Skip to content

Model prices

The per-token retail price of every model Else can run, and which model each tier uses by default.

Every figure on this page is a retail price per million tokens — what you are charged, not an estimate. Prices are the same whether a model runs in a chat or in a handed-off task.

Tokens are how models measure text: roughly 4 characters, or about ¾ of a word. A short question and answer is a few hundred tokens; a long report with source material can be tens of thousands.

You do not have to choose a model. The four options in the composer’s engine menu are Auto, Fast, Smart and Max — Auto picks for each request, and the other three each run one model by default:

Tier What it is for Default model Input / 1M tokens Output / 1M tokens
Fast Cheapest acceptable answer gemini-3.6-flash $1.80 $9.00
Smart Balanced quality, long context gpt-5.6-terra $3.00 $18.00
Max Deepest reasoning claude-opus-5 $6.00 $30.00

You can point any tier at a different model in Settings → How it worksSet an engine per speed tier, and you can pick a model for a single request from Pick an exact engine in the composer’s engine menu. See Engines and tiers for how that interacts with what you are charged.

All 60 models below are selectable today. Output tokens cost more than input tokens on nearly every model, which is why a long answer costs more than a long question.

Images in is a measurement, not a guess: yes means image input was measured working, no means it was measured not working, and untested means we have not measured it and will not claim it either way.

Repeated context is the rate for input tokens the model has already been sent in this conversation. When a request re-sends context, those tokens bill at that lower rate instead of the full input rate. It is applied by the biller automatically — there is nothing to switch on, and a dash means this model has no separate rate for it.

What we know about it links to every page that covers that model — all of them, not just one — and each link says what kind of fact is on the other end. Reads images and images untested are our own probe results; PDFs (maker’s claim) is the maker’s published specification and nothing we measured. A dash means no page covers this model beyond the row you are reading.

Model Context Images in Input / 1M tokens Output / 1M tokens Repeated context / 1M What we know about it
granite-4.0-h-micro 131K no $0.02 $0.13
llama-3.2-1b-instruct 60K no $0.03 $0.24
gpt-5-nano 128K untested $0.06 $0.48 $0.01 images untested
llama-3.2-3b-instruct 80K no $0.06 $0.40
qwen3-30b-a3b-fp8 33K no $0.06 $0.40
glm-4.7-flash 131K no $0.07 $0.48
gemini-2.5-flash-lite 1M yes $0.12 $0.48 $0.01 reads images · takes a whole document · PDFs (maker’s claim)
gemma-4-26b-a4b-it 256K no $0.12 $0.36
gpt-4.1-nano 1M untested $0.12 $0.48 $0.03 takes a whole document · images untested
gpt-4o-mini 128K yes $0.18 $0.72 $0.09 reads images
llama-3.1-8b-instruct-fp8 32K no $0.18 $0.34
gpt-5.4-nano 128K untested $0.24 $1.50 $0.02 images untested
gpt-oss-20b 128K no $0.24 $0.36
gemini-3.1-flash-lite 1M yes $0.30 $1.80 $0.04 reads images · takes a whole document
gpt-5-mini 128K untested $0.30 $2.40 $0.03 images untested
llama-4-scout-17b-16e-instruct 131K yes $0.32 $1.02 reads images
llama-3.3-70b-instruct-fp8-fast 24K no $0.35 $2.70
gemini-2.5-flash 1M untested $0.36 $3.00 $0.09 takes a whole document · PDFs (maker’s claim) · images untested
gemini-3.5-flash-lite 1M untested $0.36 $3.00 $0.04 takes a whole document · PDFs (maker’s claim) · images untested
m3 1M no $0.36 $1.44 $0.07 takes a whole document
gpt-oss-120b 128K no $0.42 $0.90
mistral-small-3.1-24b-instruct 128K yes $0.42 $0.67 reads images
gpt-4.1-mini 1M untested $0.48 $1.92 $0.12 takes a whole document · images untested
gemini-3-flash 1M no $0.60 $3.60 $0.06 takes a whole document
nemotron-3-120b-a12b 256K no $0.60 $1.80
qwen3.5-397b-a17b yes $0.72 $4.32 reads images
qwq-32b 24K no $0.79 $1.20
gpt-5.4-mini 128K yes $0.90 $5.40 $0.09 reads images
kimi-k2.6 262K no $1.14 $4.80 $0.19
kimi-k2.7-code 262K no $1.14 $4.80 $0.23
claude-haiku-4.5 200K yes $1.20 $6.00 $0.12 reads images · PDFs (maker’s claim)
gpt-5.6-luna 1.1M untested $1.20 $7.20 $0.12 takes a whole document · images untested
o4-mini 200K untested $1.32 $5.28 $0.33 images untested
qwen3-max untested $1.44 $7.20 images untested
gemini-2.5-pro 1M yes $1.50 $12.00 $0.15 reads images · takes a whole document · PDFs (maker’s claim)
gpt-5 128K untested $1.50 $12.00 $0.15 images untested
gpt-5.1 128K yes $1.50 $12.00 $0.15 reads images
grok-4.3 1M yes $1.50 $3.00 $0.24 reads images · takes a whole document
glm-5.2 262K no $1.68 $5.28 $0.31
gemini-3.5-flash 1M yes $1.80 $10.80 $0.18 reads images · takes a whole document
gemini-3.6-flash 1M untested $1.80 $9.00 $0.18 takes a whole document · PDFs (maker’s claim) · images untested
deepseek-v4-pro 131K untested $2.09 $4.18 $0.17 images untested
claude-sonnet-5 1M untested $2.40 $12.00 $0.24 takes a whole document · images untested
gpt-4.1 1M yes $2.40 $9.60 $0.60 reads images · takes a whole document
grok-4.20-0309-non-reasoning 2M no $2.40 $7.20 $0.24 takes a whole document
grok-4.20-0309-reasoning 2M no $2.40 $7.20 $0.24 takes a whole document
gpt-4o 128K yes $3.00 $12.00 $1.50 reads images
gpt-5.4 1M yes $3.00 $18.00 $0.30 reads images · takes a whole document
gpt-5.6-terra 1.1M untested $3.00 $18.00 $0.30 takes a whole document · images untested
claude-sonnet-4.5 200K yes $3.60 $18.00 $0.36 reads images · PDFs (maker’s claim)
claude-sonnet-4.6 200K untested $3.60 $18.00 $0.36 PDFs (maker’s claim) · images untested
kimi-k3 1M untested $3.60 $18.00 $0.36 takes a whole document · images untested
claude-opus-4.5 200K untested $6.00 $30.00 $0.60 PDFs (maker’s claim) · images untested
claude-opus-4.6 1M untested $6.00 $30.00 $0.60 takes a whole document · PDFs (maker’s claim) · images untested
claude-opus-4.7 1M untested $6.00 $30.00 $0.60 takes a whole document · PDFs (maker’s claim) · images untested
claude-opus-4.8 1M untested $6.00 $30.00 $0.60 takes a whole document · images untested
claude-opus-5 1M yes $6.00 $30.00 $0.60 reads images · takes a whole document
gpt-5.5 1M yes $6.00 $36.00 reads images · takes a whole document
gpt-5.6-sol 1.1M untested $6.00 $36.00 $0.60 takes a whole document · images untested
claude-fable-5 1M untested $12.00 $60.00 $1.20 takes a whole document · images untested

