Changelog

A language menu, and a shorter header

Pick English or Chinese from the header. If your browser prefers the other language, the site offers it once.

• The language menu lists each language in its own name and keeps you on the same page. The offer appears in your language, never redirects, and does not come back once you choose or dismiss it.
• The header keeps Video, Image and Text. All prices, price changes, sources, method, changelog and the agent guide moved to a Resources column in the footer.

A quieter home page, and questions on every family page

The home page reads as lists instead of grids of cards, and every model family page ends with the questions people ask about its price, answered from its own numbers.

• Today’s biggest price gaps is one ranked list: the model, the cheapest platform and price, the dearest, a bar between them, and the gap.
• The six platforms are one list with endpoints, prices, how many families each is cheapest in, and when it was last read. The separate “latest reads” panel is gone; latest price changes stand on their own.
• Each family page answers how much the model costs, which platform is cheapest, which platforms sell it, and how current the prices are, also as FAQ structured data for search.

Image and video families open on one card per version

An image or video family page now starts with the answer: for each version, its platforms at the same output, cheapest first, and how far apart they are.

• Each card shows a version, the output it is ranked at, the price gap, and a bar per platform. Green is the cheapest. Platforms priced in another size or unit are named in one line under the bars.
• The header is the title, one sentence and a row of every model. A model scrolls smoothly to its card, which is ringed for a moment so you can see where you landed.
• As a card scrolls into view it rises into place and its bars grow, so the cheapest stands out as you read. With reduced motion the cards are simply there.
• An -ext endpoint is marked with a small Ext tag. Hover it for what it means and the standard endpoint’s price; it is explained once under the cards.
• A version sold on one platform is a chip with its price, linking to its prices, instead of a table of its own. Text families keep their input and output table.

Each family lists its models

The home search opens a model a developer looks up by either name. Under a family heading, every real version is a link, newest first.

• The home search offers Nano Banana 【 Gemini Image 2.5 】, Hailuo 3 【 MiniMax H3 】, Kling O3 【 Kling 3.0 Omni 】, Nano Banana Pro, and Claude Sonnet. Each chip opens that model.
• Under a family heading, every real version is listed. A version sold on more than one platform scrolls to that version. A version on one platform opens its prices.
• On a Chinese page, Hailuo is 海螺 and Kling is 可灵 before the other name. The name inside 【 】 stays. The price library title and the family filter use the same wording.

Version titles use the names developers search

A version keeps its short name and shows the other name of the same model beside it. Searching either name opens that row. Distinct products stay on their own rows.

• The first Nano Banana is Nano Banana 【 Gemini Image 2.5 】. Nano Banana Pro, Nano Banana 2, Nano Banana 2.1 and Nano Banana Lite stay their own rows. Gemini Image 2.5, Nano Banana, and version 1 open the first row.
• Hailuo 3 is Hailuo 3 【 MiniMax H3 】. Hailuo 3, MiniMax H3 and MiniMax-H3 open it. Hailuo 3 Lite, Max and Highspeed stay their own rows. Kling O3 is Kling O3 【 Kling 3.0 Omni 】, and Kling O3 Pro is Kling O3 Pro 【 Kling 3.0 Omni Pro 】. Kling 3.0 Pro stays the other product.
• Wan 3 opens Wan 3.0. The row title stays Wan 3.0, and Wan 3.0 Prime stays its own row. fal’s Wan VACE 14B joins Wan 2.1 VACE. Wan 2.2 VACE Fun stays its own row. Wan 2.2 5B Fast is its own row beside Wan 2.2 Fast.
• LTX Video 0.9.8 Distilled and LTX Video 0.9.7 Distilled keep the second decimal. WaveSpeed’s v097 is LTX Video 0.9.7, without Distilled. The LTX 2.5 Pro endpoint whose page title said Fast is LTX Video 2.5 Pro. LTX Video 2.5 Fast stays the Fast price.

