If you searched “GPT-6 API price” or “how much per million tokens,” you probably want a table you can show finance—not a 15-section ranking. Comparison pages still split Input, Output, cache, and Batch into four unit systems, so a first-time reader cannot tell which SKU they are actually buying.
By September 2026 the public gap is huge: a flagship can cost about 50× a volume tier. Read the numbers first, then decide whether you need routing.
Every dollar figure below is an estimate from public list prices as of 2026-09-10. Refresh the vendor cards before you buy. This is not a leaderboard and it does not use unpublished internal invoices.
1. 2026 API price table for the three models
All figures are USD per million tokens. Astra is the public GPT-6 flagship, not Pro. Fable 5.1 is the coding / agent revision. Flash is Google’s volume SKU.
| Line item (as of 2026-09-10) | GPT-6 Astra | Claude Fable 5.1 | Gemini 3.8 Flash |
|---|---|---|---|
| Input | $5.00 | $6.00 | $0.10 |
| Output | $20.00 | $30.00 | $0.40 |
| Cached Input | $0.50 | $0.60 | $0.01 |
| Batch (offline) | ~50% off | ~50% off | ~50% off |
2. How each model is billed
Match the SKU before you trust a screenshot. “GPT-6” on a comparison page is almost always Astra; Pro is a different, dearer card. Fable 5 vanished from some IDE lists—put 5.1 in config, not in a hardcoded constant. Background: Fable 5 and agent infrastructure.
| Label you see | Budget against | Do not confuse with |
|---|---|---|
| GPT-6 / Astra | Astra: $5 / $20 | GPT-6 Pro, cheaper distilled SKUs |
| Claude Fable 5.1 | Fable 5.1: $6 / $30 | Retired Fable 5, Haiku-class models |
| Gemini 3.8 Flash | Flash: $0.10 / $0.40 | Gemini flagship / deep-think modes |
Input is what you send; Output is what comes back and is usually 4–5× dearer. Cache hits are about 1/10 of Input—useful for a stable system prompt. Batch is for jobs that can wait overnight, not a live IDE loop. Seat subscriptions are a different bill: 2026 AI coding cost ranking.
3. What 1M, 10M, and 100M tokens cost
Linear estimates from the public card (no tax, no gateway markup). “100M tokens” here means 100M Input + 20M Output—roughly a mid-intensity team month.
| Usage (estimate) | Astra | Fable 5.1 | Flash |
|---|---|---|---|
| 1M Input + 1M Output | $25 | $36 | $0.50 |
| 10M Input + 2M Output | $90 | $120 | $1.80 |
| 100M Input + 20M Output | $900 | $1,200 | $18 |
| 30M Input + 6M Output / month | ~$270 | ~$360 | ~$5.40 |
If 70% of Input hits cache, that monthly row falls to about $176 / $247 / $3.50. Misses still pay full Input.
4. What one coding pass costs
Treat “one review” as about 80k Input + 8k Output (a few files, one change, a short note). This is list-price math, not a lab invoice:
| One review (estimate) | Astra | Fable 5.1 | Flash |
|---|---|---|---|
| 80k In + 8k Out | ~$0.56 | ~$0.72 | ~$0.011 |
The same task from an IDE agent often replays rules and the tree, so tokens can run 2–3×. Chat-and-paste uses fewer rounds but more of your time. Run ten jobs on your repo, average the tokens, then multiply.
5. Which model fits coding, API, and CI
A low unit price does not mean the model can land a compile. Pick the default by where the request starts.
| Job | Better default | Backup | Do not default |
|---|---|---|---|
| IDE / agent, multi-file edits | Fable 5.1 | Astra | Flash alone on a refactor |
| Product chat, plans, multimodal Q&A | Astra | Fable 5.1 | Flash as the only brain |
| CI logs, classify, summarize, translate | Flash | Astra Batch | Fable sweeping every repo |
| Keep the month in tens of dollars | Flash + pinpoint upgrades | Astra + cache | Three flagships left on |
Tool calling on Gemini: Gemini 2026 agents and structured output. How the GPT-6 band moved: GPT-6 API pricing predictions.
6. How to lower API cost
- ☐ Default to Flash or “auto”; upgrade only for multi-file or failed builds
- ☐ Cache a stable system prompt and repo summary
- ☐ Daily caps on Astra and Fable; degrade instead of silent Pro
- ☐ Overnight jobs go to Batch, not a flagship IDE loop
- ☐ Keep model IDs in config, with a Flash fallback
// Upgrade by task, not by brand
if (task.kind === "classify" || task.kind === "summary") model = "gemini-3.8-flash";
else if (task.kind === "refactor" || task.touches > 3) model = "claude-fable-5.1";
else model = "gpt-6-astra";A router saves SDKs. It does not save money if the default still hits the dear SKU.
7. How to choose
| Your situation | Start here |
|---|---|
| Student / indie already on Google | Flash first; Astra for the hard reasoning turn |
| Daily multi-file agent work | Fund Fable 5.1; Flash for chores |
| You must brief finance | Use §3 with your own token counts |
One line: Flash is cheap and fast, Fable 5.1 lands the edit, Astra explains the plan. Stacking all three jobs on one default is how the bill spikes.
FAQ
How much is 1 million GPT-6 tokens?
On the 2026-09-10 Astra card, 1M Input is $5 and 1M Output is $20. A 50/50 split is about $12.50. Pro is not on this everyday table.
Is Input or Output more expensive?
Output, on all three: about 4× on Astra and Flash, 5× on Fable. Long regenerations hurt more than long prompts.
How are cached tokens billed?
Cached Input is about 1/10 of normal Input ($0.50 / $0.60 / $0.01 per million). Misses still pay full price.
Which of these APIs is cheapest?
Gemini 3.8 Flash on the public card. Use it for classify, summarize, and CI—not as your only refactor engine.
What might AI coding cost per month?
On the 30M + 6M row: about $270 Astra-only, $360 Fable-only, $5.40 Flash-only. Real teams mix. Measure tokens, then multiply.
Summary
This piece is for you if you need the 2026-09 list prices and a way to estimate a monthly bill from public math.
It is not for you if you still want one “best” champion. That answer is usually an unrouted flagship—the dearest row on the table.
Seat pricing: AI coding cost ranking or ChatGPT work-tier prices. If the real question is Xcode or a local agent runtime, read that cluster next—you do not need this price table to push a Mac order.
