PAN 2 vs Odyssey 3: The Only World Model Here With a Price Tag Also Publishes Its Worst Number

Two claims about physical reasoning, published three weeks apart, and the difference between them is not the score. PAN 2 is the Institute of Foundation Models' 29-billion-parameter open-weight world model — Apache-2.0 weights on 7 October 2026, code on 5 November, Day-0 vLLM-Omni support on 10 November — which renders up to twenty seconds of 704×1248 video at 24 frames per second from a starting frame and a described action. Odyssey 3 is the autoregressive diffusion transformer of roughly 14 billion parameters from Odyssey, released as a free research preview on 8 October 2026 at about $0.139 per generated video on the base tier and $0.267 on Pro. One is a downloadable checkpoint with no price. The other is a metered product with no weights. And the reason this pairing is worth an article rather than a paragraph is that the paid one published a benchmark caveat most labs would have buried, while the free one has not been benchmarked by anybody at all.

That inversion is the useful part. The instinct is to read "open weights" as "more transparent" and "commercial API" as "more marketing". Here it runs the other way, and understanding why tells you more about evaluating world models in 2026 than any leaderboard position does.

What Odyssey published, including the part that costs it a point

Odyssey 3 Pro reported 66.1 on Physics-IQ Verified, a video-to-video physical reasoning benchmark, and described it as state of the art. The method behind that number is the interesting bit: it came from a single run using best-of-eight selection, while the benchmark's own rules call for four runs with a reported standard deviation. Without selection, the four-run average was 63.37.

Both numbers are Odyssey's. The 2.73-point gap between them is the price of picking the best of eight samples and calling it the score, and the fact that both numbers exist in public is more transparency than the category usually offers. On the benchmark's own measure of net improvement over the track average, Odyssey 3 Pro ranks second at +11.15 percentage points, behind FLUX 3 [large] at +12.27. On WorldMark, Odyssey's own evaluation, it is first in three of four categories — 77.2 first-person stylised, 79.0 third-person real, 76.3 third-person stylised — and third in first-person real at 80.6, behind Lyra 2.0 at 84.4 and AlayaWorld at 83.0.

Read those numbers the way a buyer should and three things follow. First, the model is genuinely competitive: second on the benchmark's own improvement measure is not a marketing artefact. Second, it does not win the category it is marketed in, and its own evaluation says so in one of four tracks. Third, and most usefully, the vendor's willingness to publish a selected number and the unselected average side by side is a signal about how to read everything else they publish.

Now turn the same lens on PAN 2. IFM published considerably more evaluation than most video-model releases — 300 single-action cases across six capability categories; 100 egocentric multi-step tasks posed both ways, 200 prompts and 456 steps, scored on step completion and full-plan completion over three seeds; Physics-IQ Verified on the image-to-video setting normalised so that the variance between two real takes equals 100; PhyGround on a five-point scale judged by Qwen3.8-27B; RBench as an unweighted mean over five task and four embodiment categories; and a blind pairwise human study on 140 prompts with thirteen reviewers producing 1,400 comparisons, Elo fitted by Bradley–Terry and centred at 1500. What IFM did not publish is a selection caveat, because as far as the card shows there was nothing to select from: the numbers are reported as harness output.

Both are unaudited. Only one of them is comparable to anything else, because Odyssey ran a public benchmark and PAN 2's headline figures come from a private harness. That distinction — benchmark versus harness — is worth more than the scores themselves.

Two products, two kinds of friction

Price is where the comparison is easiest to state and hardest to act on.

• Unit price — Odyssey 3 is roughly $0.139 per generated video at 832×480 on the base tier and about $0.267 at 1280×720 on Pro; PAN 2 has no unit price because there is no hosted service, only weights and a GPU bill.

• The real cost — Odyssey 3's price is the whole cost of a clip; PAN 2's is the machine: thirteen seconds of 704×1248 video takes roughly 119 seconds on four B300s, with eight H200s in BF16 as the documented default.

• Access friction — Odyssey 3 has a free browser demo anyone can open, and API access by contact form with no self-serve keys and no published rate card; PAN 2 is 86.17 gigabytes across sixty files under Apache-2.0, ungated, no conversation required.

• Weights — Odyssey 3 publishes none and has no licence to read; PAN 2's are Apache-2.0 with no attribution requirement, no user threshold and no field-of-use restriction.

• Speed in the harness — Odyssey 3 runs at 16 frames per second at 832×480, distilled into a few steps so a person can steer in real time; PAN 2 runs 24 frames per second across fifty sampler steps at guidance 2.5.