A compared price is the charge for making the output

The price a comparison quotes is the charge for producing that model’s output. A cheaper rate that needs an extra input, an add-on, or a tool sold beside the generator stays in the full price list and no longer sets the “from” price. The same model is one group however a platform spells it.

• Reference or uploaded video, v2v, video continuation, a re-render of a clip already paid for, a minimum, web search, an SVG written by another model, and a surcharge are not the quoted price. Layer split, prompt rewrite, upscale, and apps beside a generator are tools. An image or video model’s text-token rate is not its image or video price. An off-peak window does not win against the list price.
• Gemini 2.5 Flash Image, including fal’s gemini-25-flash-image, is Nano Banana. Juggernaut FLUX and RunDiffusion Photo FLUX are fine-tunes, not FLUX’s price. One price that covers several sizes matches each of those sizes. When any compared price states a resolution, audio or quality alone does not become the ranked spec.
• kie.ai uses the credit value when the shown dollars are that value rounded. Token rates are per million tokens. A count the page does not name, such as “$0.015 / 1000 UTF-8 bytes”, stays that amount. APIMart, OpenRouter, and Replicate name the unit and the charge; those rows appear after the next collection. A column headed “Official / Fal Price” is described as an official or fal price, since the column does not say which each figure is.

fal video prices by resolution, read from each model’s page

fal’s price list gives one price per video endpoint, usually the cheapest resolution’s, and sometimes in a billing unit rather than per second. The pricing note on each model’s page gives the rest, so fal video prices now have a row per resolution, audio option or clip length wherever the note fits the price list.

• Grok Imagine 1.5 was listed at $0.01 a second, which is its billing unit; it is $0.08 a second at 480p, $0.14 at 720p and $0.25 at 1080p. Pixverse prices per “video segment” ($0.05) are now clip prices: a 5-second clip is $0.15 at 360p, $0.20 at 720p and $0.40 at 1080p. Pixverse C1 and V6 list each resolution with and without audio.
• Hailuo 3, Wan, Kling, Gemini Omni, LTX and other models priced per second now show the rate at each resolution, not only the default one. Offers with dates count only while they run: H3 Max is $0.035 a second at 480p, $0.056 at 768p and $0.112 at 1080p until November 15, then $0.05, $0.08 and $0.16.
• Where a note and the price list disagree, as on FLUX 3’s video endpoints, Luma Reframe, Pixverse v5.6 and some Kling and Wan endpoints, fal’s listed price is kept as before and the collector flags the note for a check.
• A price that only gained detail, such as fal’s “$0.05 a second” becoming “480p: $0.05 a second”, is no longer logged in Price changes as one price removed and another added.

Cleaner model pages: one table per version, fewer false “cheapest” marks

Model pages such as Hailuo, Kling and Seedance explain the comparison once, and only versions sold at more than one resolution get a table of their own. A price that doesn’t say which resolution it is for is still shown, but is no longer marked cheapest, and more prices now say what they are for.

• kie.ai’s Hailuo 2.3 “Standard” is Hailuo 2.3 Fast and its “Pro” is Hailuo 2.3, so they now sit with those versions. “Pro” alone is no longer a version of its own: kie.ai’s Hailuo 02 Pro is Hailuo 02 at 1080p.
• APIMart’s MiniMax-H3-Regeneration and fal’s H3 Max Insert and 3D to Video are tools, not Hailuo 3 itself, so they are no longer compared with it. Regeneration had set Hailuo 3’s cheapest 2K price; 3D to Video is billed per second plus tokens for the video and images sent in.
• fal prices billed per “unit” are read from the model’s page: Hailuo 2.3 is $0.28 for a 6-second 768p clip, so it now compares per second with WaveSpeed and Replicate. Replicate’s clip lengths read “6s”; WaveSpeed and OpenRouter prices name the resolution when the model has only one; OpenRouter’s minimum per generation, quoted in cents, is no longer read as dollars per second.
• A version’s table has a row per resolution. Prices with a video, image or audio going in are in a table of their own, side by side and never ranked. A version sold on one platform gets the same table, instead of a list of raw prices.
• A version with one output is all in its row of the comparison table, and its name opens all its prices. Tool endpoints such as lip sync, avatars, effects and extend are listed under “Not compared”.
• When no two platforms sell a version at the same stated resolution, its row says so and nothing is marked cheapest; before, a 720p price could be ranked against one whose resolution was not stated.
• WaveSpeed prices read in words, such as “480p · 5s: $0.125 (list $0.25)”, instead of the quote the collector sent; hover to see the text as stored.

