ROUTING
Small questions stop costing frontier prices
Every message is scored, then sent to the cheapest model that clears the bar.
0.81
Sonnet 5
−5 cr
Solutions · For you
The right model for the question, without you learning which is which.
Ask once. Avna sends it to the model that fits, names the one that answered, and shows what it cost. Pro is 899,000 Toman, about $5.00 — ChatGPT Plus is $20, and it will not take your card.
Claude Haiku 4.5Grok 4.3DeepSeek V4 ProDeepSeek V4 FlashQwen3.7 MaxClaude Fable 5Claude Sonnet 5GPT-5.6 TerraGPT-5.6 LunaGrok 4.5Kimi K2.7 CodeMiniMax M3GPT-5.6 SolClaude Opus 5MiMo-V2.5MiMo-V2.5-Pro
The models it picks between, read from the live catalogue rather than typed here. The list is the proof of the sentence above it, not the offer: you are not buying a shelf of models, you are buying not having to stand in front of one.
- Scored and routed per message, retried on another model if a provider fails
- Every reply names the model that answered and why, the full cost sits in your usage history
- Long-term memory you can read, edit and delete
- The same assistant in the browser, by voice, and on Telegram
Live demo
Pick a prompt. Watch Avna choose
Plan a 3-day Shiraz trip
GPT-5.6 Solgood at itineraries, no need for a frontier model
Day 1: Nasir al-Mulk at opening for the light, then Vakil Bazaar and the bathhouse beside it; Hafez's tomb at sunset. Day 2: leave at 7 for Persepolis and Naqsh-e Rostam, back by mid-afternoon with nothing else planned. Day 3: Narenjestan and Eram Garden, then the old quarter for dinner. Do Persepolis as a taxi day, not a tour. You'll want to leave when you're done.
This message
Credits used8 cr
one flat currency across every model
Tokens1,460
after context compression
Served from cache38%
you are not charged twice
CREDITS LEFT1,992 cr
8 credits used in this demo · this demo wallet holds 2,000
Saved 1 credits on this message
vs. routing everything to the frontier model.
Plan 3 days in Shiraz for two: Hafez, Persepolis and good food, mid-range budget.
GPT-5.6 Sol
A tight three-day plan with travel times between stops and a Persepolis start that beats the coaches. Names the garden restaurant most guides miss.
CREDITS
8 cr
TOKENS
3,800
TIME
3.2s
Sonnet 5
Three clear days in plain lists: one must-see per morning, afternoons left loose. Flags that Hafez at sunset means being there by six.
CREDITS
4 cr
TOKENS
2,900
TIME
2.1s
DeepSeek V4 Pro
A workable three days. Puts Persepolis on day one, which costs you the early light at the mosque, but its food picks are the most specific of the set.
CREDITS
2 cr
TOKENS
2,400
TIME
2.7s
Three models pulled at random from the 6 in this demo. Credits, not dollars.
The models
You never pick the model. Avna picks, and shows its work.
Every chat model from OpenAI, Anthropic, xAI and the leading open labs, on one balance. Each message is scored and sent to the cheapest model that clears the bar, and the reply tells you which one answered and why.
CACHE
You stop paying full price for repeats
Repeated context is stored once and reused. Cached tokens are billed at half price.
3,140
1,902
2,189
FALLBACK
A provider outage doesn't reach you
If a model fails, Avna retries your message on another one and tells you it did. You get an answer, not an error page.
Haiku 4.5
1 answer
none
You can always see what answered
Elsewhere you can't tell what an answer cost or why that model replied. Here the reply names the model and the reason, your profile itemises every request, and the router works for you, not on you.
- On the reply: the model, why it was chosen, and whether it fell back
- Pick a routing preference (cost, balanced, quality or speed) and the router leans that way
- Date, model, tokens and credits for every request in your profile, exportable as CSV
avna: usage · one request
- model
- Opus 5 → Sonnet 5
- reason
- reasoning score 0.81 · cheaper at equal quality
- tokens in
- 3,140 (1,902 cached, billed at half)
- tokens out
- 806
- fallback
- none
- credits
- 4 cr
- latency
- 2.4s
Every memory is yours
Read, edit, export or delete any line Avna remembers, or turn memory off entirely.