Put those five lines together and the two products are aimed at different buyers. Odyssey 3 is priced for someone who wants clips and does not want to own a datacentre — a content team, a research group running a few thousand generations, anyone whose unit economics are per-output. PAN 2 is priced for someone who already has the node, because for everyone else the licence being free is the least important thing about it.

The missing number on both sides is the same one, and it is not the price. Neither vendor publishes a success rate for the thing buyers actually care about. Odyssey ran a driving policy, a robot arm and a humanoid inside its generated worlds, and the detail of those demonstrations is worth reading closely — but the number a buyer needs is not in any of them: no success rate on the road task, no baseline against training in the real world, no ablation showing which part of the generated world did the work.

PAN 2's equivalent number is worse and public. Its model card states that about two in three simulated plans contain at least one step that is not carried out. That is a damning figure for anyone hoping to hand the output to a robot, and it is also the single most useful sentence in either release, because it tells you the failure rate of the thing you are about to build on. A vendor that publishes a two-in-three failure rate and a vendor that publishes no rate at all are not equally informative, and the one with the bad number is the one you can plan around.

What each one actually generates, and why the difference is architectural

Odyssey 3 is an autoregressive diffusion transformer of roughly 14 billion parameters, trained on internet video, gameplay footage with time-aligned input and simulated rigid-body interactions, then distilled into a few-step model so a person can navigate the output in real time: you prompt an environment, move through it in first or third person with independent camera control, and trigger events as you go. The tiers differ only in what you can see — 832×480 on the base tier at about $0.139 a video and 1280×720 on Pro at about $0.267 — with a Flash variant behind the free browser demo that anyone can open without an account.

PAN 2 is a batch renderer with a language model inside it. A conceptual reasoner writes the next world state as a structured, time-grounded description, a frozen Qwen3.5-9B text encoder turns that into conditioning features, a 29B diffusion transformer fills in the perceptual detail, and a Wan 2.2 VAE decodes twenty seconds of 704×1248 at 24 frames per second. There is no interactive loop: one call, one clip, no feedback inside the call. What there is instead is a contract — the prompt-rewriter skill, a fixed negative prompt and a validator that checks captions against the encoder's budget of 1,480 tokens soft and 1,536 hard, with the card admitting that the agent case studies relied on hand-refined prompts.

That difference decides what the two are good for far more than the price does. Odyssey 3 is built for a person in an environment: generate, look around, adjust, generate again. PAN 2 is built for a machine asking a question: if I do this, what happens, and would a checker prefer a different option. The samplers encode the intent — eight distilled steps for one, fifty steps with guidance at 2.5 on every step for the other — and so does the failure each one documents.

The demonstration neither company has turned into a measurement

Both releases lean on video evidence of a policy learning inside a generated world, and neither one has published the number that would make the evidence load-bearing.

Odyssey trained a driving policy on about 20 hours of real Indian road data with a frozen backbone inside its generated worlds, and separately showed robot-arm manipulation learned from tens of hours of demonstrations, including recovery behaviours that were not in the training data, humanoid policies built with Flexion, drone navigation, and a GTA V-trained policy transferring to Red Dead Redemption 2. The game-to-game transfer is the strongest argument for a general video world model as a training substrate, because if a policy can learn from footage of one game and work in another, the model captured something about dynamics rather than about pixels. What is missing is any published success rate on the Indian road task, any baseline against training in the real world, and any ablation showing which part of the generated world did the work.

PAN 2 does not have a policy demo of that kind, and it does have something Odyssey does not: a published failure rate. About two in three simulated plans contain at least one step that is not carried out. Its card also concedes that simulating plans which should fail is, in its own words, not yet measured — which matters for a training substrate, because a world that only renders success teaches a policy that everything works.

Line those two admissions up and the gap in the category becomes visible. One company demonstrated transfer and did not measure reliability. The other measured a failure mode and did not demonstrate transfer. Neither is the study a buyer needs, which is both models on one task with one policy and one metric, run enough times that the standard deviation is smaller than the gap.

What the access paths cost you in time rather than money

Odyssey 3's friction is a conversation. There is a free public research preview in the browser, API access by contact form, no self-service keys, no published per-token or per-clip rate card beyond the tier figures, no open weights and no licence to read. For a team that wants to benchmark a model against its own workload this afternoon, that is a different obstacle from a price: you can try the demo and you can talk to sales, and neither gets you a reproducible harness.

