Which AI video generators can an AI agent call in November 2026? A 10-second Pika clip is $0.20

$0.20. That is 10 seconds of Pika H3 Flex at 768p, from the per-second price in Pika’s public catalog on November 12, 2026.

Google’s own pricing page prints the next number a harness can paste into an eval without doing the multiplication: about $0.70 for 10 seconds of gemini-omni-1.2-flash at 720p. Runway Dev’s Gen-4.5 row is 12 credits a second at $0.01 a credit, which is $1.20 for 10 seconds once you multiply. Sora is not a fourth column. The Videos API shut down on September 24, 2026.

FreeGenVideo is the browser generator people are already opening. It is not in the price table, because the homepage fetched the same day publishes no API. The rest of the video category is the same kind of question: a listing is a product page until its own docs publish a tool an agent can call. Runway is the one in this set that prints an MCP URL. How a model calls a tool is tool use. The wire it plugs into is the Model Context Protocol. The six-month MCP calendar is MCP, the next 6 months.

Which generators publish an API an agent can call?

Pika, MiniMax, Runway, Kling, Luma’s Agents API, and Gemini Omni Flash do. Sora does not. Veo’s Gemini API price is not on the pricing page. FreeGenVideo’s API is not on the homepage.

Pika’s docs say the catalog is the price, and that there is no free generation tier. The catalog above was fetched with no API key. MiniMax’s legacy Hailuo 2.3 row is a different unit: $0.56 for a 768P clip of 10 seconds, printed as a clip price, not as a rate this note multiplied. The H3 series is the per-second table.

Kling stays in yuan. Runway Dev also sells Kling under its own credit rates. The kling_4 row at 720p, with no input video, prints 12.56 credits a second through November 30, 2026 UTC, and a list price of 15.7 credits a second. Runway’s worked example of the rounding is a 5-second 720p clip with no input video at 64 credits during the promotion: round the list charge up, take 20 percent off, and round up again. List prices apply from December 1, 2026 UTC. That 64-credit clip is Runway’s price, not Kling’s ¥1.0. Do not add the two bills.

What does the 10-second bill assume?

It assumes a text prompt, no input video, and a duration the same page allows. Change any of those and the invoice moves.

Pika H3 Flex at 480p is $0.01 a second, so 10 seconds computes to $0.10. At 768p the catalog says $0.02. The operation text says the clip can run from 3 to 15 seconds and is billed per output second. There is no minimum in that price object. Pika 2.5 at 720p is $0.04 a second both on the plain 720p tier and on the tier marked duration_s 5. The 1080p object has $0.06 and, when duration_s is 5, $0.09. This note does not pick a 10-second 1080p price out of those two.

MiniMax-H3 at 768P is $0.08 a second of output. Ten seconds computes to $0.80. At 2K the rate is $0.13, and 10 seconds computes to $1.30. The first five images on a task are free, and a text-only call has no image line. Input video is billed again, per second, at the output resolution. Audio input is free. H3-Lite at 768P is $0.04 a second, $0.40 for 10 seconds, and it does not offer 2K. The rate-limit page puts Video Generation V2, model MiniMax-H3, at 300 requests a minute and 30 tasks in flight. The older Hailuo series row is 20 requests a minute.

Gemini Omni 1.2 Flash is the model to name in the tool schema if you want Google’s 10-second example. Video output is $12.00 per million tokens. The footnote says 5,792 tokens per second of 720p video, about $0.07 a second, about $0.70 for 10 seconds, about $1.04 for 15 seconds. Free tier is “Not available.” Output runs from 3 to 15 seconds, at 24 fps, up to 4K. The input token limit is 131,072. That limit is for the prompt and the reference, not a place to paste the mp4 back into the next turn.

gemini-omni-1.1-flash is still a separate price on November 12. Video output is $17.50 per million tokens, about $0.10 a second at the same 5,792 tokens. The page says that on December 3, 2026 those requests route to 1.2 and bill at 1.2 rates. A harness that hard-codes 1.1 is buying the dearer second until that date, and then it is buying 1.2 under the old name.

Vertex Veo does not join this 10-second column. The model page lists 4, 6, or 8 seconds. Reference-image video is 8 seconds only. The price table says $0.40 per count for Veo 3.1 video plus audio at 720p or 1080p, and the Veo section never says “per second.” Eight times $0.40 would be a guess. This note does not print it. The Gemini API’s Veo 3.1 guide still describes those 4, 6, and 8 second lengths. The Gemini API pricing page, fetched the same hour, has no Veo row. A tool description that invents a Veo dollar rate is ahead of the page Google is billing from.

Luma’s current call is not the Ray 2 snippet. The Agents API video guide, read November 12, 2026, generates with ray-3.2. Duration is 5s, 10s, or omitted, and omitted means 5 seconds. A 10-second request is rejected if you also set HDR, a start frame, or an end frame. Standard 720p text-to-video is 100 credits for 5 seconds and 300 credits for 10 seconds. That 10-second total is three times the 5-second total, not two. The pricing page says the dollar value of a credit depends on the plan or contract, and that video rates can change ahead of general availability. The usage guide’s default list rate is $0.003 a credit, and it says to use that rate only when overage_rate is null and no contract sets another one. 300 × $0.003 = $0.90, computed here, and labeled as the list rate. A contract can make the same 300 credits a different dollar bill. The rate-limit page does not print a universal RPM. It says the allowance is on the dashboard and in X-RateLimit-Limit. The 30 RPM timeline on that page is an illustration.

