Five of OpenRouter's free models take a million tokens of context

I went looking for the cheapest way to throw a large codebase at a model and found this in OpenRouter's public catalog — no key needed to check:

curl -s https://openrouter.ai/api/v1/models \
| jq -r '.data[]
  | select((.pricing.prompt|tonumber)==0 and (.pricing.completion|tonumber)==0)
  | select(.context_length >= 1000000)
  | "\(.context_length)\t\(.id)"' | sort -rn

As of 2026-08-28, five text models come back at $0 per token with a million-token window or more:

1048576  thinkingmachines/inkling
1048576  thinkingmachines/inkling-small
1048576  minimax/minimax-m3
1000000  nvidia/nemotron-3-ultra-550b-a55b
1000000  nvidia/nemotron-3.5-lightning

(Two Google Lyria audio previews also return $0 at 1M; I've left them out since they aren't chat models.)

For scale: 1,048,576 tokens is roughly 4 MB of source. Nemotron 3 Ultra is a 550B-parameter MoE. Free.

The catch is throughput, not context. Free models on OpenRouter run at 20 requests/minute and 50 requests/day until the account has purchased $10 in credits at some point, after which the daily cap goes to 1,000. So this is a "read something enormous a few times a day" budget, not a "serve users" budget — which happens to be exactly the shape of a one-off repo analysis or a nightly summarization job.

Worth knowing that 119 paid models in the same catalog also clear 1M, so if you outgrow the request cap the migration is a model ID, not an architecture.

Current free-model list and how the limits work: OpenRouter free models