AY Automate / Original Research
What the major AI models actually cost per token, as of September 2026.
Every price below was re-checked against the vendor's own pricing page on 2026-09-26, not copied from a roundup or a cached summary. Each row notes what changed since the first edition of 2026-07-23. Prices change; the source link next to every row is where to check the current number.
$0.10
Cheapest input, per M tokens
$180
Most expensive output, per M tokens
3
Vendors, all live-verified
17
Models priced
What this index is, and is not
Every vendor publishes today's price on its own pricing page. What none of them publish is the trend: what a model cost last quarter, or how a price change lands against everyone else's. This page starts as a single dated snapshot, checked live against each vendor's own page rather than aggregated from a roundup, with the intent to re-check it on a regular cadence and let the history accrue on this same URL.
It is not a benchmark of model quality, and it is not a claim about which model is cheapest for a given task; a model's real cost depends on how many tokens it needs to finish the work, not only its per-token rate (see the FAQ below). It is a dated, sourced record of what each vendor's own page says right now.
Price per million tokens, across models
This chart is from the first edition (2026-07-23) and shows its 11 models at July prices. It has not been redrawn for this edition, so use the table below for the 2026-09-26 prices. Bars are sorted by output price and log-scaled so the $0.30 and $180 rows both stay readable; the dollar labels are the real, linear prices.
Bars are log-scaled for legibility across a 600x price range; read the printed dollar figures, not relative bar length, for the actual ratio between two models.
Every price, with its source
| Model | Vendor | Input $/M | Output $/M | Note | Source, as of 2026-09-26 |
|---|---|---|---|---|---|
| GPT-5.5 Pro | OpenAI | $30.00 | $180.00 | Flagship reasoning tier, prompts under 272K tokens. Unchanged since 2026-07-23 | developers.openai.com/api/docs/pricing |
| Claude Fable 5.1 | Anthropic | $10.00 | $50.00 | New since 2026-07-23. Cache reads $0.25 | platform.claude.com/docs/en/about-claude/pricing |
| Claude Fable 5 | Anthropic | $10.00 | $50.00 | Unchanged since 2026-07-23 | platform.claude.com/docs/en/about-claude/pricing |
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | New since 2026-07-23. Cached input $1.00 | developers.openai.com/api/docs/pricing |
| Claude Opus 4.8 | Anthropic | $5.00 | $25.00 | Unchanged since 2026-07-23 | platform.claude.com/docs/en/about-claude/pricing |
| Claude Opus 5.5 | Anthropic | $4.00 | $20.00 | New since 2026-07-23. Released Sep 22, 2026 | platform.claude.com/docs/en/about-claude/pricing |
| GPT-5.6 Sol | OpenAI | $4.00 | $20.00 | Promotional price, available at least through Nov 21, 2026 per OpenAI. Was $5.00 / $30.00 on 2026-07-23 | developers.openai.com/api/docs/pricing |
| GPT-5.6 Terra | OpenAI | $2.00 | $12.00 | Was $2.50 / $15.00 on 2026-07-23 | developers.openai.com/api/docs/pricing |
| Gemini 3.1 Pro Preview | $2.00 | $12.00 | <=200k-token tier; rises to $4.00 / $18.00 past 200k. Unchanged since 2026-07-23 | ai.google.dev/gemini-api/docs/pricing | |
| Claude Sonnet 5 | Anthropic | $2.00 | $10.00 | Launch intro price is now the standard price; Anthropic says the planned rise to $3.00 / $15.00 will not occur | platform.claude.com/docs/en/about-claude/pricing |
| GPT-6 Sol | OpenAI | $2.00 | $10.00 | New since 2026-07-23. Released Sep 22, 2026 | developers.openai.com/api/docs/pricing |
| Claude Haiku 4.5 | Anthropic | $1.00 | $5.00 | Unchanged since 2026-07-23 | platform.claude.com/docs/en/about-claude/pricing |
| Gemini 3.6 Flash | $0.75 | $3.75 | Through Dec 31, 2026, then $1.50 / $7.50 from Jan 1, 2027. Was $1.50 / $7.50 on 2026-07-23 | ai.google.dev/gemini-api/docs/pricing | |
| Gemini 3.8 Flash | $0.75 | $3.75 | New since 2026-07-23. Through Dec 31, 2026, then $1.50 / $7.50 from Jan 1, 2027 | ai.google.dev/gemini-api/docs/pricing | |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | Unchanged since 2026-07-23 | ai.google.dev/gemini-api/docs/pricing | |
| GPT-5.6 Luna | OpenAI | $0.20 | $1.20 | Was $1.00 / $6.00 on 2026-07-23 | developers.openai.com/api/docs/pricing |
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | New since 2026-07-23. Released Sep 22, 2026 | developers.openai.com/api/docs/pricing |
Sorted by output price, highest first. Sakana Fugu was checked for the first edition (2026-07-23) and left out because it had no published per-token rate card then; it was not re-checked for this edition. See how Fugu billing actually works for what is confirmed.
Frequently asked questions
How current is this price index?
Every price on this page was re-checked against the vendor's own live pricing page on 2026-09-26. The first edition was checked on 2026-07-23, and each row's note says whether the price changed since then. AI model prices change without much notice, so treat this as a dated snapshot, not a live feed. Click the source link next to any row to see the vendor's current number.
Why is Sakana Fugu not on this list?
This index only covers models priced on the Anthropic, OpenAI and Google pricing pages. When the first edition was checked on 2026-07-23, Sakana had confirmed a billing model (subscription plus usage-based billing with per-request cost reporting) but not an input/output rate card, so it was left out. It was not re-checked for this edition. See our separate breakdown of how Fugu billing works for what is and is not confirmed.
Why do input and output prices differ so much for the same model?
Generating a token (output) takes meaningfully more compute than reading one (input), so every vendor on this page prices output higher, typically 5x to 6x the input rate. The ratio is fairly consistent within a vendor's own model family; it is the absolute price level that varies by vendor and by model tier.
Is a cheaper per-token price always a cheaper task?
No. A more expensive model can still finish a task for less money if it uses meaningfully fewer tokens to reach a correct answer, and a cheaper model can cost more in practice if it needs several retries or a longer reasoning trace. Per-token price is one input to a real cost comparison, not the whole answer; our AI vs Human Employee Cost calculator models the actual task-level tradeoff.
Will this page track price changes over time?
Yes. This is the second edition. Every row carries a note comparing today's price with the first edition of 2026-07-23: in two months OpenAI cut GPT-5.6 Sol, Terra and Luna, Google cut Gemini 3.6 Flash until the end of 2026, and Anthropic kept Claude Sonnet 5 at its launch price. Six models are new since the first edition.
Refreshed on a regular cadence
The next snapshot will report real change against this baseline, on this same page.
No new URL for the next check-in. The series stays comparable, and this edition's numbers stay attached to the same page once the next one lands.
Picking the wrong model is more expensive than this whole page.
Our AI Strategy & Fractional CAIO practice benchmarks model spend against your actual workload, not list price, so you are not overpaying for capability you do not use.
Book a free consultation