Pick one version in All prices

Search a model in All prices, such as seedance, kling, veo, nano banana or claude, and a Version row lists its versions, each with its number of prices: Seedance 2.5, 2.0, 2.0 Mini, 1.5 Pro, 1.0 Pro Fast… Pick one to see only its prices, on every platform.

• Versions are grouped as on each model’s page, however each platform writes the id: APIMart’s seedance-1-0-pro-fast, fal’s seedance/v1/pro/fast and Replicate’s seedance-1-pro-fast are all 1.0 Pro Fast. Video and image models are split by version (Veo 3.1 Fast, Kling O3 Pro), text models by model (Claude Opus 5.5, with batch prices listed after the rest).
• The row appears when at least half of a search’s prices belong to one model, or when you pick one in the Model row. Platform, route and unit counts follow the version you pick. Models with many versions, such as Kling and GPT, show the newest 16 first; “more” shows the rest.
• “No version named” holds the prices whose model id and name give no version: an original model with no number, or avatar, lip-sync and effect endpoints.
• On a model’s page, the link to every price of a version now opens exactly that version’s prices in All prices; before, it searched the version’s name, which could also find its neighbours.
• Each version has its own address, such as /prices?q=seedance&version=2-0-mini, to bookmark or share.

Every price says who serves it

The same model can reach you by very different routes, even on one platform: the maker’s own API, or a channel the platform itself says is not the official one, such as APIMart’s “Ext” models. Every price now shows its route, as the platform states it and never guessed from the price, and All prices can show one route at a time.

• Five routes. Official: the platform says the request goes to the maker’s own API, as OpenRouter does for Veo through Google Vertex. Official partner: the platform resells the maker’s API as a partner, like fal’s partner APIs. Platform-hosted: the platform runs the model on its own GPUs. Third-party: the platform says another company serves it, or that it is not the official channel. Not disclosed: the platform doesn’t say.
• In All prices, the comparison pages and the Video, Image and Text pages, a route mark sits under each platform name; hover it for the platform’s own words and the basis. A model’s page states its route and basis, with a link to the archived page it was read from.
• All prices has a Route filter: Official / partner, Platform-hosted, Third-party and Not disclosed, each with its number of prices. Route words in a search set it rather than match model names: “official” or “官方” keeps official and partner routes, “ext”, “third-party” or “逆向” keeps third parties. Try “veo 3.1 fast official”.
• Where routes come from: OpenRouter’s list of providers for each model (with several, the least certain one counts), fal’s and Replicate’s model pages, and APIMart’s model pages, which mark models “official channel” or “Ext”; APIMart models it doesn’t mark are Not disclosed. kie.ai and WaveSpeed don’t say, so all their prices show Not disclosed.
• On Replicate, “Official” alone doesn’t mean the maker’s API. Official models that run on Replicate’s GPUs, or whose page links open weights on Hugging Face or ModelScope, count as Platform-hosted; only those that run on CPU with no published weights, where the maker’s API does the generating, count as Official partner.
• Every row from the JSON API and the MCP tools has route and route_label. /api/v1/search, /api/v1/cheapest and the search_prices and cheapest tools take route=official,partner to keep the maker’s own API only; with a route set, listings not checked yet are left out.

Cleaner APIMart names, and sharper search

Page notes such as “Prices are estimates” no longer end up in APIMart model names, and where the page gives a model id, the model goes by it: “GPT-4o Mini” is gpt-4o-mini, “DALL-E 3” is dall-e-3. No price changed.

