Sixteen AI video tools on seven published criteria

Every tool here is rated one to five on the same seven criteria, given only from what the vendor publishes. Two tools leave cost transparency unrated because neither puts a usable rate in public. The spread is widest on assembly, which most tools say nothing about. As of 2026-09-12.

Inclusion rule. Tools with a public English product page that addresses at least four of the seven criteria. A tool that publishes almost nothing cannot be rated on this rubric and is left out rather than rated low. Order. Descending by average of rated criteria, then alphabetical.

1One criterion at a time

Each criterion has its own page, because each one splits the field differently: some separate almost nothing, one is nearly empty, and one is sharply polarised. The averages above hide all of that.

• Script handling — twelve of sixteen tools publish nothing about text longer than a prompt
• Cast consistency — almost every vendor claims consistency; the ratings turn on whether a mechanism is described
• Shot control — seven of sixteen offer a prompt box and nothing more
• Picture quality — the flattest column on the rubric: nothing sits below three and six reach five
• Sound — two entries generate speech with the picture in named languages
• Assembly — the most polarised column here: six entries publish nothing about joining segments, three document it as a priced step with everything to do afterwards
• Cost transparency — two entries publish no usable rate at all and are left unrated

2Each tool on its own

Every row above has a page of its own: the same seven ratings, what each one rests on, and where that tool sits against the rest on its strongest and weakest columns.

• SceneMixer — the narrowest spread here: nothing below four, and fives on five of the seven
• LTX Studio — a tight band between three and five, with cost transparency left unrated
• Hailuo (MiniMax) — a spread of four points: strongest where generation happens, silent on everything around it
• invideo AI — strongest at the two ends of the job and mid-table where the pixels are made
• Morph Studio — the flattest profile in the table, with nothing at either extreme
• Veo / Flow (Google) — peaks on both generation columns, a floor on script handling
• CapCut — the assembly column at five and the two generation-adjacent columns at one
• Wan (Tongyi Wanxiang) — solid generation columns, nothing published about assembly or script handling
• Sora (OpenAI) — a peak on picture quality, a floor on script handling, and one unrated column
• Kling AI — top of the field on picture, bottom on both workflow columns
• PixVerse — nothing above four, nothing below one, and most of the row in the middle band
• Dreamina — one strong column and a long middle, with script handling on the floor
• Runway — a peak on picture and a floor on script handling, four points apart
• Vidu — a strong consistency column, a weak sound column, and nothing on assembly
• Luma Dream Machine — the widest spread in the table: a five and two ones on the same row
• Pika — the lowest average here, carried by a single strong column

3Where two tools look alike and are not

Twelve pairs in this table land within half a point of each other on average while disagreeing by two points or more on at least three criteria. Those are the pairs where the average is actively misleading, so each one has a page.

• CapCut and Luma Dream Machine — 4 of seven criteria apart.
• CapCut and Morph Studio — 4 of seven criteria apart.
• CapCut and Runway — 4 of seven criteria apart.
• CapCut and Veo / Flow (Google) — 4 of seven criteria apart.
• Hailuo (MiniMax) and invideo AI — 4 of seven criteria apart.
• invideo AI and Kling AI — 4 of seven criteria apart.
• CapCut and Hailuo (MiniMax) — 3 of seven criteria apart.
• CapCut and Kling AI — 3 of seven criteria apart.
• CapCut and Vidu — 3 of seven criteria apart.
• invideo AI and Morph Studio — 3 of seven criteria apart.
• invideo AI and Sora (OpenAI) — 3 of seven criteria apart.
• invideo AI and Veo / Flow (Google) — 3 of seven criteria apart.

4The averages hide more than they show

An average across seven criteria treats joining segments and generating a shot as equally important, which no production does. A team that never assembles anything can ignore the assembly column entirely, and for them the ordering above is close to meaningless. Read the column that matches the work.

That is also why the per-criterion numbers are published beside the average rather than behind it. The average is a way into the table, not a verdict, and the rows underneath are where the answer is.

5Where the field actually splits

Picture quality is the flattest column here: most vendors publish named models, resolutions and clip lengths, and most land in the same band. The columns that separate tools are the ones about the work around generation, which is where a tool either has a documented answer or has nothing.

Script handling is the sharpest of those. A majority of entries say nothing at all about text longer than a prompt, which is a real gap for anyone starting from a script rather than an idea.

6How this fits the ledger

The ledger records what a second costs on each channel that publishes a rate. These ratings record what each vendor documents about the rest of the work. A cheap second on a tool that documents none of the surrounding steps is not cheap once the surrounding steps are counted.

The two pages answer different questions and are kept apart on purpose. Cost transparency appears in both because it is the one criterion the ledger can verify directly: a rate that can be read in public is the same fact in either table.

Ratings are given against each vendor's own public pages, read 2026-09-12. They rate what is published, not output quality, and no vendor is consulted before a rating appears. How each criterion is defined is on how scoring works.