For paid models ranked by price, see cheapest OpenRouter models.

Results: every free model, 10 days

As of 2026-11-06 from our daily run against https://openrouter.ai/api/v1/chat/completions. "Answered" is a reply on the first try, with no retries. "Passed" counts only answered requests, graded automatically. Speed is the median of output tokens divided by total request time, on answered requests. Models are sorted by answered, then passed.

How we computed it: the task columns show passed / answered for that task. Each model got two of the three tasks a day, rotating, so a model sampled on all 10 days has 20 requests, with 6 or 7 per task. Poolside's models were sampled on 3 days before their announced withdrawal, Inkling on 8 and Content Safety on 9, because the script tests whatever the catalog lists that day. Across all 282 requests, 138 were answered (49%).

What the numbers mean

• Availability decided the ranking, not intelligence. Ten of the 16 models answered half of their requests or fewer. Every Gemma 4 and Xing4.0 miss was a 429 from a full shared pool, every Inkling miss was a 403 because those endpoints only serve agentic coding harnesses, and Flint's 10 misses between October 30 and November 3 were no upstreams available. Flint came back on November 4 and passed 6 of 6 since.
• Uptime is not availability. On November 6, OpenRouter's endpoint API reported 100% uptime over the last day for both Gemma 4 free endpoints, and 99.7% for Xing4.0. Our free-tier key still got through to Gemma 4 zero times in 10 days. Uptime says the provider was up, not that a free request gets a turn.
• Tool calling worked whenever a model answered. 44 of 44 answered tool-call requests across nine models named the right function with the right arguments. If a free model fails your agent, the cause is more likely a 429 than a bad call.
• Two reasoning models burned the JSON budget. Dots3-Note Preview and Nemotron 3.5 Lightning answered all 14 JSON requests, and all 14 replies used exactly the 400-token cap without parseable JSON. The same two models passed 13 of 14 coding tasks, where the cap was 1,500 tokens. We did not retest with a larger budget or reasoning off, so treat their JSON score as "needs a bigger budget", not "can't do JSON".
• The smallest model was the fastest and the least useful. LFM2.5-2.6B produced a median 537 tokens per second but answered 5 of 20, and 3 of its 5 answers hit the 1,500-token cap with no code.
• Nemotron 3 Ultra was flawless and slow. 20 of 20 passed, at a median 29 tokens per second and 6.7 seconds a request. The coding task took it a median 12 seconds, against 3.9 for Flint and 5.6 for Super VL.

Which free model to use

Data-policy notes come from each free endpoint's OpenRouter page as checked on October 28, 2026, and are listed per model in our free coding models guide. The four NVIDIA free endpoints and LFM2.5 may use prompts for training. If you have turned off training in your OpenRouter privacy settings, OpenRouter won't route to those endpoints, and the models that answered every request in this test will fail for you.

What changed in the free catalog during the test

OpenRouter listed 16 free text models on October 28 and 11 on November 6. The five that left:

• poolside/laguna-s-2.1:free and poolside/laguna-xs-2.1:free: withdrawn on October 31, as OpenRouter had announced. Laguna XS passed all 3 tasks it answered before leaving.
• thinkingmachines/inkling:free and thinkingmachines/inkling-small:free: last listed in our November 4 run. Both answered every request with a 403: "only available on agentic harnesses". Our script is not a coding agent, so these two were never testable here.
• nvidia/nemotron-3.5-content-safety:free: last listed in our November 5 run. It is a moderation classifier, not a chat model. From October 30 its endpoint returned a 404, and its three answers before that were neither valid JSON nor working code.

Nothing was added. If you use a free model in anything that runs on a schedule, check the catalog for an expiration_date and keep a paid fallback. Dots3-Note Preview and North Mini Code, the two models in the top five whose free endpoints don't train on prompts, have no paid version on OpenRouter, and Dots3-Note Preview is listed only until December 31, 2026.

How the benchmark works

• Who sent it: a free-tier OpenRouter account with no credits, which is the situation of most people searching for free models. Free-tier keys share each model's free pool and are limited to 50 free requests a day, so each model got two of three tasks per day.
• When: one run a day, starting between 12:34 and 14:40 UTC on nine days and 16:16 UTC on October 28. Free pools fill up at different hours. A model that was full for us at 12:34 UTC may be open at 03:00 UTC.
• Tasks: (1) extract an order into JSON with exactly four keys, graded on exact values; (2) call a get_weather tool for Bratislava in Celsius, graded on function name and arguments; (3) write a JavaScript run-length encoder, run against five test cases in a macOS sandbox. temperature 0, max_tokens 400 for JSON and tool, 1,500 for code. One attempt, no retries, no system prompt.
• What it does not measure: writing quality, long-context recall, multi-turn behaviour, image input, or anything at the sizes real agent runs use. Three tasks on 20 samples separate "answers and follows instructions" from "doesn't", and nothing finer.

The raw rows and today's catalog are saved next to the script that produced them, in the repo under docs/marketing/data/free-openrouter-models-benchmark-2026/. The next refresh adds the new days and rewrites the changed rows.

Run a model that answers on real work in Teamday

None of these free models is in Teamday's model picker. Teamday's OpenRouter models are paid ones, such as Kimi K3 and GLM-5.3 Flash, which don't share a free pool, so an agent's scheduled task doesn't stop on a 429. Save your OpenRouter key in runner settings and usage bills to your own OpenRouter account. See current pricing.

Related guides

• OpenRouter free models: the full catalog with limits
• Best free OpenRouter models for coding
• Free OpenRouter models for roleplay and Janitor AI
• Cheapest OpenRouter models by capability
• OpenRouter top models: usage rankings and prices