The older Dream Machine page still shows POST https://api.lumalabs.ai/dream-machine/v1/generations for ray-2 and ray-flash-2, with a 5-second example, and a banner that points current guides at the Agents API. That page has no dollar price. Its rate-limit sibling, for the Build tier, lists Ray (Video) at 10 concurrent generations and 20 create requests a minute, and a $5,000 monthly usage cap. The page says it was updated 4 months ago. It does not name ray-3.2.

Which of these can the harness reach through MCP?

Runway’s, with a URL. The page is runway.com/mcp. It says to paste https://mcp.runwayml.com/mcp and sign in with a Runway app account. No API key. MCP uses your Runway credits, just like the Runway app, and the cost depends on the model and the resolution.

The Dev API is a different document. There, a credit is $0.01 and Gen-4.5 is 12 credits a second. The MCP page does not say those credits are the Dev API credits, and it does not say an app plan with unlimited generations makes an MCP call free. The help-center article linked from the page returned HTTP 403 on November 12, 2026, so this note does not quote it. A harness that adds the MCP server and also calls api.dev.runwayml.com should log which balance moved. Guessing that they are one pool, or two, is ahead of the pages.

Luma’s usage guide says spend in the Luma app, or over MCP, is not in GET /v1/usage. It does not print an MCP server URL. No official video MCP URL turned up on the Google, OpenAI, Kling, Pika, MiniMax, or FreeGenVideo pages read for this note. A docs site that offers “copy as MCP” for its own documentation is a documentation tool. It does not generate the video.

What should a harness wire?

Wire one REST call whose duration you set and whose price you can assert. Put the per-second rate in the tool description. Keep the mp4 out of the context window. Return a URL, the duration, and the cost.

For a 10-second clip with a published dollar rate, the cheapest row in this set at a 720p-class resolution is Pika H3 Flex at 768p, $0.20 computed. MiniMax-H3-Lite at 768P is $0.40 computed. The rate-limit page names MiniMax-H3 at 300 requests a minute and 30 tasks in flight. It does not name H3-Lite or H3-Highspeed, so those two do not inherit that cap on the evidence of this page. Luma Ray 3.2 at standard 720p is $0.90 only if you are allowed to use the $0.003 list rate, and the pricing page says those video rates can change before general availability. Gemini Omni 1.2 is the row to pick when you want Google’s own 10-second example, about $0.70, and you can live with no free tier. Runway Gen-4.5 is $1.20 computed on the Dev API. Use the MCP URL when the team already pays for Runway app credits, and log that balance separately from the Dev API until Runway says they are the same.

Call gemini-omni-1.2-flash, not gemini-omni-1.1-flash, if the budget is the $0.70 clip. The 1.1 id is about $0.10 a second until December 3. After that date the old id bills as 1.2, so a tool list that mentions both is one model with two names and, until December 3, two prices.

Do not leave sora-2 in the tool list. The deprecations page names September 24, 2026, and leaves the replacement blank. A retry loop on a removed model is a failed eval that looks like a timeout.

Do not budget a 10-second Vertex Veo clip. Ask for 4, 6, or 8 seconds, and do not multiply $0.40 until Google says what a count is. Do not point a tool call at FreeGenVideo. One video a day, resetting at 00:00 UTC, with no sign-up, is a guest allowance. The homepage fetched November 12 has no API in the text, and the sitemap has no API path.

Kling belongs in the tool list when the invoice can be in yuan. Kling 4.0 with no input video is ¥1.0 a second at 720p, the duration field accepts 3 through 30, and the default when you omit it is 10 seconds. The eval has to read output_duration. A prompt that says “make it longer” and returns 15 seconds is a ¥15 charge on that rate, not the ¥10 you wrote in the tool description. Kling sets no QPS limit. The cap is parallel tasks on the resource pack, and the error is code 1303. That is a queueing problem, not a price problem.

The eval is three assertions. The returned duration equals the duration you sent. The billed amount equals that duration times the rate in the tool description, plus any input-video seconds the price page adds. The run stops when the provider returns a shutdown, a missing price, or a count whose unit you cannot name. A pretty clip at the wrong length is a failed task. The token bill for the model that decided to call the tool is a separate meter. That one is the agent cost calculator. It does not include these per-second video rates.

What is still not a number?

A Luma contract rate, a universal Agents API RPM, a maximum length for Runway Gen-4.5, a Pika request cap, the meaning of a Vertex count, a dollar price for Dream Machine ray-2, and a per-second price for FreeGenVideo. Those seven are blank on the pages read November 12, 2026. The priced rows above are the ones a tool description can assert.

The video category has more browser products than this table. This note did not promote any of them to a tool call. Submit a listing if the product is missing, or claim the page that is already there. A listing with an API still needs the price page linked. A slogan that says “API access” is not a rate.