• Midjourney, Suno and FlowMusic on APIMart keep only each action’s own price, such as midjourney@imagine, instead of repeating them under a summary card.
• Searching All prices for an APIMart id such as veo3.1-fast-ext no longer also finds video “extend” endpoints.
• Searching All prices ignores words like “price”, “价格” or “多少钱”; before, “veo 3.1 价格” found nothing.
• LTX Retake endpoints, which redo part of an existing video, count as video editing rather than text to video.

Filter all prices by model

Pick Video, Image or Text on All prices and a Model row appears: Seedance, Veo, Kling and the rest, each with its number of prices.

• The home page’s latest price changes now show only cuts and rises: how much, which price (input, output, per second) and from what to what. New and delisted models stay on the changes page.
• On the changes page, hovering “Flagged” says why a change was flagged.

Connect your agent from the home page

“Copy for agent” on the home page copies one line that has Claude Code, Codex, Cursor or any coding agent add our MCP server by itself. The live board can show what the agent sees: the real answer of the MCP tool for the same model.

• The menu has each client’s own command or install link: Claude Code, Codex, Gemini CLI, Add to Cursor, Install in VS Code, and the steps for Claude.ai and ChatGPT.
• New /setup.md: setup steps written for the agent that reads them.
• The agents page now gives each client its own config. The single mcp.json we showed before was wrong for Claude Code, VS Code and Gemini CLI.
• Search treats a short whole number as a version: “Claude Opus 5” no longer returns Opus 5.5, and “Veo 3” no longer returns Veo 3.1.
• On phones the live board fits the model on show, with no empty space under short lists, and stays on it until you tap another.

Price gaps count exact prices only

The gap on the home page, the live board and the comparison pages now compares exact prices for the same job only. It used to include approximate routes, which made Veo 3.1 Fast look 8.6× apart when the exact prices differ 1.9×.

• The home page now opens with “AI API Pricing Comparison” and the widest exact gap between platforms.

Clearer page titles for search

Page titles, headings and descriptions now lead with what each page compares, and fit in a search result without being cut off.

• Model pages: sections are now named for the model, such as “Claude API pricing by platform” and “Veo 3.1 API price by resolution”.
• Single-endpoint price pages stay where they are, but are kept out of search engines for now; the sitemap lists the main pages only.

A new logo

The site has a new icon: cobalt-blue flowing curves for many API routes meeting in one choice. It is in the header, footer and browser tab.

Prices with a video going in are their own row

Several video models charge a different per-second rate when you upload a video (Seedance 2.0 on kie.ai: $0.095/s at 480p, $0.057/s with video input). Those rates are now never mixed with the plain price: each version’s resolution table has its own “480p · Video input” rows, platform by platform. They are not ranked, because platforms bill them differently: kie.ai charges the rate on input plus output seconds, Replicate on output seconds only.

• More wordings are recognised as video input: APIMart’s “720P+Video”, “720P-refvideo” and “480P & uploaded videos”, and OpenRouter’s “video continuation”. Before, APIMart’s Kling O3 “720P+Video” and SkyReels V4 “refvideo” prices counted as ordinary 720p prices.

Exact WaveSpeed prices, real fal token prices, and rows ranked at one resolution

WaveSpeed video and image prices were a "from" figure that could be far off (Veo 3.1 Fast showed $0.15 a run; a default run costs $1.20). They now come from WaveSpeed’s own price calculator, per resolution, with and without audio. fal models priced at a placeholder "$1 per unit" now show the token prices written on their model pages.

