How we test

We ran six test sessions between 2026-10-21 and 2026-10-27, all on one vendor, Higgsfield. Together they charged 197.28 credits net (our charge) for 76 runs: 61 images and 15 video clips. This page explains the method behind every “we tested” claim on the site.

What we test and what we don’t

As of 2026-10-27 we test image and video models inside Higgsfield only. We submit every run through Higgsfield’s MCP connection (Model Context Protocol) and pay with credits from our Plus plan account. Each session has a credit cap set before it starts.

Checked 2026-10-27 by editorial review.

Image models so far: Soul 2.0, Soul Cinema, Nano Banana Pro, Nano Banana 2, Seedream 5.0 Lite and GPT Image 2. Video models: Seedance 2.0, Kling 3.0 with its Motion Control mode, Kling 2.6, Hailuo 02, Hailuo 2.3, and Veo 3.1, which returned no clip.

Every other vendor number on the site is a labeled vendor claim. A claim we could not confirm reads “not verified as of” with a date; a run we did not do reads “not tested”.

Each result is usually one charged sample per model, not a ranking; a different seed can give a different picture.

Our fictional test character

Every person in our tests since 2026-10-21 is fictional, and most tests use one subject: our fictional, AI-generated test character. Soul 2.0 generated it from text in our first run, with no reference; of 4 variants (0.48 credits) we chose variant 3 as the canonical image.

Later runs use the canonical image as reference or start frame, or start from an image made with it. In our old photo restoration test, the “old photo” is synthetic: Seedream 5.0 Lite turned the canonical image into a scratched sepia print (H9-dmg, 2026-10-27). The one exception is the eight Soul 2.0 preset runs of 2026-10-27. A Soul style preset cannot be combined with a reference image, so those prompts describe the character in text.

Scenes with two people add a second fictional adult, a man who reads 33 to 38, made with Soul 2.0 from text on 2026-10-27; the same page explains the choice. In our couple photo test each model got both characters as references, in the same order. Our hug video test started from one of those photos.

We use generated people so that no example needs a real person’s face. Neither character is ever an author, reviewer or user.

Three subjects show no person, and we generated each from text. For our pet dance video test on 2026-10-27, Nano Banana Pro made a fictional dog (2 credits, 1 try), the start frame for both the Kling 3.0 and the Hailuo 02 clip. The driving clips for our motion control test show mannequins that Kling 3.0 generated on 2026-10-26. H6a, a Kling 2.6 rainy-street clip (2026-10-27) logged for a page not yet published, has no person. No real person, pet or third-party clip went into any of them.

Same input, same prompt, every try counted: the protocol

Each test since 2026-10-21 follows six rules.

1. Every person in a test is one of our two fictional characters; the runs without a reference image are named above.
2. Every model in a comparison gets the identical prompt, word for word. Image comparisons also share the aspect ratio, and video comparisons share one start frame.
3. Settings are logged per run. Nano Banana Pro ran at 1k, except one 4k run in a 1k-versus-4k test (2026-10-26). GPT Image 2 ran at 1k on its low quality setting; Seedream 5.0 Lite ran at default quality.
4. We count every submission. A failed or rejected one stays in the record, and so does a refusal, with any refund next to it.
5. A refusal is a result, and we publish it. We do not reword the prompt or swap the image to get past it.
6. We log the outcome before judging it. On the page, “by eye” notes and logged verdicts are kept apart.

In the first session the aspect ratio was 3:4 on 11 of the 16 charged runs, 9:16 for the two video start frames, 16:9 for the anime film pair and 4:3 for the Ghostface scene. Eleven of the 15 clips are vertical; four, H6a among them, are 16:9.

One start frame, H0, serves our video comparison on the character: a full-body shot that Nano Banana Pro made from the canonical image on 2026-10-26 (2 credits, 1 try). Seedance 2.0 got it that day; Hailuo 2.3 and Veo 3.1 got the same frame and prompt, word for word, on 2026-10-27.

The first session logged 0 content refusals in 17 submissions, the Ghostface costume prompt included: Nano Banana Pro accepted it on the first try.

The first refusals came on 2026-10-27. Veo 3.1 refused H0 three times: twice on Veo 3.1 Fast, then once on Veo 3.1 Lite. Each job ended with status nsfw and no error text. We kept the prompt and the frame unchanged, so the refusal is the published result.

A failure is logged apart from a refusal. On 2026-10-26 two Kling 3.0 Motion Control jobs ended with status failed and no error text. The third attempt completed after we changed only the mannequin driving clip, and the log records that change.

What the first session ran, model by model

The table counts charged runs in our log (2026-10-21). One run is one job that Higgsfield charged for.

