Which AI Video Model Is Best at the End of 2025?

TL;DR

Google's VEO 3.1 won both the interpretive K-pop squid dance test and the anime fruit duel test, with Sora 2 close behind and Hailuo 2.3 strong on detail but prone to anatomical morphing. LTX-2 finished last in both, though it is the only one whose weights are slated for an open-source release in early 2026. Hailuo 2.3 still has no built-in audio.

Transcript

What's up everybody? Welcome back to the Matt Vidpro AI YouTube channel. Today we are having an AI video model showdown at the end of 2025. We've got four models in the lineup. Today, yes, we do have a sponsor for today's brawl. It is Miniax aka Haleo AI. Yes, their model is going to be competing today, but as you know, I don't exactly like to go e... Read More

Key Insights

• VEO 3.1 won both scored tests in this comparison, taking the interpretive K-pop squid dance round and the reference-image anime duel round, judged on the best overall combination of prompt adherence, animation quality, and watchability.
• Hailuo 2.3 has no support for audio at all. Pairing it with Minimax audio through the Hailuo video creation agent gives some level of audio generation, and Minimax told the reviewer that future plans include upgrading the video model to include built-in audio.
• VEO 3 was the first AI video model to natively generate audio alongside video, including characters talking to each other and generated music, but VEO output is limited to 8 seconds of video.
• LTX-2's distinguishing feature is not quality but openness: it does native 1080p and its weights are supposedly going to be released open source in early 2026, after an original release window that slipped, making it a preview of what open-source video generation may offer.
• Anatomical morphing is Hailuo 2.3's main failure mode. In the dance test the model handled a backward step and head swoop by turning the dancer's back into the front of her body, and a retry showed instant morphing at the start followed by roughly 90 percent coherent footage.
• Reference images materially change results. The same fruit anime duel prompt run without a reference image made VEO 3.1 drop the fruit theme and generate characters resembling copyrighted anime, while the reference-image version stayed on concept.
• Animation style differs by model on anime prompts: Hailuo 2.3 reads more like Flash animation that moves whole characters rather than animating every new frame, and Sora 2 exhibits a similar stretch style while taking creative liberty with character designs.
• LTX-2 placed last in every test covered here. Its audio quality is much lower than the other three, its fighting movements ran so fast that the characters' arms were hard to see, and its body movement showed poor anatomical performance.

Questions & Answers

Q: Which AI video model won the end of 2025 comparison?

Google's VEO 3.1 won both of the tests scored in this comparison. It took the interpretive K-pop squid dance test, though the reviewer called it very close with Sora 2, and it also won the anime fruit duel test that used a reference image, credited with the best combination of the qualities being looked for. Hailuo 2.3 placed just behind the leaders in the dance test with better interpretive dancing but more anatomy errors, and ranked second in the anime duel ahead of Sora 2. LTX-2 finished last in both tests.

Q: Does Hailuo 2.3 generate audio?

No. Hailuo 2.3 has no support for audio at all, and every Hailuo clip in the comparison plays silent. There is a workaround: the Hailuo video creation agent can pair Hailuo 2.3 with Minimax audio, which effectively gives it some level of audio generation. Minimax also told the reviewer that future plans include upgrading the video model to include built-in audio, which the reviewer said he was happy to hear. By contrast, VEO 3.1, Sora 2, and LTX-2 all do native audio generation, though LTX-2's audio quality is described as much lower than the other three.

Q: What is LTX-2 and why include it if it loses every test?

LTX-2 is included because of what it represents rather than how it scores. It does native 1080p, it has native audio, and its actual model weights are supposedly going to be released open source, with the stated timing being early 2026 after an earlier planned release slipped. That makes it a preview of the best that can be expected from a potential open-source offering. In the tests themselves it placed last every time, with weak audio, poor anatomical performance, and fight movements so fast the arms were hard to make out.

Q: What settings and platforms were used to test each model?

Hailuo 2.3 was used on the Hailuo AI website at a resolution of 768p and 10 seconds long. Sora was used on the Sora website in landscape orientation with a duration of 10 seconds. LTX-2 was used in the LTX AI playground, always set to Pro, at a duration of 8 seconds in 1080p at 25 fps. Most VEO 3.1 clips were generated with the Gemini app, but a few toward the end were made through Hailuo AI's website, which integrates other AI video generators, because the reviewer ran out of credits on his Gemini Ultra plan. Every model received the same prompt.

