Agent Exchange Media · pricing

Pricing

Every route is paid per call in USDC on Base with x402; there is nothing else to buy. Prices are the amounts in the 402 challenges and in the manifest, which are generated from the same table.

Our price against what Workers AI charges us

"Workers AI cost" is Cloudflare's list price for the model at a typical input and at the route's limit, from developers.cloudflare.com/workers-ai/platform/pricing/ (page dated 2026-09-17, fetched 2026-09-30 23:17 UTC; $0.011 per 1,000 neurons; the first 10,000 neurons a day are free). Tokens are estimated at 4 characters each. Margin is at the route's limit; typical calls are far below it.

Route and limits Price Workers AI cost (typical, max) Margin at max Basis

POST /image
prompt <= 1,000 characters; steps 1-8; seed 0-4294967295; one image per call $0.02 $0.000634
max $0.0011 95% FLUX 1024x1024 (4 tiles) at 4 steps; max at 8 steps

POST /caption
image <= 5 MB (JPEG, PNG, WebP, GIF); image_url fetched in 8 s from a public https host; prompt <= 500 characters $0.005 not published
max not published n/a LLaVA has no published unit price (beta, billed in neurons); the Llama 3.2 Vision fallback would cost about $0.0003 per image

POST /speech
text <= 2,000 characters; lang one of en, es, fr, zh, ja, ko (default en) $0.01 $0.000120
max $0.000500 95% MeloTTS $0.0002 per audio minute; 2,000 characters is about 2.5 minutes; the Aura-1 fallback costs $0.015 per 1k characters, so it is used only under 600 characters ($0.009)

POST /transcribe
audio <= 10 MB by URL, <= 4.5 MB inline (6 MB request cap); fetched in 8 s from a public https host $0.01 $0.000500
max $0.0050 50% Whisper $0.0005 per audio minute; 10 MB of MP3 at 128 kbps is about 10 minutes

POST /translate
text <= 5,000 characters; language codes ISO 639-1 (en, es, fr, de, ja, ...) $0.002 $0.000103
max $0.000855 57% M2M100 $0.342 per M tokens in and out; 5,000 characters in and about as many out

POST /summarize
text 20 to 20,000 characters; max_words 20-400 $0.002 $0.000198
max $0.0010 49% Llama 3.1 8B fp8 $0.152/M in, $0.287/M out; 20,000 characters in, <= 900 tokens out

POST /embed
1-32 texts, each <= 2,000 characters $0.001 $0.000006
max $0.000192 81% bge-m3 $0.012 per M tokens; 32 texts of 2,000 characters

POST /chat
<= 20 messages, <= 12,000 characters in total; max_tokens <= 1024; roles system, user, assistant $0.005 $0.000130
max $0.000750 85% 8B: $0.152/M in, $0.287/M out; 12,000 characters in, 1,024 tokens out

POST /classify
text <= 5,000 characters (<= 2,000 with labels); 2-10 labels of <= 60 characters $0.001 $0.000003
max $0.000099 90% DistilBERT $0.026 per M tokens; the zero-shot path on Llama 8B is the max

POST /ocr
image <= 5 MB; image_url fetched in 8 s from a public https host $0.005 not published
max not published n/a LLaVA has no published unit price (beta, billed in neurons); the Llama 3.2 Vision fallback would cost about $0.0003 per image

POST /detect-language
text <= 20,000 characters; 'und' when nothing matches $0.001 $0
max $0 100% no model; Worker CPU only

POST /batch
1-100 items; ops embed (<=32 texts of 2,000 chars), translate (<=5,000 chars), summarize (<=20,000 chars), classify (<=5,000 chars); total text <= 500,000 characters; a single bad item fails the whole batch before payment $1.00 $0.0021
max $0.1018 90% 100 summarize items at their limit is the worst case

POST /image-set
1-20 prompts of <= 1,000 characters; steps 1-4; always 20 images (prompts cycle); response about 2-4 MB of base64 $5.00 $0.0127
max $0.0127 100% 20 FLUX images at 4 steps

POST /chat (model llama-3.3-70b)
<= 20 messages, <= 12,000 characters in total; max_tokens <= 1024; roles system, user, assistant $0.02 $0.000687
max $0.0032 84% 70B: $0.293/M in, $2.253/M out; 12,000 characters in, 1,024 tokens out

