Hailuo 3 vs Hailuo 2.3: Should You Upgrade?
Which One Should You Choose?
| Your situation | Better current choice | Reason |
|---|---|---|
| You need to ship through EvoLink now | Hailuo 2.3 | Model routes, documented modes, pricing, and operational behavior can be tested. |
| You want to prepare for MiniMax's next generation | Track Hailuo 3 | Build a replayable evaluation, but keep it outside production dependencies. |
| You need a stable API model ID | Hailuo 2.3 | Hailuo 3 has no verified public model ID. |
| You need predictable cost | Hailuo 2.3 | Current route cost can be checked; Hailuo 3 price and billing are unknown. |
| You need the best visual quality | Run matched tests after release | Promotional clips and unverified specifications cannot establish a production winner. |
| You are planning a migration | Use Hailuo 2.3 as control | Promote Hailuo 3 only after quality, reliability, latency, cost, and compatibility gates pass. |
Need to generate video now? Open Hailuo 2.3 API and use a real workload to establish quality, latency, and cost baselines. Not migrating yet, but want to test early? Follow the Hailuo 3 product overview for verified information and access status.
What Can We Confirm on July 21, 2026?
| Area | Hailuo 3 | Hailuo 2.3 | Production implication |
|---|---|---|---|
| Official MiniMax API status | Teased in MiniMax's January 1, 2026 annual review; not publicly listed in the API | Listed and documented | An official name mention is not a model or API release. |
| EvoLink route | Not available | Available | Use Hailuo 2.3 for current traffic. |
| Public model ID | Not published | MiniMax-Hailuo-2.3 and MiniMax-Hailuo-2.3-Fast in MiniMax docs | Do not invent or copy a third-party Hailuo 3 ID. |
| Documented generation modes | Unknown | Text-to-video and image-to-video; Fast is positioned for image-to-video | Mode parity must be verified after launch. |
| Documented output | Unknown | 768P at 6/10 seconds; 1080P at 6 seconds; 24fps in the official model guide | Do not assume Hailuo 3 limits from rumors. |
| Price and billing | Unknown | MiniMax packages deduct video points by model and output; EvoLink uses its current route price | Do not treat provider video points, third-party credits, and EvoLink billing as one price. |
| Production evidence | None on an EvoLink route | Can be measured now | Hailuo 2.3 wins readiness, not necessarily future capability. |
Why Would You Still Choose Hailuo 2.3?
The simple reason is that Hailuo 2.3 can be tested instead of imagined. Your team can generate clips, observe queue and failure behavior, measure prompt adherence, and work out the real cost of an accepted result.
MiniMax documents Standard and Fast models. Standard supports text-to-video and image-to-video; Fast supports image-to-video only and is positioned for efficiency and cost. Current documentation lists 768P at 6 or 10 seconds, 1080P at 6 seconds, and 24fps.
prompt_optimizer enabled by default, optional fast_pretreatment, and asynchronous callbacks. Those fields describe the current MiniMax channel. They do not prove that Hailuo 3 will preserve the same contract, and they should not be copied as EvoLink parameter names without checking the live route documentation.
Hailuo 2.3 is the better starting point when:
- a release schedule depends on a callable route;
- text-to-video or image-to-video covers the workflow;
- output duration and resolution fit current limits;
- the team needs observable task states and billing;
- a baseline is required before evaluating Hailuo 3.
Its status does not guarantee that every prompt will succeed. It means failures, latency, and quality can be measured instead of inferred.
What Does Hailuo 3 Still Need to Prove?
MiniMax has said only that it plans to bring its understanding of “next-generation multimodal intelligence” through Hailuo 03. That confirms a name and direction, not an architecture, generation mode, price, resolution, duration, or audio capability. Any quality jump remains a hypothesis that matched workloads need to test.
