Answers
Straight answers about generating with AI
One question per page, answered in the first paragraph. Model limits, error codes, refunds, agent access and prices — measured against the live platform, not copied from a press release.
Maintained by BananaBanana TeamLast updated
When something goes wrong
Why was my AI image blocked by the safety filter?
Because one of two independent filters fired. The first checks your prompt before anything is drawn and can be relaxed; the second inspects the finished pixels and cannot be turned off by anyone. A relaxed filter only affects the first. Since every run draws a different picture, a borderline prompt that fails the second stage often passes on a retry — and a blocked generation is refunded, so the retry costs only time.
Read the answerWhat does 'The service is currently experiencing high load' mean in Veo?
It is a capacity error from Google, raised inside a generation that already started — your prompt is fine. It hits hardest on the most expensive configurations, especially 4K with a last frame. The job is restarted automatically up to four times on the same key; if all attempts fail you get the error code HIGH_LOAD and a full refund.
Read the answerDo I get my money back when an AI generation fails?
Yes, automatically and in full. The charge happens when a generation starts and is reversed the moment it is marked failed — upstream errors, capacity errors and content-filter rejections all refund without a support ticket. A batch that is only partly rejected is refunded proportionally. There is nothing to claim and no waiting period.
Read the answerWhy must an edited AI video keep the exact length of the source?
Because the length is validated after the clip is generated, not before. If the result does not match the requested duration to within a tenth of a second, the whole job is failed upstream — minutes of compute wasted. So an edit inherits the source's length instead of taking it from the form, and sources are trimmed to whole seconds before they are sent.
Read the answerWhat the models actually do
How long can a Gemini Omni Flash video be?
Three to ten seconds per clip, set to the exact second, with a synchronised audio track that cannot be turned off. Asking for 3 seconds really does produce a 3.0-second file and bills a third of a 10-second one. Since the 1.1 generation (September 2026) a finished clip can be extended: each pass appends 3–10 seconds to the end, and the chain stops at 40 seconds. Edits still keep the source length.
Read the answerHow many reference images can I attach to one prompt?
It depends on the model. Nano Banana Pro takes 11 reference images, split as up to 6 objects and 5 characters; the other image models take up to 14. Veo 3.1 accepts 3 reference images, and they are mutually exclusive with a first frame. Gemini Omni Flash allows 10 images per request and, unusually, lets a start frame and references coexist.
Read the answerCan I cancel an AI video generation that is already running?
No, and not because of a missing button. Google's video APIs expose no working cancel operation — the request answers 501 Unimplemented, and delete is rejected outright. Once a generation starts it runs to completion or fails. A failure is refunded in full, so a mistake costs you the wait rather than the money.
Read the answerDo seeds actually reproduce the same AI image?
Partly. On the image models a seed is a real control: reusing one with the same prompt and settings keeps composition close, which is how you iterate on wording without losing the frame. On Gemini Omni Flash video it is accepted but ineffective — two runs with an identical seed and prompt came out visibly different in our tests. For consistency across generations, reference images beat seeds.
Read the answerWhich AI model generates true 4K images, and is it worth it?
Nano Banana Pro is the 4K model — $0.20 an image, and the one to use when the picture carries legible text. Nano Banana 2 also reaches 4096 px for $0.13, with 512 px drafts from $0.03. Nano Banana 2 Lite is 1K only, at a flat $0.03. Resolution buys detail, not composition: draft at low resolution, commit at high.
Read the answerWhat it costs
How much does one AI-generated video actually cost?
Between $0.10 and $4.40 per clip, depending on model, resolution, length and whether the audio track is generated. A 4-second silent 720p clip on Veo 3.1 Lite is $0.10; an 8-second 4K clip with audio on Veo 3.1 is $4.40. Gemini Omni Flash bills per second — $0.10 a second at 720p, so $0.30 to $1.00 per clip. You pay per generation, not per month.
Read the answerHow much does AI text-to-speech cost per script?
One cent per started 200 characters. A 1,000-character script — roughly 150 spoken words — costs $0.05, and a 60-second voice-over usually lands under $0.10. Thirty voices are available, single speaker or a two-person dialogue, returned as a 24 kHz WAV in the same response. You are charged only if the audio is produced.
