Odyssey-3 vs Hailuo 2.3: The Incumbent Video Model Its Own Maker Now Calls Legacy

H​ailuo 2.3 is the M​iniMax video model that shipped on October 28, 2025 and held a reputation for physics and motion quality that most generators of its generation could not match. It is also, as of this month, filed under Legacy Models on M​iniMax's own documentation, superseded by MiniMax H3. Odyssey-3 is ten days old, published by Odyssey on October 8, 2026 as a research preview, and it does something H​ailuo 2.3 was never designed to do: generate an environment from a prompt and keep predicting it in real time as you move through it or trigger an event. Comparing them is not a fair fight in either direction, and the reason is worth spelling out, because it is also a description of how fast this category moved in twelve months.

H​ailuo 2.3 is still callable and still priced. That is the part that matters if you are choosing today: one of these two models has a rate card and a working API, and the other one has a browser preview and a contact form.

What H​ailuo 2.3 actually is, in M​iniMax's own specification

M​iniMax's model list describes M​iniMax H​ailuo 2.3 as text-to-video and image-to-video with state-of-the-art instruction following and what the company calls extreme physics mastery, at 1080p for six seconds or 768p for six or ten seconds, all at 24 fps. M​iniMax H​ailuo 2.3Fast is the cheaper variant, image-to-video only, same resolutions and durations. The 768p ten-second mode is the one that distinguishes it from most of its contemporaries: a ten-second single generation at 24 fps was, and largely still is, a real capability rather than a marketing line.

The arrival post on M​iniMax's Chinese news feed is dated October 28, 2025 and frames the release as an upgrade to H​ailuo 02 focused on dynamic expressiveness and realism. There is no open-weight release, no technical report that we can find, and no third-party evaluation of H​ailuo 2.3 anywhere on the public boards we checked — including Physics-IQ Verified, where neither H​ailuo 2.3 nor its Fast variant appears at all.

What it costs, per clip and per second

M​iniMax's pricing page lists the H​ailuo line as flat per-clip prices, which is unusual and unusually easy to plan with:

• M​iniMax-H​ailuo-2.3 — $0.28 per 768p six-second clip, $0.56 per 768p ten-second clip, $0.49 per 1080p six-second clip.

• M​iniMax-H​ailuo-2.3-Fast — $0.19 per 768p six-second clip, $0.32 per 768p ten-second clip, $0.33 per 1080p six-second clip.

• M​iniMax-H​ailuo-02, the predecessor, is still listed at the same 768p and 1080p rates plus a 512p tier at $0.10 for six seconds and $0.15 for ten.

Divided out — this is our arithmetic on M​iniMax's list prices, not a M​iniMax figure — that is roughly $0.047 per second at 768p and $0.082 per second at 1080p on H​ailuo 2.3, and about $0.032 and $0.055 per second on the Fast tier. Those are cheap numbers for a model in the 1080p class.

Odyssey-3 has no list price to divide. Its only public cost figures are the normalized estimates on the Physics-IQ Verified cost column: $0.139 per generated video for Odyssey-3 and $0.267 for Odyssey-3 Pro. Those are not comparable to M​iniMax's rates, because the leaderboard normalizes its cost figures to 24 fps and 1280-wide output where the inputs allow while M​iniMax's figures are what the invoice says. Putting the two next to each other as if they measured the same quantity would be the kind of comparison that looks rigorous and is not.

The physics claim, and where it can be checked

H​ailuo 2.3's central claim was physical accuracy — M​iniMax's own phrase is extreme physics mastery — and the benchmark that exists specifically to test that claim is Physics-IQ Verified, run by Anates Labs with DeepMind. H​ailuo 2.3 is not on it. MiniMax H3, the open-weight successor released on July 31, 2026, is: rank 9 at +0.95 pp net improvement over the track mean on the benchmark's own prompts, 24 fps, normalized cost $0.438 per video. MiniMax H3 Max, the high-speed tier of the same model, sits at rank 11 at −2.63 pp.

So the family's successor is on the board and slightly positive; its own generation was never submitted. Odyssey-3, on the same board and the same prompt set, sits at rank 8 at +2.15 pp and $0.129 per video, and at rank 3 with +9.99 pp when submitted with custom prompts. Odyssey-3 Pro, on custom prompts, is rank 2 at +11.15 pp.

Read that honestly and it says three things at once. A world model built for interaction is scoring above a family of purpose-built video models on a physical-plausibility benchmark. The comparison is asymmetric, because the incumbent's own generation was never tested there. And the margin, +2.15 against +0.95 on the matched prompt set, is well inside the ±0.84 and ±0.35 point error bars those two rows carry. On the evidence that exists, Odyssey-3 and M​iniMax's best entry are not measurably far apart on physical plausibility.

The difference that does not show up in any score

Every row above measures whether a model produces a physically plausible clip. H​ailuo 2.3's answer is a clip: it takes a prompt or a still image, it returns six or ten seconds of 24 fps video at a resolution you chose, and the transaction ends there. It is a good answer, and M​iniMax's per-clip price makes it a cheap one.

Odyssey-3's answer is not a clip. Odyssey describes it as a multi-step video diffusion transformer extended autoregressively through teacher forcing and causal masking, distilled into a few-step variant fast enough to interact with, and the preview exposes first-person navigation, third-person navigation and independent camera movement. You do not get a video; you get an environment that responds to what you do inside it. Nothing in H​ailuo 2.3's documentation has an equivalent, and nothing in its release history suggests M​iniMax was aiming for one — M​iniMax went toward open weights and multimodal references with H3, not toward interactivity.

That difference is why Odyssey can publish results about a robot arm recovering from a missed grasp and a driving policy trained on twenty hours of simulated data, and M​iniMax publishes clips. They are not two points on one axis. They are different axes that happen to both be labelled video generation.

Getting to them

H​ailuo 2.3 is reachable through M​iniMax's own platform API, and OrcaRouter does not host it. Odyssey-3 is not on any hosted platform and has no published rate for anyone to route.

What one OrcaRouter key does reach is the part of the family that is current: minimax/minimax-h3, the open-weight successor, is routed at $0.08 per second of 768P output — M​iniMax's own list price, passed through at 0% markup, so any change to that rate is live on your invoice the same day. The K​ling video line sits on the same key at $0.084 to $0.28 per request, alongside more than 200 models, with automatic failover across providers and a routing DSL that puts several video models behind one interface so a bake-off is a config change. If Odyssey opens a self-serve endpoint, that is the shape to add it to.

What this matchup is really about

H​ailuo 2.3 is not a model anyone would now choose as a frontier option — its own vendor has moved on twice — but it is still a good, cheap, well-specified clip generator at a price per second that most new entrants do not beat, and its ten-second 768p mode is more than Odyssey-3's preview will give you in a single continuous output. If your requirement is finished footage at a predictable cost, H​ailuo 2.3 is the more useful model today, and no amount of physics-benchmark positioning changes that.

Odyssey-3 is the more interesting one, and it is not yet a product. It is the model to watch if your problem is prediction rather than production: what happens next, given what I just did. The honest summary of Odyssey-3 against H​ailuo 2.3 is that the incumbent is retiring and the challenger is not yet on sale, which makes this a comparison between a service you can buy and a capability you can only try.