For an upgrade decision, require evidence in five layers:
| Gate | What to measure | Promotion condition |
|---|---|---|
| Quality | Prompt adherence, identity, motion, physics, artifacts, accepted-output rate | Meaningful lift on priority workloads, not just selected demos. |
| Reliability | Success, moderation, timeout, retry, queue, callback behavior | Meets the production SLO with fallback tested. |
| Latency | Queue, generation, polling, delivery, p50/p95 | Fits user experience or batch deadlines. |
| Cost | Price, retries, rejected outputs, reviewer time | Accepted-output cost is defensible against Hailuo 2.3. |
| Compatibility | Inputs, parameters, task states, output format, safety behavior | Adapter changes are understood and reversible. |
Claims such as 4K, 60 fps, 30-second generation, or native audio should enter this table only when official documentation defines their exact contract.
How Should You Test the Quality Difference?
Start with a fixed set of real jobs, not prompts chosen to flatter either model. Send the same text, image, aspect ratio, duration target, and negative constraints to both routes. Then randomize reviewer order so the model name does not influence scoring.
Cover these categories:
- single-subject motion with precise camera direction;
- two or more interacting subjects;
- hands, faces, and object transfer;
- consistent wardrobe, products, and logos;
- first-frame conditioning and large camera movement;
- typography and UI-like compositions;
- difficult physics such as liquids, cloth, and collisions;
- style continuity across multiple clips.
MiniMax's own 2025 review provides unusually useful stress cases. It describes distant faces becoming blurred or over-sharpened, left/right-hand mistakes, common-sense behavior overriding an explicit instruction, and object deformation during actions such as opening a door. MiniMax also says that a clip requiring a character not to pour wine into a glass took 89 attempts. These examples do not prove that Hailuo 2.3 always fails. They do define better upgrade tests than a hand-picked demo reel.
| Stress test | Do not score only | Record instead |
|---|---|---|
| Distant and wide-shot people | One attractive frame | Face stability, flicker, sharpening, and identity drift |
| Left/right hands and object transfer | Final pose | Direction compliance, contact, fingers, and object integrity |
| Counter-intuitive instructions | Whether a clip exists | Whether the model follows the instruction instead of restoring the common action |
| Doors, liquids, and collisions | Selected best clip | First-pass acceptance, retries, and deformation |
| Camera commands | A general “cinematic” impression | Whether each command runs, when it runs, and whether the subject stays stable |
Score every run, not only the best candidate. A model that needs five attempts to produce one acceptable clip may be less useful than a model with slightly lower peak quality and a higher first-pass rate.
How Should You Compare Speed and Cost?
Here is the trap to avoid: comparing Hailuo 2.3's known price with a guessed Hailuo 3 price. The same goes for comparing a Fast 768p result with a hypothetical Hailuo 3 maximum-quality result and calling the difference a model advantage.
Use:
accepted_output_cost = generation_price + retries + failed_or_moderated_charges + reviewer_timeRecord resolution, duration, mode, queue, retries, and billing outcome for each result. MiniMax's current video package deducts 1, 2, and 2 video points for Standard at 768P/6s, 768P/10s, and 1080P/6s. The corresponding Fast deductions are 0.7, 1.1, and 1.3 video points. The package also says failed generations and videos under security review are not deducted.
Those values apply to MiniMax's official video package. They are not per-generation dollar prices, and they are not EvoLink or another aggregator's credits. Normalize every channel to one billing unit before calculating accepted-output cost; identical words such as “points” or “credits” do not make the units comparable.
Do the same for the data contract. Record how long task status and output URLs remain available, whether requests or results are cached, how duplicate submissions are handled, and when uploaded assets and generated files are deleted. Those details can break a migration even when generation quality improves.
These are also the issues most likely to derail a migration: changes in credits or billing units, 1080P cost, billing for failed or moderated jobs, queue throttling, character consistency across shots, and the attempts required for one accepted clip. A generational upgrade should produce a measurable improvement on at least some of those dimensions.
How Do You Migrate Without Breaking Production?
1. Keep Hailuo 2.3 as the control
Preserve prompts, input-asset versions, output acceptance rules, timeout settings, task states, and cost logs. Do not change the control prompt while evaluating the challenger.
2. Add Hailuo 3 behind configuration
Use a separate challenger route key. Do not let one model alias silently point to two generations. Map only verified parameters and reject unsupported combinations explicitly. Never silently reinterpret a duration, resolution, or reference input.