Q: How did the models handle the interpretive K-pop squid dance prompt?

The prompt was a single dancer performing an interpretive K-pop inspired life of a squid dance with a slider cam and shallow depth of field, run with no reference image. VEO 3.1 produced hip-hop style pop dancing but invented an unprompted squid-themed jacket whose tentacles flip around realistically. Hailuo 2.3 gave the most genuinely interpretive dance with strong hair physics, then morphed the dancer's back into her front during a head swoop. Sora 2 showed no morphing but lower overall detail and fidelity than Hailuo 2.3. LTX-2 bent and twisted its subject in ways no human body moves.

Q: Why did VEO 3.1 fail the anime duel prompt without a reference image?

When the fruit themed anime duel prompt for lemon versus blueberry was rerun with no reference image, VEO 3.1 generated anime characters that resemble copyrighted anime characters rather than fruit themed ones. The reviewer noted they looked like Dragon Ball Z, apparently the same character in different forms including a Super Saiyan version. The fighting itself looked very cool, but the model totally missed the prompt. Hailuo 2.3 handled the same no-reference prompt better, returning a simple, shape-based animation of an actual lemon character duelling an actual blueberry character.

Q: How does Sora 2 compare with VEO 3.1?

Both do native audio generation and both scored highly. Sora 2 has great consistency and understanding, can be used through the Sora app to essentially create AI TikToks, and was seen by many at release as a step up over the original VEO 3. In the dance test it finished a very close second to VEO 3.1, showing no morphing but less detail and fidelity than Hailuo 2.3. In the reference-image anime duel it dropped to third behind Hailuo 2.3, because its animation was harder to read and mushier, and it took creative liberty with the character designs, including one shot with an extra finger.

Q: Was this comparison influenced by the sponsor?

The video is sponsored by Minimax, also known as Hailuo AI, and the sponsor's own model competes in it. The reviewer states directly that he does not go easy on models, that he shows all of the direct outputs, and that Minimax has no say on his opinion or on the clips featured in the testing. In practice the sponsor's model did not win either scored test: Hailuo 2.3 was placed behind VEO 3.1 in both, docked points for anatomical morphing in the dance test and for less compelling fight animation in the anime duel.

Q: What were the reference images in the tests generated with?

All reference images used in the testing were generated with Nano Banana Pro. The anime duel test used a Nano Banana Pro image of a lemon character and a banana character already clashing in battle, which the reviewer felt looked good and set the models up for success by establishing both characters before the video prompt ran. The prompt paired with it asked for a fruit themed anime duel, lemon versus banana, three clean hits, with smear frames. The description also notes that Nano Banana Pro is free on Hailuo AI as part of a promotion.

Summary & Key Takeaways

• Four AI video generators are put head to head at the end of 2025: Minimax's Hailuo 2.3, Google's VEO 3.1, Sora 2, and LTX-2. The same prompt goes to every model, sometimes with a shared reference image, and every raw output is shown. Hailuo sponsored the video but had no say over the clips or the verdicts.

• Test settings differ per platform: Hailuo 2.3 ran on the Hailuo AI site at 768p for 10 seconds, Sora ran on the Sora website in landscape at 10 seconds, and LTX-2 ran in the LTX AI playground on Pro at 8 seconds, 1080p, 25 fps. Most VEO 3.1 clips came from the Gemini app, with a few made through Hailuo AI's site after the Gemini Ultra credits ran out.

• In the K-pop squid dance test, VEO 3.1 took the win by inventing an unprompted squid-themed jacket with tentacles that flip realistically. Hailuo 2.3 produced the most genuinely interpretive dance and strong hair physics but morphed a dancer's back into her front. Sora 2 avoided morphing with slightly lower fidelity, and LTX-2 bent its subject in anatomically impossible ways.

• The anime duel tests used a lemon versus banana reference image generated with Nano Banana Pro, then a no-reference rerun with lemon versus blueberry. VEO 3.1 won the reference version, while its no-reference attempt abandoned the fruit theme entirely and produced characters resembling copyrighted anime. Hailuo 2.3's no-reference attempt delivered simple but on-prompt lemon and blueberry characters.