Split the bill first: GPT-Live-1 charges $0.05 per voice minute, and GPT-6 Astra charges $10 / $50 per million tokens on standard short context. Stacking a voice front-end and a flagship back-end on the same default chain is the usual reason an Agent invoice blows up.
Three checks: are you talking, can the task be downgraded, and does a single Input cross 272K. Figures are estimates from the OpenAI public price card. Refresh the site before you buy.
How much do GPT-Live-1 and GPT-6 Astra cost?
| Line item | GPT-Live-1 | GPT-6 Astra | How to choose |
|---|---|---|---|
| Primary unit | $0.05 / voice minute | $10 Input / $50 Output | Voice is minutes; reasoning is tokens |
| Cached Input hit | N/A | $1.00 / million | Only saves if the system prompt is stable |
| Cache write | N/A | $12.50 / million | A write is 1.25× Input |
| Batch / Flex | Live session | About half of standard | Overnight jobs only; do not mix half-price into the daily card |
| Fast / Priority | Queue separately | About 2× standard | Speed up only the last hop |
| Backend / tools | Billed on the model you pick | Web Search and similar extra | The voice layer is never an all-in price |
Astra Batch / Flex is $5 / $25. Copy that half-price row into a daily budget and you undercount the month by half. Long context (Input > 272K) reprices the whole job at $20 / $75 — not just the overflow slice.
Why can’t you compare voice-minute fees and token fees directly?
Live-1 uses POST /v1/live/sessions. Listen and speak run at the same time, billed by the second, with no rounding up to a full minute. Creating a WebRTC session bills 15 seconds of setup; that charge is credited once the session is actually running, so do not add another 15 seconds in an estimate.
Astra uses Responses / Chat Completions and bills tokens. The official recommendation is to hand reasoning to a backend model: keep the voice layer on Live-1, and pick GPT-5.6 Terra or Luna for thinking and tools. One session is those two charges added together.
| Name you hear | Budget against this | Do not mix it with |
|---|---|---|
| GPT-Live / Live-1 | $0.05 / minute voice layer | gpt-realtime-2.1 audio token rates |
| GPT-6 / Astra | $10 / $50 standard short context | Batch $5/$25, Fast $20/$100 |
| GPT-5.6 Terra / Luna | $2/$12 or $0.20/$1.20 | Treating “GPT-5.6” as if it were Astra |
Data-residency endpoints add about 10% on eligible models released on or after March 5, 2026. Astra Fast is unavailable under EU residency. For a three-vendor table, see Astra vs Fable 5.1 vs Flash.
How much for 1 million, 10 million, 100 million tokens and 1 hour of voice?
Estimates below assume standard short context, no cache, not Batch, not Fast. A typical Agent mix is 80% Input / 20% Output.
| Usage | Astra (80/20) | Terra (80/20) | Luna (80/20) | Read as |
|---|---|---|---|---|
| 1 million tokens | $18 | $4.00 | $0.40 | Trial, eval |
| 10 million tokens | $180 | $40 | $4.00 | A heavy personal week |
| 100 million tokens | $1,800 | $400 | $40 | Everyone defaulting to Astra hits this first |
| Voice length | Live-1 voice layer | Read as |
|---|---|---|
| 90 seconds | $0.075 | One short call |
| 10 minutes | $0.50 | One full support session |
| 1 hour | $3.00 | Billed by the second; 61 minutes = $3.05 |
| 40 hours / month | $120 | Backend tokens not included yet |
Which shortcuts get an Agent bill wrong?
The usual miss is to grab one “flagship price,” then multiply a fuzzy “a few million tokens a month.” What actually moves the number is the default model, whether context crosses 272K, and whether the voice layer stays open.
| Shortcut | What you undercount | Fix |
|---|---|---|
| Treat $5/$25 as everyday Astra | That is Batch/Flex; a live IDE pays $10/$50 | Budget at standard; list half-price separately |
| Default Astra every turn + always-on Live-1 | Minutes and flagship tokens stack | Voice on Live-1; backend default Terra |
| Stuff the whole repo into every request | Past 272K, the whole job is 2× Input and 1.5× Output | Summarize into cache; over the limit, switch to Luna |
| Multiply Input only and forget Output | Astra Output is 5× Input | Use your real in/out mix, not a 50/50 guess |
Web Search is about $10 per thousand calls, and retrieved text still bills model tokens. Hosted Shell / Code Interpreter bills container sessions (about $0.03 per 20 minutes at 1 GB). Tool fees never sit on the “per million tokens” row, but they do hit the invoice.
How do you estimate one Agent session from the public list?
This is a calculation protocol, not a lab test. Open the bill or a Usage export and run the same addition against your own numbers.