Seedream 5.0 Lite ran in four same-prompt pairs against Nano Banana Pro: action figure, polaroid, anime film style and candy cane. GPT Image 2 ran once, in the caricature pair.

Checked 2026-10-21 by editorial review.

What every test log records

The table shows every field of the first log, filled from one real run: G9b, Seedream 5.0 Lite in the candy cane pair.

The first log keeps tries inside the outcome sentence and has no queue time or watermark status. From 2026-10-26 each line adds a tries count, the billed charge, the file’s pixel size and the time from submit to finished file. Video lines add duration, frame rate, audio and watermark.

Our 15 clips run 5.03 to 5.875 seconds. Eight are logged with no visible watermark; the other seven have no watermark entry. Only H6a has sound, requested for that test. Nobody listened to it: ffprobe shows a stereo AAC track, and its spectrogram reads as rain-like noise with no music or speech.

Request and result can differ; the log keeps both. All three Hailuo clips, including our Hailuo 2.3 clip, came back at 5.875 seconds for a requested 6. Three clips planned on Kling 2.5 Turbo, absent from the MCP catalog on 2026-10-27, ran on Kling 3.0 (one) or Kling 2.6 (two). The restored portrait (H9-anim) asked for 9:16 and came back 3:4.

Every job we requested as Nano Banana Pro came back labeled nano_banana_2 in the job record, 25 of 25 by 2026-10-27. The billing record, logged from 2026-10-26, names them “Nano Banana Pro”. On 2026-10-21 the other three models matched the request; Soul 2.0 appears as text2image_soul_v2. We report the label as it stands and draw no conclusion from it.

How credits become cost per image or clip

As of 2026-10-21 we log credits first and convert to dollars second, because credits are the unit the vendor charges.

The worked example uses the Higgsfield pricing page and our first session’s 24.98 credits.

The dollar figures are our own calculation, rounded to the cent; the price per credit is rounded to four decimals. A 2026-10-27 server message names our plan “plus (monthly)”; monthly Plus costs $59 for 1,200 credits, about 25% more per credit, so the site’s annual-rate Plus figures are a lower bound.

The same method works per image: one Nano Banana Pro image, charged 2 credits on 2026-10-21, comes to about $0.08 on Plus and $0.14 on Starter.

Charged credits include billed runs that failed, such as the 2-credit Christmas polaroid that drifted. A rejected submission with no charge, like the 429, adds 0, and so does a refunded job, such as each Veo 3.1 refusal. Prices rotate, so current rates sit with Higgsfield credit costs per plan.

Checked 2026-10-21 by editorial review.

What we have not tested yet

Four things are missing as of 2026-10-27, and each has a reason.

• Genjutsu video. On 2026-10-21 we considered one Genjutsu preset, Chandelier Lounge Performance. Which social trend it reproduces is not verified as of 2026-10-21. We did not run it because its driving video shows real people: our output would carry their likeness and an unlicensed performance. The rules are on the content policy page. On 2026-10-27 Genjutsu’s replace mode quoted 42 credits for 5 seconds at 720p, above that test’s cap, so we skipped it. Genjutsu: not tested.
• Face swap and Character Swap. Neither was exposed through the MCP connection on 2026-10-21 or 2026-10-27. Both: not tested.
• Recast app. The MCP connection did not expose it on 2026-10-27 and routes “recast” requests to Kling 3.0 Motion Control. We ran that route once with the second character (10 credits) and do not present it as a Recast result. Recast app: not tested.
• Other vendors. We hold no account with Kling, OpenArt or any other vendor as of 2026-10-27; our Kling clips ran inside Higgsfield only. Their output quality: not tested.

Re-test schedule and how changes are logged

Since 2026-10-21 we re-verify prices and free limits at least every 30 days, and sooner when we see a change, because promotions rotate. We re-run effect test tables every 60 to 90 days. Each change goes into the correction log with date, page, old value, new value and source.

A correction to a test log is a new dated entry, and the original line stays as written. On 2026-10-27 the log called one Soul 2.0 image, H13-b4, the only head-to-feet result. A full-size re-check that day found one foot cut by the bottom edge, and the correction now sits below the unchanged row.

How some links can earn us money is on the affiliate disclosure page. As of 2026-10-21 no commission program is active for us, and a commission would not change what we test.

Sources checked

• Higgsfield pricing: plan prices on annual and monthly billing, monthly credits, prices without VAT. Checked 2026-10-20.
• Our test logs, Higgsfield Plus account: sessions of 2026-10-21, 2026-10-26 and 2026-10-27 (four sessions), including the “plus (monthly)” server message in session 4; fields under what every test log records.
• Calculations: dollars = plan price / monthly credits x credits charged.