Why treat these separately

Vendor free tiers are rarely a single pool. They are usually split by model category: text generation, embeddings, image, speech (TTS/STT), and on some platforms video or rerank. Many accounts exhaust their text credits while their embedding and speech quotas sit completely untouched. Activating these categories separately gives you an extra layer of usable tokens for free.

Steps

1. Inventory categories before registering

Before signing up anywhere, list the models you need in four buckets:

  • Text generation (chat / completion)
  • Embeddings (vectorization, essential for RAG)
  • Speech (STT transcription, TTS synthesis)
  • Image (generation or vision understanding)

Then check each vendor's console under "models / usage" to see whether these categories have independent free quotas or share one pool. Platforms with a shared pool can be skipped; only independent quotas are worth splitting.

2. Create separate API keys per category

On platforms that support multiple keys, create one key per category and tag it (e.g. embed-free, tts-free). Two benefits:

  • The usage dashboard shows which category is consuming quota;
  • Exhausting one category does not affect the others.

3. Build a minimal "quota probe" script

For each category, write a minimal call that sends one request and logs the usage field. One run tells you whether the free quota is actually enabled, what the rate limit is, and whether a payment method is required.

Embedding probe (pseudo-code):

```python

resp = client.embeddings.create(model="<embedding-model>", input="ping")

print(resp.usage) # log token count and success

```

Speech is usually billed per second or per character, so the probe sends the shortest audio clip or a short sentence.

4. Route embedding quota into your RAG pipeline first

Embedding free quota is usually token-based, and vectorizing a document costs far fewer tokens than having a model read the same content. Use embedding quota for:

  • Vectorizing document chunks;
  • Vectorizing queries;
  • Small-scale reranking.

This consumes a lot of otherwise-idle free tokens while saving text generation quota for actual reasoning.

5. Batch speech and image quota offline

STT and image understanding often have separate free minutes/images. Use them to process a backlog of meeting recordings, screenshots, and scans in one pass, then feed the structured text to a text model. Confirm the free quota's duration first — usually stated as "about X minutes/month" or "about X images/month"; trust your console.

Caveats

  • Different units: text is tokens, speech is seconds/characters, images are counts. Don't estimate them the same way.
  • Independent rate limits: a category's RPM/TPM may be stricter than text; add backoff retries for batch jobs.
  • Region and account type: some categories are not available on the free tier in some regions or account types; trust what the console shows.
  • Data compliance: free tiers often allow usage for service improvement; don't send sensitive audio/images through free channels.
  • No fabricated numbers: quotas change; this article gives no fixed figures. Always check your console's usage/quota page.

When this applies

  • Teams building RAG apps where vectorization cost is high but text reasoning needs are low;
  • Individual developers with large backlogs of recordings and screenshots to transcribe;
  • Indie developers who want to max out free quota without paying separately for embeddings.

Wrap-up

Treating official free tiers as "several small pools split by category" rather than "one big pool" is one of the most underrated free-token sources in 2026. Inventory first, split keys second, validate with probes third, then allocate by category. Set it up once and reuse it long-term.