• WaveSpeed: Veo 3.1 Fast is $0.10/s without audio and $0.15/s with it; Kling 3.0 Pro $0.112/s and $0.168/s; Seedance 2.0 $0.216/s at 720p. Prices are WaveSpeed’s public sale price; the list price is kept in the evidence.
• fal: GPT Image 2.6 and similar models show input, cached and output token prices (for example $8 / $2 / $30 per million image tokens) instead of "$1 per unit".
• Each row of the model table names the resolution, and for video the audio, it is ranked at (“Ranked at 720p · No audio”). A platform without that output is shown in grey, not ranked. A bare price beside a “+Sound” one on the same platform counts as silent; a platform that never mentions audio is still compared, marked “audio not stated”.
• Tool endpoints (upscalers, lip-sync, background removal, try-on, video extend and the like) are no longer counted as a model version, so they can’t win a version’s cheapest price.
• Replicate prices that depend on a yes/no input now say so: “without audio · standard” instead of “false · standard”.
• Dated text snapshots are merged with their model: gpt-5-2025-08-07 is GPT-5, claude-4.5-sonnet is Claude Sonnet 4.5, -latest aliases join their model.

Only the same model is compared: a row per model, a column per platform

Family and task pages used to rank every endpoint on one list, so GPT-6 Luna sat next to gpt-oss-120b and Seedance 2.0 next to Seedance 1.5 Pro. Now each model or version gets its own row and each platform its own column; green marks the cheapest platform for that model only. Inside a video or image version, a second table lines up the same resolution across platforms.

• Text models are matched across platforms by model id: openai/gpt-6.1-sol on OpenRouter, gpt-6-1-sol on kie.ai and GPT-6.1 Sol on APIMart are one row, with input and output prices shown separately. Batch pricing is its own row.
• Sub-models are no longer merged into their parent: GPT Image Flare and Sunburst, FLUX Klein, Kling Motion Control, Wan Animate and VACE, LTX Distilled, Recraft Vector, Hailuo Highspeed and others are separate rows.
• Each video and image version has a resolution-by-resolution table (480p, 720p, 1080p, 4K…) when two or more platforms sell it; the full price list sits under it.
• A model sold on one platform only is listed apart, with nothing to rank it against. Prices in another unit (tokens, a platform’s own unit) are shown in grey and never marked cheapest.
• Cheaper "-ext" endpoints (APIMart’s gpt-image-2.6-ext, veo3.1-fast-ext…) are compared with the rest but labelled “stability may vary”, with the platform’s standard-endpoint price beside them when it has one.
• The same tables are in the Markdown pages and the agent API.

Prices by model version, and per-clip prices ranked per second

Family pages now list every version on sale, such as Seedance 2.5, 2.0, 2.0 Fast, 2.0 Mini and 1.5 Pro, each with its own cheapest price. Prices billed per clip or per minute are converted to per second and ranked with the rest, so a cheap per-clip offer is no longer hidden in a separate table.

• A “By version” table at the top of each video and image family page, newest first, then one section per version. The family’s headline price is its cheapest named version, never a lip-sync or avatar endpoint.
• Per-clip prices count when the clip length is known: from the label (“5s”, “10.0s”), or Veo 3.x’s 8-second clip, marked “default length”. APIMart’s Veo 3.1 Lite at $0.07 per clip is $0.00875/s, the cheapest Veo price we have.
• Task pages such as Text to video API rank every exact price on one per-second scale; units that can’t be converted are listed below it.
• Fixed: kie.ai’s “no video input” prices for Seedance 2.0 were treated as input charges and left out, while the cheaper “with video input” rate, which bills the input video on top, was shown as the price. Add-ons like “Get 1080P Video” no longer count as a generation price.
• For agents: the cheapest API and MCP tool take a version (for example “2.0 Mini”) and return the per-second price and clip length with each row.

Full breadcrumbs, and page titles that say what each page is

Every inner page now opens with a full breadcrumb trail, also on phones, and each page’s title names the one thing it prices, such as “Kling API pricing” or “Lip sync API pricing”.

