Today we tell the story of our own failure. Nothing terrible happened, but it is instructive.

On 13 October at 15:37 UTC a post about Noam Shazeer moving from Google to OpenAI went out to our Telegram channel and our site. The text of the Russian version was in Polish: "współautor… opuszcza firmę na rzecz OpenAI" — it is in the picture at the top. Nobody ordered Polish. We did not fix the post — let it stay as a reminder.

How it works on our side

News on aiexpertise.pro is translated automatically by free models through the OpenRouter service. Once an hour our code takes every model marked ":free", sorts them by context size and response length and tries them one by one: if the first fails, it takes the next. Quality and response language play no part in the choice. We described how we work in general in How our AI writes articles.

What went wrong

1. The first models in the queue stopped answering. Two Thinking Machines models, Inkling and Inkling Small, were returning 403 errors to us. According to the company, free access to them is offered only inside "agentic" programs (such as Claude Code or Codex), only for a limited time, and the data sent through it is used to improve the model (Thinking Machines post, model page on OpenRouter). Our script is apparently not one of those programs, so every story first collects two refusals.

2. Next in the queue was a small, fast model. That is nvidia/nemotron-3.5-lightning:free, which appeared on OpenRouter in August 2026. In recent days it translated almost all our news without trouble: about 43 of the last 50 stories. On 13 October it answered in Polish.

3. Nothing checked the result. Russian text must be in Russian, but we never verified that anywhere. The model did not report an error: it produced a confident, fluent text in the wrong language. For a program, that counts as success.

4. The story was old. We presented the post as fresh, but the item from one of the services we take news from was published back on 18 June 2026: Shazeer announced his departure from Google in June. Four months later we reported it as if it had just happened.

What we learned

  • A free offer can vanish without warning. The company described free access to Inkling as temporary (its post says "for the next few weeks") and limited to certain programs. What is free today may close tomorrow, or need a different way in.
  • Free requests may feed training. When a service is free, you may be paying with data. For Inkling this is stated outright: prompts and outputs are logged and used to improve the model. We translate public news, so it does not matter to us. If you upload a contract or a private conversation to a free model, think twice.
  • Limits and refusals under load are normal. Free models on OpenRouter have their own caps on requests per minute and per day (OpenRouter docs), and the service itself says free models are usually not suitable for production use (OpenRouter FAQ). We built part of our pipeline on them. That was our choice, and these are the consequences.
  • A model can be wrong silently. A neural network does not say "I am not sure". It writes in Polish as smoothly as in Russian.
  • The model list shifts. Models come and go, and ordering "by size" can accidentally put the one that does not answer today at the top and the one that answers oddly in the middle.

What can be done about it

We have not fixed anything yet: one Polish post is not a reason to rebuild everything. But the cures are well known:

  1. Automatic result checks. If a "Russian" post contains almost no Russian letters, treat the answer as a failure. Format can be checked the same way: length, hashtags, empty replies.
  2. A fallback model. If the check fails, take the next model; if none passes, leave the post to a human.
  3. Choosing models by quality, not size. A list of tested models instead of "everything that is free".
  4. Checking the source date. Do not treat a story older than a few days as fresh.

What is useful for you

You do not need to run a news site to step on the same rakes.

  • Read the output. Especially if it is in a language you know less well than the model pretends to.
  • Do not send a free model anything you would not show a stranger. Look at the terms: what is stored and where it goes.
  • Do not build on a free tier what you cannot do without. Keep a backup in case access closes.
  • Check the date. "News" can be old. Open the source.
  • Remember what worked. If a task repeats, pin the model that handled it and its settings instead of choosing anew every time.

If you have similar stories, share them in the comments in our Telegram channel, Pro AI Expertise.

Facts about the services are taken from their official pages as of 13 October 2026; the terms of free models change.