- ☐ Note session length in seconds, divide by 60, multiply by $0.05; write 0 if there is no voice
- ☐ Split out the backend model ID: Astra / Terra / Luna — do not write “GPT-6”
- ☐ Record Input, cache hits, cache writes, and Output separately
- ☐ Did Input exceed 272K? If yes, use the long-context rates for the whole job
- ☐ Fast or Batch? If yes, ×2 or ×0.5
- ☐ Add tool calls × the public tool rate
Total ≈ (voice seconds ÷ 60 × 0.05)
+ Input_M × model Input rate
+ CacheHit_M × cache rate
+ CacheWrite_M × write rate
+ Output_M × model Output rate
+ tool fees| Example (estimate) | Live-1 | Backend | Total |
|---|---|---|---|
| 8 min support + Terra 8K/1.5K | $0.40 | about $0.03 | about $0.43 |
| Same call, backend swapped to Astra | $0.40 | about $0.16 | about $0.56 |
| 8 min + Astra stuffed with 200K context | $0.40 | about $2.20 | about $2.60 |
Row three, ten times a day, 22 days a month, is about $572 from that pattern alone. Live-1 is not the expensive part — feeding the repo to Astra on every turn is.
How much is a light, mid, or heavy month?
The tiers exist so finance has a table you can talk from. They are not an internal invoice. Text-only uses a 5:1 in/out mix; voice support adds minutes.
| Tier | Assumed usage | All Astra | All Terra | How to read |
|---|---|---|---|---|
| Light · text | 5M in + 1M out | $100 | $22 | Personal trial, small tool |
| Mid · text | 30M in + 6M out | $600 | $132 | A small team can default to Terra first |
| Heavy · text | 100M in + 20M out | $2,000 | $440 | Add routing before everyone is on Astra |
| Light · voice | 20 hours + Terra 5M/1M | — | about $82 | $60 minutes + $22 tokens |
| Mid · voice | 80 hours + Terra 20M/4M | — | about $328 | $240 minutes + $88 tokens |
| Heavy · voice | 300 hours + Astra 30M/6M | about $1,500 | — | $900 minutes + $600 tokens |
The same mid-tier text mix on Luna is about $13. Subscriptions (ChatGPT Work / Pro) are a different bill from the API; see ChatGPT Work pricing.
Should a voice Agent back onto Astra or Terra?
The official starting advice is GPT-5.6 Terra; use Luna when cost is the constraint. Save Astra for multi-file refactors, long-document reasoning, and computer-use work — jobs that cost more to redo than to get right once.
| Entry | Voice layer | Backend | Call |
|---|---|---|---|
| FAQ / booking / status | Live-1 | Luna | Default this |
| Multi-step tools, need the right function | Live-1 | Terra | Official start tier |
| Multi-file code, long research, computer use | Voice can be off | Astra | Promote only this hop to flagship |
| Overnight scan, batch summary | No Live-1 | Terra Batch | Half price; do not occupy a live seat |
| Need tail latency down | Live-1 | Terra Fast | 2× price; last hop only |
// Upgrade by task, not by brand
if (task.voice) session.model = "gpt-live-1";
if (task.kind === "faq") backend = "gpt-5.6-luna";
else if (task.kind === "tool_agent") backend = "gpt-5.6-terra";
else if (task.kind === "hard_reason") backend = "gpt-6-astra";If you ship a voice product, then Live-1 + Terra, and keep Astra as an upgrade key. If you ship an IDE Agent that touches many files every turn, then put a daily cap on Astra and send chores to Luna. If context often sits near 272K, then summarize before you request — do not bet on a whole-job repricing.
gpt-6-astra only when redo cost is higher than the flagship delta.FAQ
How much does the GPT-Live-1 API cost?
The voice layer is $0.05/minute on the public list, billed by the second with no rounding up. Ten minutes is about $0.50. Backend models, tools, and phone lines are extra.
How much is 1 million GPT-6 Astra tokens?
Standard short context: Input $10, Output $50, cache hit $1, cache write $12.50. A 50/50 mix is about $30; 80/20 is about $18. Pro is not on this everyday table.
Which is more expensive, GPT-Live-1 or Astra?
The units are different, so you cannot compare them directly. On a short call the voice layer is often cheaper than the backend. In this set, the expensive combo is “Live-1 + Astra every turn + very long context.”
Which OpenAI API tier is cheapest?
In the tiers compared here, GPT-5.6 Luna has the lowest public token rate. It is fine for classify and FAQ work, not as the only engine for a multi-file refactor.
How much does an AI Agent cost per month on the API?
From the light / mid / heavy table above: text-only all on Astra is about $100 / $600 / $2,000. Export your own minutes and tokens, then multiply by the public rates.
Conclusion
This table is for you if you need to see Live-1 minute fees and Astra token fees as two lines, then use the same addition for an Agent month.
This table is not for you if you are still hunting for “one cheapest model that does everything.” That answer usually means an unrouted flagship default — the most expensive row on the card.
If the entry is voice, then Live-1 + Terra/Luna. If the entry is hard reasoning, then Astra, and hold the 272K line. Next, compare three flagships in Astra vs Fable vs Flash, or the subscription layer in ChatGPT Work pricing. You do not need to jump from this API card into a rental page.
