OdysseysMega measures whether an agent can turn extended work across live websites into a usable, review-ready result. Tasks require sustained browsing, evidence gathering, comparison, and often document or spreadsheet construction; every rubric is judged against the complete screenshot history rather than the final answer alone.
OdysseysMega measures whether an agent can turn extended work across live websites into a usable, review-ready result. Tasks require sustained browsing, evidence gathering, comparison, and often document or spreadsheet construction; every rubric is judged against the complete screenshot history rather than the final answer alone.
OdysseysMega measures whether an agent can turn extended work across live websites into a usable, review-ready result. Tasks require sustained browsing, evidence gathering, comparison, and often document or spreadsheet construction; every rubric is judged against the complete screenshot history rather than the final answer alone.
OdysseysMega measures whether an agent can turn extended work across live websites into a usable, review-ready result. Tasks require sustained browsing, evidence gathering, comparison, and often document or spreadsheet construction; every rubric is judged against the complete screenshot history rather than the final answer alone.
OdysseysMega measures whether an agent can turn extended work across live websites into a usable, review-ready result. Tasks require sustained browsing, evidence gathering, comparison, and often document or spreadsheet construction; every rubric is judged against the complete screenshot history rather than the final answer alone.
A task passes only when the judge finds every one of its rubrics satisfied. Rubrics are the task author's checklist — the specific facts, comparisons and deliverables a complete answer has to contain — and each is judged against the run's recorded screenshots and actions.
Measured on 100 tasks · 399 runs · 4 models
The strictest reading on this page: a run counts only where the judge found every one of the task's rubrics satisfied. The rubric score on the board gives partial credit instead — the share of a task's rubrics a run met.
The strictest reading on this page: a run counts only where the judge found every one of the task's rubrics satisfied. The rubric score on the board gives partial credit instead — the share of a task's rubrics a run met.
The strictest reading on this page: a run counts only where the judge found every one of the task's rubrics satisfied. The rubric score on the board gives partial credit instead — the share of a task's rubrics a run met.
The strictest reading on this page: a run counts only where the judge found every one of the task's rubrics satisfied. The rubric score on the board gives partial credit instead — the share of a task's rubrics a run met.
The strictest reading on this page: a run counts only where the judge found every one of the task's rubrics satisfied. The rubric score on the board gives partial credit instead — the share of a task's rubrics a run met.
Claude Opus 5.5
41%
best model's pass rate
Across 100 tasks, Claude Opus 5.5 — the strongest model — satisfies every rubric on 41% of its runs.
Claude Opus 5.5
41%
pass rate
Across 100 tasks, Claude Opus 5.5 — the strongest model — satisfies every rubric on 41% of its runs.
GPT-6 sol
20%
pass rate
Across 100 tasks, GPT-6 sol — ranked 2nd of 4 — satisfies every rubric on 20% of its runs.
GPT-5.6 sol
18%
pass rate
Across 100 tasks, GPT-5.6 sol — ranked 3rd of 4 — satisfies every rubric on 18% of its runs.
Claude Opus 5
15%
pass rate
Across 100 tasks, Claude Opus 5 — ranked 4th of 4 — satisfies every rubric on 15% of its runs.
Recorded runs, frame by frame. Pick a model, then its task failures or strong runs; the checks beside the frame appear at the first frame the judge's evidence cites.
The tasks where the models disagree most. Open one for the request, a replay of every run, and how each rubric was judged.
The tasks where the models disagree most. Open one for the request, a replay of every run, and how each rubric was judged.
The tasks where the models disagree most. Open one for the request, a replay of every run, and how each rubric was judged.
The tasks where the models disagree most. Open one for the request, a replay of every run, and how each rubric was judged.
The tasks where the models disagree most. Open one for the request, a replay of every run, and how each rubric was judged.
Travel and Tourism
Plan a 10-day international vacation from India for a first-time international traveler (indian nationality and passport, obviously). Research and compare suitable destinations, verify entry requirements, build the route, find accommodation, research…
Home and Garden
Design an indoor air-quality and humidity-control system for a 12 m² bedroom study in Enfield with a 2.4 m ceiling, opening window, central heating and no air conditioning. I use the room for long study sessions and sleep, have no diagnosed respiratory…
Travel and Tourism
Plan a realistic 9-day family trip through France for two adults and one child, using a travel window at least 30 days from when the task starts. The trip should combine Paris with two additional destinations reachable by practical rail connections and should…
Health
I am setting up a strength training corner of my garage and want a spreadsheet tracking real candidates before I buy anything. Find 4 candidate adjustable dumbbell pairs, 4 candidate adjustable weight benches, and 4 candidate power racks, each category…
Food and Drink
I want to plan a practical one-week food-focused trip to Seoul that combines well-known Korean dishes with local markets, neighborhood restaurants, and a few distinctive food experiences. The goal is to create a realistic itinerary that balances sightseeing…
Travel and Tourism
Plan a realistic 9-day family trip through France for two adults and one child, using a travel window at least 30 days from when the task starts. The trip should combine Paris with two additional destinations reachable by practical rail connections and should…
News & Media Publishers
Compile a review-ready comparison of the main official Cuban news agencies and media publishers suitable for staying informed on current events and local developments. Focus on digital access options and any notes on print or multi-platform availability that…
Community and Society
A colleague's family is formalizing their annual charitable giving and wants a rigorously vetted shortlist of ocean-conservation and marine-research nonprofits with a proposed $50,000 allocation across them not a list of well-known names, an actually verified…
Science and Education
Develop a 6-month language learning plan for an English speaker who wants to achieve conversational proficiency in Japanese. Research textbooks, online courses, language exchange platforms, vocabulary tools, grammar resources, pronunciation practice and mock…
Travel and Tourism
I will be studying Advanced Software Engineering at King’s College London’s Strand Campus and live near Enfield Town station. Design a resilient commuting system for four campus days per week across 30 teaching weeks. Compare exactly six viable door-to-campus…
Travel and Tourism
Plan a 10-day international vacation from India for a first-time international traveler (indian nationality and passport, obviously). Research and compare suitable destinations, verify entry requirements, build the route, find accommodation, research…
Arts & Entertainment
I’m considering launching or partnering with a streaming and online TV service in South Korea. Research the current streaming market, identify the major platforms and their content strategies, compare subscription models and audiences, examine local…
Food and Drink
Build a reusable Beverage Storage, Shelf-Life & Quality-Control Decision System that helps a household, hospitality operation, small producer, or food-service team determine how beverages should be stored, monitored, assessed, and handled when conditions…
Vehicles
I’m helping a group of EV drivers decide which public charging networks and charging locations are genuinely dependable for everyday and longer-distance use. I don’t want to judge a network simply by how many chargers it operates or how impressive its…
Arts & Entertainment
I want to explore France’s visual-arts and design scene in a way that goes beyond simply listing famous museums. Research major art institutions, contemporary galleries, design destinations, exhibitions, festivals, fairs, educational opportunities…