Read the answerIs there a free trial, and what can I do with it?
Every new account starts with $0.20 of real balance, no card and no trial clock. That is six images on Nano Banana 2 Lite, or two 4-second silent clips on Veo 3.1 Lite, or a mix. It is ordinary balance rather than trial credit: it does not expire, and whatever you generate with it is yours to use.
Read the answerCan I get my unused balance refunded?
Yes. You can request a refund of the unused part of your balance at any time while the account is active — there is no deadline, and the balance itself never expires. We return what you topped up and have not spent, minus promotional bonus credits. Generations that fail or are blocked by the safety classifier are refunded to your balance automatically, with no request needed.
Read the answerConnecting an agent
Why does an MCP server list tools fine but return 401 when I call one?
That is lazy authorisation, and it is deliberate. Catalogue methods — initialize, ping, tools/list — answer without credentials so directories can read the tool list. Only a paid action, tools/call, requires a token, and it answers 401 with a WWW-Authenticate header. That 401 is exactly what triggers the OAuth sign-in flow in Claude; a 200 there would leave the header ignored and the client stuck.
Read the answerWhat is the correct MCP server URL, /mcp or /api/mcp?
Use https://bananabanana.pro/api/mcp. That is the JSON-RPC endpoint. The bare /mcp path is the human documentation page, and posting JSON-RPC to it historically returned HTML — the single most common misconfiguration. Requests to /mcp are now rewritten to the canonical route, but configure clients with /api/mcp so nothing depends on that.
Read the answerCan an AI agent generate images without an account or API key?
Yes, over x402. An agent holding a wallet posts a request, receives HTTP 402 with the exact price for those arguments, pays in USDC on Base and repeats the request with a payment header. No sign-up, no API key, no prepaid balance. Images and speech settle only after the file exists, so a failure costs nothing.
Read the answerYour account and your output
Who you are buying from
Who operates BananaBanana?
BananaBanana is operated by Rari LTD, registered at 11 Leo Kiacheli St, Tbilisi, Tbilisi 0108, Georgia. It is an independent company, not a Google product or partner, and it sells metered access to Google's image, video and speech models. The same details appear in the terms of service, the privacy policy and the site footer, and support runs through one inbox that answers billing questions within three business days.
Read the answerIs BananaBanana affiliated with Google?
No. Rari LTD, the company behind BananaBanana, is an independent company with no affiliation, sponsorship or endorsement from Google. It buys access to Google's production models — the Gemini image models, Veo 3.1, Gemini Omni Flash and Gemini Flash TTS — and sells generation on a prepaid balance. Gemini, Veo and Google are trademarks of Google LLC. You need no Google account, and your billing relationship is with Rari LTD alone.
Read the answerWhere do my prompts and uploaded images go?
Your prompt, any reference images or source video, and the generation parameters are sent to Google's models through the access channels we operate. Your email address, account id and balance stay with us — nothing identifying you travels with the request. Results come back to your private account and live there for 30 days before automatic deletion. Prompts and output are never used to train models.
Read the answerIs BananaBanana safe to use for client work?
For most client work, yes — with two limits worth knowing. There is no SLA: an outage or an upstream capacity problem can interrupt generation, though an interrupted generation is never billed. And generation requests leave our servers to reach Google's models, so material you are contractually required to keep confidential deserves a decision before you upload it rather than after.
Read the answerWhy one question per page?
Because that is the shape an answer engine can lift. A question buried in the middle of a 3,000-word guide is hard to quote and easy to misread; a page that opens with the answer, states the numbers, and then explains itself can be cited whole.
Every page here carries something that is not on a vendor datasheet: a raw upstream error code, a limit we measured against the API, a price taken from the same table the checkout uses. When a fact changes, the date under the answer changes with it.
Missing a question?
The list grows from real support conversations and agent sessions. If your question is not here, the MCP documentation and the price table cover the mechanical detail, and support answers at [email protected].
Answer it by trying it
New accounts start with $0.20 of balance — no card, nothing expires.