Odyssey Launches Odyssey-3 World Model With 66.1 Physics-IQ Score
Odyssey on 8 October 2026 launched Odyssey-3, saying Odyssey-3 Pro scored 66.1 on Physics-IQ Verified video-to-video and opened a free Flash research preview.
PromptCrates Editorial
Staff Writer
Odyssey, an AI research company building foundation world models, launched Odyssey-3 on Thursday, 8 October 2026, describing it as an autoregressive diffusion transformer that predicts how objects, environments and actions evolve for robots, humanoids, vehicles, drones, agents and games. The company post said Odyssey-3 Pro scored 66.1 on Physics-IQ Verified’s video-to-video benchmark—the highest reported score in Odyssey’s write-up—and ranked first in three of four WorldMark categories in the lab’s own evaluations. A free Flash research preview with first- and third-person navigation is available in-browser, while API access for physical-AI developers is by request. Business Wire carried the same-day announcement and named Flexion among early humanoid adopters.
Physics benchmarks and what Odyssey claims they measure
Physics-IQ, credited to Anates Labs and DeepMind, asks models to continue videos of real physical experiments across fluid dynamics, optics, solid mechanics, magnetism and thermodynamics, then compares predictions with what happened. Odyssey reported Odyssey-3 Pro at 66.1 video-to-video with best-of-8 sampling and 54.7 image-to-video, while the standard Odyssey-3 tier improves the tradeoff between physical accuracy and generation cost so teams can run more simulations on the same budget. WorldMark scores, using the benchmark’s captions and the mean of thirteen metrics, placed Odyssey-3 first in first-person stylized, third-person real and third-person stylized splits in Odyssey’s evaluation tables.
Those figures are vendor-reported against a public leaderboard that can change as other models submit runs. Odyssey’s post included cost assumptions such as one dollar per MI355X GPU-hour excluding some prompt-rewriting fees, and resolution notes of 832×480 for Odyssey-3 versus 1280×720 for Pro. Independent ranking pages may list competing systems ahead on some submetrics; readers should treat the 66.1 claim as Odyssey’s stated Physics-IQ Verified video-to-video result at launch, not as a permanent league table. PromptCrates has covered adjacent generative-media releases such as Runway’s Claude Motion explainers and Adobe Firefly’s music and speech tools; Odyssey-3 is aimed at physical prediction and control rather than marketing stills.
How Odyssey-3 adapts to robots cars and agents
The launch post describes training an action decoder or policy on paired observations and actions while keeping the Odyssey-3 backbone as a shared world prior. With tens of hours of robot demonstrations, Odyssey said the system completed manipulation tasks and showed recovery behaviours absent from those demos, including reorienting a gripper after a missed grasp and retrieving a dropped object. Flexion’s humanoid policies built on Odyssey-3 reportedly beat tested vision-language-action baselines under environmental and lighting changes that broke those baselines. A driving policy trained on about 20 hours of Indian road data with a frozen backbone predicted waypoints for closed-loop driving. An early multi-camera experiment generated three synchronized front-camera views after about 100 training steps on an autonomous-driving dataset.
For agent research, Odyssey showed a natural-language goal pursued inside the generated world so developers can study task completion under a learned dynamics model. Training data mixes internet video with schema-verified event annotations, gameplay with keyboard and mouse traces, and simulated rigid-body interactions with captions. The architecture is a multi-step video diffusion transformer extended autoregressively with teacher forcing and causal masking, then distilled for real-time interaction. Co-authors Oliver Cameron and Jeff Hawke linked the work to a decade building driverless cars, where predicting the world was a prerequisite for choosing the next control.
Availability limits and what to watch next
The research preview is free for interactive browsing; open weights and public pricing were not published in the launch materials, and public API docs still referenced Odyssey-2 Pro in some secondary write-ups at handoff. Odyssey invited physical-AI developers to contact the company for API access. Claims about state-of-the-art scores and humanoid transfer should be read as company evaluations until third parties replicate them on shared protocols. Amazon’s Strands Decider work shows another industry push toward decision-oriented agents; Odyssey’s bet is that a shared physics-aware world model can shorten per-robot data collection.
The verifiable product news from 8 October 2026 is that Odyssey shipped Odyssey-3 with a browser Flash preview, published Physics-IQ and WorldMark numbers it says lead those tables, and demonstrated adapters for arms, humanoids and driving without releasing open weights. Teams building sim-to-real stacks should compare those numbers against their own sensor suites and against competing world models as leaderboards update.
Odyssey’s founders argue that world models have lagged language models by roughly two orders of magnitude in scale, which is why physical agents still learn many tasks from scratch. Odyssey-3 is positioned as an early sign that a single pretrained dynamics model can transfer across bodies when a thin action head is trained on modest demonstration data. That transfer story is central to the Flexion humanoid results and the twenty-hour driving policy: if the backbone already encodes fluids, contacts and optics well enough, less robot-hours are needed per deployment. Buyers should still ask for task-level success rates, failure modes and safety interlocks; Physics-IQ scores measure video continuation fidelity, not warehouse throughput or miles between disengagements.
Secondary coverage on 8–9 October repeated the 66.1 figure and noted that the public Physics-IQ Verified board can reorder as new proprietary entries appear, including systems that lead some display columns while Odyssey leads others. Odyssey’s own charts emphasise best-of-8 sampling and prompt-enhanced variants, so apples-to-apples comparisons require matching decode budgets. For simulation buyers, the Flash preview is the near-term evaluation path; for robotics teams, the unanswered commercial questions are API pricing, rate limits, data residency and whether fine-tunes of the backbone will be offered. Until those terms ship, the launch is best read as a research-preview product with strong vendor benchmarks and concrete demos across arms, humanoids and driving—not as a fully packaged enterprise platform.