• Breadcrumbs wrap instead of collapsing to “…” on narrow screens. Model pages now sit under their category and model family: API Pricing › Video › Veo › google/veo-3.1.
• Task pages use the words people search for: Text to video API, Image to video API, Lip sync API, Image upscaler API.
• Platform pages are titled by their pricing, for example “OpenRouter pricing”.
• Fixed: phone screens could scroll sideways because of the footer links.
• Fixed: clicking a step in the breadcrumb, such as “Video”, stayed on the same page.

Prices your agent can read, and a new home page

Every page now has a Markdown version, prices are in a free JSON API, and an MCP server lets Claude or Cursor ask directly. The home page now starts with search and walks through tasks, price gaps, platforms and recent activity.

• “Copy for agent” on every model, family, task and category page: copies the page as compact Markdown, or opens it in ChatGPT or Claude.
• Add .md to any address, or send Accept: text/markdown. /llms.txt lists the whole site for language models.
• JSON API at /api/v1: search, cheapest by task or family, and one model on one platform. No key.
• MCP server at /mcp with four tools. One command adds it to Claude Code.
• Home page: a search box with examples, browse by task, today’s six widest gaps between exact prices, the six platforms side by side, latest reads and changes, how it works, and common questions.
• Collector status moved from the home page to Sources.

Price tables sorted by spec, and this changelog

Each price now says what it is for in a few tags, instead of a link to the raw page it came from.

• Tags under every price: task (links to the task page), resolution, duration, audio, inputs, quality tier, and for text models cache reads, cache writes, web search and long-prompt rates.
• Words that only repeated the unit (“per seconds”, “output image”, “Default parameters, one run”) are gone. The platform’s exact wording is still beside every price.
• The “Evidence” column is removed from price tables and change logs. Each model page ends with one short “Where these prices come from” list: the address, when it was last read, and an archived copy.
• This page: every change to the site, in plain words.

Compare by task: text to video, image editing, upscaling…

Nine task pages group every endpoint by what it does, whoever made the model.

• Video: text to video, image to video, video editing, lip sync and avatars, upscaling. Image: text to image, image editing, upscaling, background removal.
• Each page shows the cheapest model families for the task, the cheapest price on each platform, and every price underneath.
• Task chips on the home page and category pages, and links in the footer.

Archived pages you can actually read

An archived copy used to be raw JSON or HTML. It now opens as a page.

• When it was read, the HTTP status, size and fingerprint at the top, then the models that cite it.
• JSON is shown as one card per record, web pages as text with tables kept as rows, and price fields are pinned first. One search box filters everything.

Video, image and text hubs, and 47 model family pages

Browse by category or by model family (Kling, Veo, Flux, GPT…), each with prices from every platform.

• The home page now explains what the site does in one screen: what each category covers, the cheapest families, and four ways to use it.
• A softer dotted texture behind the headline replaces the grid.
• “From” prices no longer pick free tiers, input-image charges or near-zero rows, which had shown things like $0 per image.
• Kling O1 and O3 no longer show the OpenAI icon.
• A sitemap for search engines, with English and Chinese versions of every page.

A faint grid behind the headline

A fine blueprint grid behind the price board. Replaced by dots the next day.

A live price board on the home page

The models with the widest price gaps, each platform’s price as a bar, and how many times more the dearest costs.

• It switches model every 6 seconds and stops when you hover or pick a tab.

The headline leads with the price gap

The title now carries the widest gap found, and a line shows how many prices are being tracked right now.

New colours

An indigo accent and cool greys, in light and dark mode.

Maker icons next to every model

Kling, Veo, Flux, GPT and 70 more makers are recognisable at a glance, across tables, search and history.

• Covers about 3,300 of 4,150 models; the rest are makers without a public icon.

API Pricing launches

One place to compare what AI APIs cost on OpenRouter, kie.ai, fal, WaveSpeed, Replicate and APIMart, in US dollars.

• Side-by-side comparisons for popular models, worked out for a concrete job.
• Search across every collected price, by model, platform and unit.
• Prices are read automatically every day and never overwritten: each model has its full price history.
• A price-change log with big moves flagged.
• English and Chinese.