Every request itemised
Date, model, tokens and credits for every call, downloadable as a CSV.
No card on file
Nothing is stored, nothing auto-renews. You top up when you decide to.
Toman or crypto
Iranian gateways, or USDT and USDC on BEP20. No international card needed.
Persian, properly
Right-to-left layout, Persian digits and Jalali dates, not a translated English app.
It says when it looked things up
When Avna searches the web to answer you, the reply says so.
You write a rough thought. It sends a real prompt
A model suggestion with its reason before you send, a prompt optimizer that shows its rewrite, saved styles and templates, and two of them you can try right here: the caveman answer style, and the memory list you can edit.
Model suggestion
Before you send, Avna tells you which model fits and why.
One tap to override
Prompt optimizer
Rewrites your rough thought into a prompt with the context the model actually needs.
You see the rewrite before it sends
Templates & styles
Save a tone once ("blunt", "client-safe") and apply it to any thread.
8 styles + 7 templates built in
One-click summary
Turn any thread into a brief, an email or a task list without re-prompting.
One tap, no re-prompting
Caveman Mode
YOUWhy is this query slow?
🦴 CAVEMAN12 words · 4s read
Missing index. Add one on (customer_id, created_at). 8.2s → 340ms.
✓ No preamble✓ No hedging✓ Stops when done✓ Numbers over adjectives
On: no preamble, no "great question", no three-paragraph wind-up. Short words, straight answer, and it stops when the answer is done.
Memory you can edit
Everything Avna remembers is a line you can read, change or delete. No invisible profile.
- Writes in British EnglishAll chats
- Ships to production on ThursdaysWork
- Learning Spanish: trip in NovemberGoals
- Hates em-dashes in emailsWriting
Try deleting one. Nothing is hidden from you.
Same account, same memory, three different places.
It is a browser tab, a Telegram contact and a voice you talk to, one subscription behind all three, and what you told it in one is known in the others.
- One subscription, one bill. The browser, Telegram and voice are the same account.
- What you tell it in one place is known in the others, including what not to repeat.
- Each answer arrives in the shape of its channel: buttons in Telegram, a voice note out loud.
- It listens and speaks in Persian, including a sentence that switches to English halfway.
09:40 · at your desk
Draft a reply to Sana Group about the delay, apologetic but not grovelling.
Here's a draft. I've kept it short, led with the new date rather than the apology, and left out the warehouse detail you told me not to share with customers.
GPT-5.6 · chosen for this message0.4 credits
Every reply carries what answered it and what it cost. Nothing is estimated.
18:10 · on the metro
Avna
online
remind me to follow up Sana Group on Thursday
Done, Thursday at 10:00, “follow up Sana Group on the delayed order”. I've attached it to the draft you wrote this morning, so you'll have it in front of you.
✓ DoneSnooze to Friday
The same assistant, as a contact. It knows about the morning's draft because it is the same account, not a copy of one.
21:25 · hands full
0:07
Heard: “What's left for tomorrow?”
Three things: the Sana Group follow-up at ten, the wallet load test, and you said you'd book the dentist. Only the first one has a time on it.
Spoken back as a voice note · Persian and English
A voice note in, a voice note out, in Persian, which is the part most assistants get wrong.
The rest of your day's AI
Voice mode
Talk instead of typing, record, hear the answer, keep going hands-free. Persian included.
Web search on auto
A classifier decides per message whether live results are needed, answers get badged when it searched.
Images in the same thread
Generate and discuss images without switching tools.
Compare lab
Run one prompt across models side by side, history, votes and a leaderboard built from your own results.
Conversations under control
Pin, search, archive, share a read-only link, reset context mid-thread, and jump around long threads with an outline.
Pay from Iran, or in crypto
Toman gateways or a USDT/USDC checkout, feeding one credit balance either way. Spark codes redeem at sign-up.
The right model, without thinking about models
No card. Persian or English. · what we keep and throw away →