Generate Anki flashcards with any model, one Base URL.

Quick answer: the three fields, wherever the add-on puts them.

Anki's AI add-ons are independent projects, so the settings layout varies, but the pattern is the same three fields every OpenAI-compatible tool uses: an API key, a Base URL or endpoint field, and a model name. Set the Base URL to https://api.apisrouter.com/v1, paste your APIsRouter key, and type the exact catalog id as the model. A concrete, verified example: the add-on "Anki AI - Enhance Your Flashcards with ChatGPT and other LLMs" (AnkiWeb code 643253121) added a custom Base URL setting in February 2026, explicitly so it can run against services matching the OpenAI API, and it accepts a free-text custom model name. Its own page flags the feature as new, so treat the exact field labels as version-dependent and the pattern as stable.

API Key:     sk-YOUR-APISROUTER-KEY
Base URL:    https://api.apisrouter.com/v1
Model:       claude-haiku-4-5-20251001
             (free-text field: any catalog id works)

How AI add-ons use the endpoint.

Anki is the spaced-repetition app with millions of users, and its add-on ecosystem is where AI arrived: add-ons that generate example sentences for language decks, fill definition and mnemonic fields from a note's front side, draft cloze deletions from pasted text, or explain a card you got wrong. Under the hood these are chat-completions calls with a prompt template per field, sent once per note. That per-note shape is why the endpoint choice matters more here than in a chat app. A deck run is a batch job: a two-thousand-note deck with one generated field is two thousand requests, each small, all shaped identically. Volume-tier models handle this beautifully, and the difference between a flagship and a floor-rate id across a big deck is the difference between a meaningful bill and a rounding error. Honesty about the ecosystem: not every Anki AI add-on exposes a Base URL. Some hardcode OpenAI's hosts, and each add-on names its fields its own way. Before topping up anything, open your add-on's config and confirm an endpoint or Base URL field exists; if it only takes an API key with no endpoint setting, it cannot route to a gateway, and you should pick an add-on that can.

Choosing models for card generation.

Batch jobs make A/B testing trivial: run fifty notes through two ids, read the cards, and check the per-model spend in the console. The winner is usually obvious within one screen of cards, and the cost of the experiment is effectively nothing at volume-tier rates.

• deepseek-v4-flash for bulk deck runs: definitions, translations, and example sentences at the lowest rate in the catalog.
• claude-haiku-4-5-20251001 when generation quality on nuanced fields, mnemonics, usage notes, subtle sense distinctions, is worth a step up.
• gpt-5.4-mini as the GPT-family volume option if your prompt templates were tuned against OpenAI behavior.
• gemini-3.5-flash for decks generated from long source texts, where a big context window earns its rate.
• glm-5.2 worth an A/B for Chinese-language decks, where its bilingual output is a genuine strength.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Failure modes specific to Anki add-ons.

Since the add-on layer varies, the reliable debugging move is the same curl pair as always: models listing, then one completion with your exact id. If both pass, the gateway is fine and the issue is in the add-on's config or version.

• No Base URL field in the config: the add-on hardcodes its vendor. No gateway can help; switch to an add-on that exposes an endpoint setting, like the verified example above.
• Model dropdown without free text: some add-ons list only vendor model names. Look for a custom model option; the verified example accepts typed-in model names.
• 401 mid-batch: the key was pasted with whitespace or truncated. Curl the models listing with the same key to confirm outside Anki.
• Batch stalls or errors partway: per-minute rate limits meeting a burst of requests. Rerun the remaining notes; batches are resumable in most add-ons, and pacing settings help where offered.
• Field labels do not match this page: expected. Add-ons rename fields across versions; match the pattern (key, endpoint, model) rather than the exact label, and check the add-on's AnkiWeb page for current instructions.

Who routes Anki through a gateway.

• Language learners batch-generating example sentences and translations across thousand-note decks at volume rates.
• Medical and law students filling mnemonic and explanation fields on dense decks, where card quality justifies a Claude id.
• Deck authors who publish shared decks and regenerate fields at scale each revision.
• Students without a vendor billing relationship, since the prepaid balance starts free with no card.
• Tinkerers already running other tools through a gateway who want deck generation on the same key and usage log.

Verify the endpoint before a big deck run.

Run the two standard checks with the id you plan to use, then generate a handful of cards before committing a full deck. The three-field config plus a five-note test run catches every common failure in under five minutes. After a batch, the APIsRouter console shows the run as it actually billed: request count, tokens per request, spend per model. For deck generation this is worth reading once, because it turns "roughly how much does a deck cost" into an exact number you can scale linearly to any deck size.

curl -s https://api.apisrouter.com/v1/models \
  -H "Authorization: Bearer $APISROUTER_API_KEY" | head -50

curl -s https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $APISROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash",
       "messages":[{"role":"user","content":"Example sentence for the word Fernweh"}]}'

FAQ