OpenRouter Auto Router: check your budget after the update
Your AI workflow still runs, but the usage bill is creeping up. If it uses OpenRouter’s Auto Router, check which setting it uses: the updated router has a more expensive default policy. Compare a small batch of results before deciding whether to keep that default, try a lower cost preference or return to a model you already trust.
The same connection now makes different choices
TLDR AI’s November 9 edition highlighted OpenRouter’s new Auto Router and Auto Code. OpenRouter announced them on November 6. Auto Router chooses which AI model handles a request; Auto Code is the separate option aimed at coding work.
Existing Auto Router connections still use the name openrouter/auto. OpenRouter says its new default costs more per task to allow a wider choice of models. Its documentation recommends setting cost_tier to low if you want costs closer to the previous behaviour. That is a preference, not a promise about your next bill.
If someone connected your workflow to Auto, ask them to confirm the setting your app sends. A rise in spending alone does not prove the router caused it: more requests or longer material can also change the total.
Discovery: TLDR AI, November 9OpenRouter’s November 6 announcementCurrent Auto Router settings and limits
Compare completed work, not the cheapest reply
Check whether your team can use the result. Paying less for a draft is no help if someone must rewrite it. Paying more only makes sense if the answer helps them finish the job.
Compare the same prompts and material using your current Auto setting, an explicit low preference, and your existing fixed model if you have one. Separate routine tasks from awkward ones so easy successes do not hide expensive failures. This is a proposed test, not a result we have measured.
The fictional shop below is preparing copy from public product information. Use public material or examples created expressly for testing. Write down what a useful answer must contain before you read the answers.
| Task | A usable result | What to record |
|---|---|---|
| Summarise a clear product sheet | The answer keeps the supplied dimensions and included accessories. | Record the model, charge and review time. |
| Compare two conflicting product descriptions | The answer points out the conflict and asks for clarification. | Record missing or invented facts and correction time. |
| Draft copy when a specification is absent | The answer asks for the missing specification. | Check whether it makes an unsupported claim. |
A routing fee and an AI bill are different things
As checked on November 10, OpenRouter lists automatic routing on Standard, Business and Enterprise, but not Free. Standard and Business use prepaid credits. The published credit-purchase fee is 5.5% on Standard and 8% on Business.
Auto Router adds no separate routing fee: requests are charged at the selected model’s rate. That still leaves a usage bill, plus the credit-purchase fee. Do not treat a zero router price as free AI work.
Keep the API charge and staff correction time as separate columns in your comparison. Add charges for repeated attempts that return answers you cannot use. Saving someone time can be worthwhile without pretending their salary has disappeared.
Check whether your account and workflow fit
The current Auto Router documentation says it is available only in the global data region. If you require processing within a specific region, do not assume a Business or Enterprise plan makes Auto suitable. Check the available routing options before switching.
Auto can respect allowed models and provider restrictions. Have the person maintaining the connection keep those requirements the same across your comparison. Otherwise, you may be comparing a restricted setup with an unrestricted one.
OpenRouter’s chat interface can show why Auto chose a model. Use it to explore the choices, then test through your actual app. You also need to check how that app sends instructions and uses the answer.
Keep the setting that earns its cost
Keep the new default if the awkward tasks become meaningfully easier to finish and the extra charge is acceptable. Try low if routine work stays useful at a lower observed cost. Keep a known fixed model when it already handles a narrow, repeated task well and changing the choice adds little.
Wait on a wider switch if you cannot yet see which model ran or what each result cost. Ask the person maintaining the workflow to make those details visible first.
Sources
Review one AI workflow.
Describe a recurring task to forvertz and explain what a useful answer should contain. We can discuss a comparison using a public sample or a fictional example.
Discuss your AI workflowUpdated