A model appears here only while it has real capacity behind it, so this list changes. If a model you picked stops being available, the tier it was bound to falls back to its default and tells you it did.

Most models charge one rate however long your prompt is. One model on this list does not: past a threshold, every input and output token in that request bills at a higher rate — not just the tokens past the line.

Model Runs as When it applies Input / 1M tokens Output / 1M tokens
m3 Exact engine pick over 512,000 input tokens $1.44 $5.76

Below the threshold you pay the rate in the main table above. The receipt shows what you were actually charged either way.

Generated images are priced per image, by size, not by tokens:

Model Size Price per image
gemini-2.5-flash-image default $0.0585
gemini-3-pro-image 1K $0.201
gemini-3-pro-image 2K $0.201
gemini-3-pro-image 4K $0.36
gemini-3.1-flash-image 0.5K $0.0675
gemini-3.1-flash-image 1K $0.1005
gemini-3.1-flash-image 2K $0.1515
gemini-3.1-flash-image 4K $0.2265
gemini-3.1-flash-lite-image 1K $0.0504
gpt-image-1.5 1024x1024 $0.051
gpt-image-1.5 1024x1536 $0.075
gpt-image-1.5 1536x1024 $0.075
gpt-image-1 1024x1024 $0.063
gpt-image-1 1024x1536 $0.0945
gpt-image-1 1536x1024 $0.0945
gpt-image-2 1024x1024 $0.0795
gpt-image-2 1024x1536 $0.0615
gpt-image-2 1536x1024 $0.0615

Three other things can appear on a receipt. Each is metered per unit of real use, and the amount charged is always itemized on the receipt itself:

Item Metered by Price
Reading a web page per page rendered $0.12 per page
Running code in the secure workspace per minute of run time $0.0048 per minute
Transcribing audio you send per minute of audio itemized on the receipt

Two notes on those, because a price is only useful if you know what it is measuring.

Reading a web page is a flat rate per page, not per second spent on it. A page that takes a long time to load costs the same as one that loads instantly.

Run time is measured while your code is actually executing, not for the whole time a run is open. A run that spends most of its time waiting on a model is not accruing workspace time.

Audio transcription is the one item without a rate on this page. It is charged at the metered rate we are charged for it, with nothing added — so the figure is not ours to publish, and quoting it here would be publishing someone else’s rate card. The amount is always itemized on your receipt, and the unit is per minute of audio.

See How billing works for what triggers each of these, and Reading a receipt for where they appear.

Every price on this page is generated directly from the same pricing code that bills your runs — it is not transcribed by hand, and it cannot disagree with what you are charged.

Prices last changed 2026-07-29 (source ced6a4f). If that date looks old, that is because nothing has moved, not because the page was forgotten.