Pull to refresh
Logo
What Changed
Following Why Sign Up
Odyssey releases Odyssey-3 world model as free public research preview

Odyssey releases Odyssey-3 world model as free public research preview

New Capabilities

The 14-billion-parameter model turns text into interactive environments and now runs a simulated fruit fly brain, as Odyssey pitches it for robots, cars, and games

October 17th, 2026: Odyssey posts fruit fly brain demo

Overview

Updated Oct 22

Odyssey's Odyssey-3 turns a text prompt into a world you can walk through. The California company opened the 14-billion-parameter model as a free research preview on October 8, and on October 17 it posted a video of a simulated fruit fly brain steering through one of its generated environments.

Odyssey pitches the same model as a foundation for robots, cars, and drones. A driving policy trained on 20 hours of simulated data drove on real roads in India, though it went about 77% as far as a policy trained on real footage before a safety driver intervened. Controllers trained in GTA V also produced movement in Red Dead Redemption 2.

Why it matters

If world models work as Odyssey claims, simulated environments become training grounds for robots and vehicles, cutting the cost of physical-world testing.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

14B
Base model parameters
Odyssey-3's base model has 14 billion parameters and generates 832×480 video.
66.1
Physics-IQ Verified score (best-of-eight)
Odyssey-3 Pro's top score on the video-to-video benchmark, from a single run with best-of-eight selection. The Physics-IQ record rule requires four runs with standard deviation, so this does not qualify as a record claim.
63.37
Physics-IQ Verified four-run average
The four-run average without the best-of-eight selection method, according to The Decoder. Both scores appear on the official leaderboard but were submitted by Odyssey.
77%
Simulation-only vs. real-footage driving distance
On real roads, policies trained entirely in simulation traveled about 77% as far between safety-driver interventions as policies trained on real driving footage, Odyssey says.
3 of 4
WorldMark categories ranked 1st
Odyssey-3 Pro ranks first in three of four WorldMark categories in company evaluations.
20 hours
Simulated driving data for policy training
Hours of simulated driving data used to train a policy that drove on real roads in India, with the Odyssey-3 backbone frozen.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

Organizations Involved

Odyssey
Odyssey
AI company
Released Odyssey-3 as a free public research preview; posted a fruit fly demo on October 17

California-based AI company building foundation world models for physical intelligence.
Flexion
Flexion
Robotics company
Partner on humanoid controllers

Swiss robotics company working with Odyssey on humanoid control.
Rockstar Games
Rockstar Games
Video game developer
Games used in Odyssey-3 experiments

Developer of GTA V and Red Dead Redemption 2, used in Odyssey-3 gameplay experiments.
Square Enix
Square Enix
Video game developer
Sleeping Dogs used in transfer experiments

Developer of Sleeping Dogs, used in Odyssey-3 transfer experiments.
PW
Poke & Wiggle
Robot data and benchmarking company
Collaborating with Odyssey on robot evaluations

Robot data, policy analysis, and benchmarking company working with Odyssey on cross-body evaluations.

Timeline

September 2026 October 2026

5 events Latest: October 17th, 2026 · 4 weeks ago
Tap a bar to jump to that date

  1. Odyssey posts fruit fly brain demo

    Latest Research

    Odyssey posted video to X showing a simulated fruit fly brain moving through Odyssey-3-generated worlds. The company did not publish a technical description or say which connectome it used.

  2. Driving policy reported on real Indian roads

    Milestone

    TechTimes reported a policy trained on 20 hours of Odyssey-3 simulated video drove on real streets in India, with the model's parameters frozen.

  3. Free public preview goes live

    Product Launch

    The Decoder reports on the free public research preview, letting anyone generate and explore interactive environments from text prompts.

  4. Odyssey launches research preview

    Product Launch

    The company announced the research preview was available, with a free demo running on the Odyssey-3 Flash variant.

  5. Odyssey first unveils Odyssey-3

    Announcement

    Odyssey showed the model and said a public release would come in the coming weeks.

Scenarios

1

Odyssey-3 becomes a validated robotics platform

Possible Resolves by Apr 11, 2027

Discussed by: The Decoder, AI in Context

Independent labs replicate Odyssey's robot and driving demos. API access opens with disclosed pricing, and developers build controllers on the frozen backbone. The four-run Physics-IQ average holds up under external testing.

2

Benchmark scrutiny limits Odyssey-3 to research-preview status

Possible Resolves by Apr 11, 2027

Discussed by: TPS Report

The best-of-eight Physics-IQ score draws criticism, and without independent verification the model stays a demo. Robotics and autonomous-systems use requires further work, as Odyssey itself notes. The free preview remains popular but commercial adoption stalls.

3

World models become the next AI frontier, competitors follow

Likely Resolves by Oct 11, 2027

Discussed by: AI in Context

Odyssey's public preview pushes other labs to release their own world models. The category moves from closed research to open competition, with multiple free previews and API offerings within a year.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

February 2024

DeepMind Genie (2024)

DeepMind showed Genie, a model that generated playable 2D platformer worlds from a single image or text prompt. It was a research demonstration, not a product.

Then

Genie remained a research paper with no public release.

Now

It established the concept of generative interactive environments but left the field open for others.

Why this matters now

Odyssey-3 is the first major public release of a similar capability, moving from paper to playable preview.

February 2024

OpenAI Sora (2024)

OpenAI unveiled Sora, a text-to-video model that generated minute-long clips with consistent physics. It was shown to select creators, not the public.

Then

Sora generated intense interest but limited access.

Now

It took over a year for Sora to reach public availability, showing the gap between demo and product.

Why this matters now

Odyssey-3's free preview skips that gap, letting anyone test the model immediately.

January 2025

NVIDIA Cosmos (2025)

NVIDIA released Cosmos, a family of world foundation models aimed at robotics and autonomous vehicle training, with open weights.

Then

Cosmos gave developers a base for synthetic data generation.

Now

It established world models as a robotics training tool.

Why this matters now

Odyssey-3 competes directly in this space, adding real-time interactive generation and a free preview.

Sources

(17)
+11
The Decoder Odyssey AI in Context TPS Report TSN Media RoboticA Guide FinancialContent (via Business Wire) Exact Reveo News AI.info TechTimes OrcaRouter The Next Web Glitchwire Digg Odyssey OrcaRouter (Korean)