3. Run offline paired evaluation
Replay representative tasks. Store raw outputs, request metadata, latency, failure states, moderation, and billing. Keep the model identity hidden from reviewers when practical.
4. Shadow and canary
Generate Hailuo 3 outputs without serving them first. Confirm task-state and billing behavior, then send a small percentage of eligible work. Use workload-specific rollback thresholds.
5. Promote by workload
Even if Hailuo 3 eventually wins a high-motion workload, it may not fit cost-sensitive drafts or every other job. Promote by workload rather than forcing one family-wide default.
What Must Be Verified in Configuration?
| Contract area | Hailuo 2.3 baseline | Hailuo 3 launch policy |
|---|---|---|
| Model selection | Standard and Fast are separate | Use a separate route key; never silently replace 2.3 |
| Input modes | Standard is T2V/I2V; Fast is I2V | Reject unverified modes instead of silently degrading |
| Duration and resolution | Documented 6/10-second and 768P/1080P combinations | Expose only combinations listed in the verified route contract |
| Prompt handling | MiniMax documents optimizer controls and camera commands | Save original and transmitted prompts so optimization does not bias the comparison |
| Async jobs | Create, query, download, and callback form the lifecycle | Verify status values, callback retry, timeout, and duplicate delivery |
| Billing and failure | MiniMax package rules and EvoLink route rules are separate | Verify success, failure, moderation, cancellation, and timeout billing separately |
| Data lifecycle | Tasks, output URLs, and uploads need separate checks | Record TTL, caching, idempotency, retention, and deletion before launch |
Rollback should not depend on a vague sense that results are worse. Use current Hailuo 2.3 data to set an allowed range for each workflow. If Hailuo 3 crosses that range for accepted-output rate, p95 latency, failure rate, moderation rate, or accepted-output cost, route back to Hailuo 2.3. Ad drafts, continuity-heavy character clips, and premium final footage should not share one threshold.
What EvoLink Routing Policy Makes Sense?
| Role | Route now | Future action |
|---|---|---|
| Production MiniMax video | Hailuo 2.3 Standard | Keep until a challenger passes gates. |
| Faster image-to-video drafts | Hailuo 2.3 Fast | Compare latency and accepted-output cost with any Hailuo 3 fast tier if officially released. |
| Next-generation evaluation | Hailuo 3 waitlist | Activate only after route verification. |
| Cross-provider fallback | Seedance 2.0 or Kling 3.0 | Test before an incident, not during one. |
For most EvoLink users, the useful next step is not waiting for an unverified new default. Run one real Hailuo 2.3 job and save the baseline. When a verified Hailuo 3 route arrives, you will already have reproducible tasks, acceptance rules, and a rollback target, so the upgrade decision can be tied to retries and accepted-output cost rather than launch hype.
FAQ
Is Hailuo 3 better than Hailuo 2.3?
There is not enough official or route-level evidence to answer. Hailuo 2.3 is available; Hailuo 3 has not been published in MiniMax's public API docs.
Which model should I use today?
Use Hailuo 2.3 when you need a callable EvoLink route. Track Hailuo 3 and prepare matched tests if a future migration matters.
What are Hailuo 2.3's documented modes?
MiniMax documents text-to-video and image-to-video for Hailuo 2.3, with a Fast image-to-video route.
Does Hailuo 3 support longer or higher-resolution video?
No official public Hailuo 3 limits have been verified. Treat detailed third-party numbers as unconfirmed.
Is there a Hailuo 3 API price?
No official MiniMax or EvoLink route price is published.
Will Hailuo 3 replace Hailuo 2.3?
MiniMax has not published a replacement or deprecation notice. Keep Hailuo 2.3 until lifecycle and migration guidance is official.
How should I compare costs?
Compare accepted clips under the same duration, resolution, mode, prompt set, retry policy, and review standard.
What is the safest migration approach?
Offline paired tests, shadow generation, workload-specific canary routing, and a rehearsed fallback to Hailuo 2.3.
Will Hailuo 3 use the same cache and retention rules as Hailuo 2.3?
Do not assume so. Verify task retention, output URL lifetime, caching, idempotency, uploaded-asset retention, and deletion behavior on each EvoLink route before migration.