PAN 2's friction is a machine. The weights are 86.17 gigabytes across sixty files under Apache-2.0, ungated, no conversation required — and then you discover that the documented serving configuration is a node, that the sampler needs a hundred forward passes per clip, and that a twenty-second 704×1248 clip is its own decoding problem before parallelism is even discussed. The licence is free and the hardware is not, which is the same trade the whole open-weight category makes and the one buyers most often get wrong.

Neither friction is a flaw. They are the two shapes a research release takes in 2026: a product with a price and a form, or a checkpoint with a licence and a bill. Knowing which shape you can absorb is a more useful filter than any benchmark position, and it is the filter most teams apply last instead of first.

What a 16-fps product and a 24-fps research model are each for

Both companies use the phrase real time and neither means the same quantity. Odyssey 3 at 16 frames per second is 62.5 milliseconds per frame, fast enough that a person steering a camera does not feel the lag — that is what a product needs, and it is why the model is distilled into a few steps rather than fifty. PAN 2's twenty seconds at 24 frames per second is not interactive at all; it is a batch render whose cost is measured in node-minutes, and its purpose is to produce a clip an agent or a researcher then inspects.

The consequence for anyone planning work is that the two are usable in different loops. Odyssey 3 fits a human-in-the-loop workflow: generate, look, adjust, generate again, at a price low enough that iterating is cheap. PAN 2 fits an automated loop: generate five candidate consequences, score them with a checker, act on the best, at a price high enough that you think before you spawn the batch. One optimises for latency per clip, the other for fidelity per clip, and the sampler settings make the trade explicit — fifty steps with guidance at 2.5 on every step is a hundred forward passes through a 29B transformer, versus eight distilled steps through an 8-to-14B-class one.

There is also a structural difference in what the output can be used for. Odyssey 3's generated worlds have trained a driving policy and a robot arm, which is the strongest public evidence that a general video world model can serve as a training substrate for embodied control. PAN 2's output has not, and its card is explicit that simulating plans which should fail is not yet measured — which matters, because a training substrate that only shows success teaches a policy that everything works. Odyssey's demonstrations do not resolve that either: a policy trained in generated worlds that has never been evaluated against a real-world baseline is a promising result, not a validated method.

Where OrcaRouter fits into either is the evaluation layer rather than the generation layer, and it is worth being exact. Neither PAN 2 nor Odyssey 3 is in our catalogue: one is a 29B diffusion model you run on your own node, the other is a hosted product sold through a contact form, and no router can serve either. What can be routed is the machinery around them, which in both cases is where the real spend accumulates — a vision-language judge scoring generated clips against a rubric, a planner turning a research question into a batch of prompts, a caption writer producing the structured world-state descriptions PAN 2's text encoder consumes inside a budget of 1,480 tokens soft and 1,536 hard. Those are thousands of ordinary calls against a model whose output cost you a GPU minute each, which is exactly the shape of workload where a single key over more than 200 models, provider list price passed through at 0% markup and automatic failover between providers stops being a convenience and starts being the difference between a batch that finishes and one that dies at 80%.

How to read the next one of these

Three questions, in order, and none of them is "what did it score".

Is the number from a public benchmark or a private harness? Odyssey's 63.37 four-run average can be compared to other entries on the same track. PAN 2's step-completion figures cannot be compared to anything, because nobody else runs that harness. Both are legitimate; only one is portable.

Was the score selected? A best-of-eight number and a four-run average measure different things, and a vendor that publishes both has told you which one to quote. A vendor that publishes only the first has told you something too.

What is the failure rate on the task you care about? PAN 2 answers this with two in three plans dropping a step. Odyssey does not answer it for the Indian road experiment. The absence is not dishonesty — nobody has run the study — but a buyer who does not ask is buying a demo.

Applied to the choice in front of you: Odyssey 3 is the accessible instrument. A free browser demo, a contact form, a per-clip price you can multiply, and a benchmark number with a documented method behind it. If you need to try world models this week and your budget is per-clip, it is the only one of the two you can start with, and the caveat is that "accessible" means a sales conversation rather than a key.

PAN 2 is the auditable instrument. Weights, code, a serving recipe, a caption contract with a token budget, a guardrail with false-positive numbers attached — 11.5% of benign text-to-video requests and 3.3% of benign image-to-video requests rejected, with unsafe text-to-video output falling from 33.9% to 1.5% — and a limitations section that leads with the failure rate. If you have the hardware and you need to know how the thing works rather than that it works, it is the more complete release in the category by a wide margin.

What would settle the comparison is the study neither company has run: both models, one task, one policy, one metric, enough runs that the standard deviation is smaller than the gap. Nobody has proposed it. Until someone does, the accurate summary is that the paid model published a better-documented number and the free model published a worse one, and the free model is the one you can check.