AICredits publishes its whole model catalog, with prices in rupees, at a public URL. That is rare, and it makes an honest data post possible: no screenshots, no hand-copied tables, just a script and the numbers it prints. This post analyses the catalog as of 2026-10-16 and points out what the numbers do and do not say.
Disclosure: AICredits links on this site are affiliate links. Everything below comes from the public endpoint, not from an authenticated account. I make no claim about model quality. Prices are per million tokens in rupees, as the catalog listed them, at the exchange rate in the data (95.10014 on every entry that day).
How do you get the data?
curl -s https://api.aicredits.in/v1/models -o models.json
No key was needed when I did this on 2026-10-16. The response is {"object": "list", "data": [...]}, 432 entries. Fields per entry include id, name, mode, context_length, supported_apis, is_active, is_free, last_updated, input_cost_per_token, output_cost_per_token, cached_input_cost_per_token (and the _inr versions), exchange_rate, and, for image models, an image_cost_matrix.
What is in the catalog?
models: 432 | active: 432 | free flag: 2
mode: {'chat': 378, 'image': 24, 'embedding': 22, 'audio': 6, 'tts': 2}
providers (id prefix), top 10: [('openai', 76), ('qwen', 48), ('google', 39), ('mistralai', 21), ('anthropic', 20), ('z-ai', 16), ('deepseek', 13), ('black-forest-labs', 10), ('groq', 8), ('meta-llama', 8)]
bare ids (no provider prefix): 24
supported_apis: {'chat_completions': 388, 'responses': 80, 'embeddings': 22, 'images_generations': 14, 'messages': 13, 'audio_transcriptions': 6, 'audio_speech': 2}
last_updated range: 2026-05-28 to 2026-10-13
What stands out:
- It is mostly chat. 378 of 432 entries. The other modes are small: 24 image, 22 embedding, 6 audio, 2 text-to-speech.
- Two providers dominate the count. OpenAI has 76 entries and Qwen 48, followed by Google at 39. Count of entries is not count of useful models; many entries are dated snapshots and variants.
- Only 13 models list the
messagesAPI. All are Anthropic models. Others, includingclaude-sonnet-5.5when I checked, list onlychat_completions. If your tool uses a provider's native API shape, check this field per model. - 80 models list the
responsesAPI in addition to chat completions. - Entries were refreshed recently: the oldest
last_updatedwas 2026-05-28 and the newest 2026-10-13. - The homepage count differs. The page I read says 300+ models in some places and 400+ in others. The live catalog says 432. Trust the endpoint.
How long are the context windows?
For the 375 chat models with an INR price, the median context length was 262,144 tokens, the maximum 2,000,000, and 103 models listed a context of at least 1,000,000 tokens. A large listed context is a ceiling, not a promise that quality holds across it; see long-context prompting for why.
Which chat models are cheapest, and which are most expensive?
Cheapest eight among chat models with context of at least 128,000 tokens, sorted by output price (rupees per million tokens):
mistralai/mistral-nemo 2.00 in 3.00 out ctx 131072
meta-llama/llama-3.1-8b-instruct 4.99 in 7.99 out ctx 131072
google/gemma-3-4b-it 4.99 in 9.99 out ctx 131072
mistralai/ministral-3b-2512 9.99 in 9.99 out ctx 131072
nex-agi/nex-n2.5-mini 2.50 in 9.99 out ctx 262144
ibm-granite/granite-4.0-h-micro 1.70 in 11.18 out ctx 131000
poolside/laguna-xs-2.1 5.99 in 11.98 out ctx 262144
upstage/solar-pro4 3.00 in 11.98 out ctx 524288
Most expensive five by output price:
openai/o1-pro 14978.27 in 59913.09 out
openai/gpt-5.4-pro 2995.65 in 17973.93 out
openai/gpt-5.5-pro 2995.65 in 17973.93 out
openai/gpt-5.2-pro 2096.96 in 16775.66 out
openai/gpt-5-pro 1497.83 in 11982.62 out
Across all chat models with a price, the output price ran from Rs 3.00 to Rs 59,913.09 per million tokens, a ratio of about 20,000 to 1, with a median of Rs 194.72. That spread is the real lesson. Model choice moves your bill by orders of magnitude, far more than any discount or gateway fee. It also means "which is cheapest" is the wrong question until you have defined the task. A model that costs 20 times more per token but finishes a hard task in one try can beat a cheap model that needs five tries.
What do embeddings cost?
The six cheapest embedding models all listed Rs 0.50 per million input tokens, for example baai/bge-base-en-v1.5, intfloat/e5-base-v2 and several sentence-transformers models. At that price, embedding cost rarely matters; retrieval quality and index maintenance do. See embedding models compared and hybrid search.
Are any models free?
Two entries carried is_free: true, both Google Lyria music-generation previews (google/lyria-3-clip-preview and google/lyria-3-pro-preview). No text model was flagged free in this snapshot. Do not read "free" as a promise it will stay that way: these are preview models.
How are INR prices derived?
For the 413 entries with a USD input price, the INR price equalled the USD price times the exchange rate times 1.05 on every one. The USD prices for five Anthropic models I compared matched Anthropic's pricing page that day. Details and the caveat about wallet debits are in the cost-per-feature post.
How should you use this list?
- Shortlist by constraints, not by price. Required context, required API (
messages,responses), modality, and budget per call. - Test on your own prompts. Use a small evaluation set and compare cost per successful task.
- Pin exact IDs. Look them up with the model ID lookup script and map old names with the migration checklist.
- Re-fetch before budgeting. Prices and the exchange rate move.
For a broader market view, AI model pricing 2026 covers vendors' own price lists. If you want to try the models with UPI, the AICredits signup is an affiliate link, and the AICredits review covers the service in more detail.
What I could not verify
- Quality, speed or reliability of any model listed. The catalog says none of that.
- Whether wallet charges match catalog prices (needs real usage).
- Any number on any day but 2026-10-16. The catalog is live.