Route and limits	Price	Workers AI cost (typical, max)	Margin at max	Basis
`POST /image` prompt <= 1,000 characters; steps 1-8; seed 0-4294967295; one image per call	$0.02	$0.000634 max $0.0011	95%	FLUX 1024x1024 (4 tiles) at 4 steps; max at 8 steps
`POST /caption` image <= 5 MB (JPEG, PNG, WebP, GIF); image_url fetched in 8 s from a public https host; prompt <= 500 characters	$0.005	not published max not published	n/a	LLaVA has no published unit price (beta, billed in neurons); the Llama 3.2 Vision fallback would cost about $0.0003 per image
`POST /speech` text <= 2,000 characters; lang one of en, es, fr, zh, ja, ko (default en)	$0.01	$0.000120 max $0.000500	95%	MeloTTS $0.0002 per audio minute; 2,000 characters is about 2.5 minutes; the Aura-1 fallback costs $0.015 per 1k characters, so it is used only under 600 characters ($0.009)
`POST /transcribe` audio <= 10 MB by URL, <= 4.5 MB inline (6 MB request cap); fetched in 8 s from a public https host	$0.01	$0.000500 max $0.0050	50%	Whisper $0.0005 per audio minute; 10 MB of MP3 at 128 kbps is about 10 minutes
`POST /translate` text <= 5,000 characters; language codes ISO 639-1 (en, es, fr, de, ja, ...)	$0.002	$0.000103 max $0.000855	57%	M2M100 $0.342 per M tokens in and out; 5,000 characters in and about as many out
`POST /summarize` text 20 to 20,000 characters; max_words 20-400	$0.002	$0.000198 max $0.0010	49%	Llama 3.1 8B fp8 $0.152/M in, $0.287/M out; 20,000 characters in, <= 900 tokens out
`POST /embed` 1-32 texts, each <= 2,000 characters	$0.001	$0.000006 max $0.000192	81%	bge-m3 $0.012 per M tokens; 32 texts of 2,000 characters
`POST /chat` <= 20 messages, <= 12,000 characters in total; max_tokens <= 1024; roles system, user, assistant	$0.005	$0.000130 max $0.000750	85%	8B: $0.152/M in, $0.287/M out; 12,000 characters in, 1,024 tokens out
`POST /classify` text <= 5,000 characters (<= 2,000 with labels); 2-10 labels of <= 60 characters	$0.001	$0.000003 max $0.000099	90%	DistilBERT $0.026 per M tokens; the zero-shot path on Llama 8B is the max
`POST /ocr` image <= 5 MB; image_url fetched in 8 s from a public https host	$0.005	not published max not published	n/a	LLaVA has no published unit price (beta, billed in neurons); the Llama 3.2 Vision fallback would cost about $0.0003 per image
`POST /detect-language` text <= 20,000 characters; 'und' when nothing matches	$0.001	$0 max $0	100%	no model; Worker CPU only
`POST /batch` 1-100 items; ops embed (<=32 texts of 2,000 chars), translate (<=5,000 chars), summarize (<=20,000 chars), classify (<=5,000 chars); total text <= 500,000 characters; a single bad item fails the whole batch before payment	$1.00	$0.0021 max $0.1018	90%	100 summarize items at their limit is the worst case
`POST /image-set` 1-20 prompts of <= 1,000 characters; steps 1-4; always 20 images (prompts cycle); response about 2-4 MB of base64	$5.00	$0.0127 max $0.0127	100%	20 FLUX images at 4 steps
`POST /chat (model llama-3.3-70b)` <= 20 messages, <= 12,000 characters in total; max_tokens <= 1024; roles system, user, assistant	$0.02	$0.000687 max $0.0032	84%	70B: $0.293/M in, $2.253/M out; 12,000 characters in, 1,024 tokens out

The two whole-dollar units

POST /batch is $1.00 for up to 100 items across embed, translate, summarize and classify in one request, each with the single-route limits (total text 500,000 characters). Bought one by one the same 100 items would cost between $0.10 and $0.20; the unit exists for buyers who settle in whole dollars and want one receipt. POST /image-set is $5.00 for twenty 1024x1024 JPEGs from up to 20 prompts at up to 4 FLUX.1 [schnell] steps each: schnell-grade quality, prompts cycled to fill 20 slots, no upscaling, no SDXL fallback. Twenty single /image calls would cost $0.40; the set is priced as one dollar-denominated job, not as a discount.

Fallbacks and substitutions

When the primary model errors, the routes fall back as listed on the routes page (FLUX → SDXL base; LLaVA → Llama 3.2 11B Vision; MeloTTS → Aura-1 under 600 characters; Whisper turbo → Whisper under 3 MB; M2M100 → Llama 3.1 8B; bge-m3 → bge-base; Llama 8B → Llama 3.2 3B); the response names the model that actually ran in model and sets fallback: true. The 70B chat tier has no fallback: it fails rather than downgrade. @cf/meta/llama-3.1-8b-instruct was not in the catalog on 2026-09-30 and is replaced by @cf/meta/llama-3.1-8b-instruct-fp8; @cf/facebook/bart-large-cnn was not in the catalog on 2026-09-30 and is replaced by @cf/meta/llama-3.1-8b-instruct-fp8 with a summarization prompt; @cf/unum/uform-gen2-qwen-500m was not in the catalog on 2026-09-30 and is replaced by @cf/meta/llama-3.2-11b-vision-instruct as the caption fallback.

How payment works

Send the request; the answer is 402 Payment Required with an accepts list (scheme exact, network Base eip155:8453, asset USDC 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913, payTo 0xc91cE6291eDC0713ec753BAFBA002506ffb2b95c, maxTimeoutSeconds 300). Sign the EIP-3009 authorization for the amount, resend the identical request with X-PAYMENT, and the facilitator verifies and settles it on chain before the model runs; the response carries PAYMENT-RESPONSE. Settlement is final. A malformed request with a payment header is rejected (400 or 413) before anything is settled. If the model fails after settlement the call answers 502 with code: model_failed and the same X-PAYMENT is accepted once more within 10 minutes without a new payment.

Free

/samples (a request and response per route), /live.json, /health, /openapi.json, /.well-known/x402, the agent card, /a2a (A2A JSON-RPC; free skills pricing and samples, paid skills return the x402 terms), /llms.txt and /mcp (tool listing; paid tools return payment instructions). Unpaid requests to paid routes are limited to 20 per minute per address; paid requests are not limited.