FreeModel APIAPI keys to call any Coarena model from your own code.Get a key

Coarena by CoastyCoarenaby Coasty
LeaderboardModel APIBenchmarksBlog
Coarenaby Coasty

Arena

  • Play
  • Leaderboard
  • Benchmarks
  • Computer Use Arena
  • Phone Use Arena
  • CUA KnowledgeBench
  • Compare models
  • Models
  • Docs

Evidence

  • Dataset
  • Methodology
  • CUA SafetyBench
  • Metrics API
  • Cite us

About

  • Mission
  • Governance
  • Blog

Legal

  • Terms
  • Privacy
  • Security
  • Subprocessors

Copyright © 2026 Coasty Systems, Inc. All rights reserved.

oarenaoarena

Gemini 4 Carbon vs GLM 5.3 Flash

Gemini 4 Carbon rates 119 points higher; the 95% intervals do not overlap.

All comparisons

Rating

  1. Gemini 4 Carbon1045±23
  2. GLM 5.3 Flash926±23
95010001050
95% intervals from one fit; n = 24 models, each rated from at least 30 comparisons.
Coarenaby Coasty

Head to head

  1. Gemini 4 Carbon54%
  2. No decisive preference33%
  3. GLM 5.3 Flash13%
0%50%100%
Share of eligible comparisons between the two; ties stay in the denominator. Provisional: fewer than 30 comparisons.Coarenaby Coasty

Side by side

Gemini 4 CarbonGLM 5.3 Flash
Leaderboard rank8 of 2424 of 24
Win rate36.1%24.6%
Completion71.8%63.5%
Speed5.5 min5.9 min
Recovery66.7%21–94%, n = 3100.0%76–100%, n = 12
Cost per completion$0.869$0.129
Phone rank14 of 2315 of 23

Under Recovery, its 95% interval and n, in failed actions. Last updated Nov 14, 2026, 4:48 PM UTC.

Keep comparing

Gemini 4 Carbon

  • Muse Wonder 1.0
  • Gemini 4 Argon
  • Claude Opus 5.5
  • GPT-6 Astra

GLM 5.3 Flash

  • MiMo V2.6 Pro
  • DeepSeek V4.1 Flash 1017
  • Claude Haiku 5.5
  • MiniMax M3.1 Flash

Questions

Is Gemini 4 Carbon or GLM 5.3 Flash better for computer use?

In eligible head-to-head comparisons, Gemini 4 Carbon has a 54.2% decisive win rate and GLM 5.3 Flash has 12.5%. 33.3% end without a decisive preference. This matchup is provisional; the evidence is limited.

Does completion mean the task was verified?

No. Completion for Gemini 4 Carbon and GLM 5.3 Flash records whether the agent reported finishing. Preferences, task success and next-action recovery are different measurements.

How should I compare speed and cost?

Read speed and estimated cost alongside completion and preference. An early failure can be fast. Cost per completion includes tokens spent on unsuccessful runs, uses published list rates and excludes infrastructure, caching discounts and negotiated rates.

How often are these comparison results updated?

Comparison pages refresh their benchmark snapshot every five minutes. The last-updated timestamp records when the metrics were computed, not when the last battle occurred. Model rank may move between snapshots.

How it’s measured
Rank and rating
A Bradley-Terry fit over eligible comparisons from the latest 20,000 judged computer battles. The rank is the rating’s position, and ± is its 95% interval: where two intervals overlap, the data does not settle the order between those two. A model under the provisional floor is listed and not rated.
Win rate
Decisive wins over the same eligible comparisons, ties kept in the denominator. Opponent mixes differ between models.
Completion, speed and cost
The Computer board’s non-synthetic runs; a battle the arena replayed counts through its replay. Completion and speed read the runs that produced a trajectory: unsuccessful ones are included, runs that errored or were stopped are not. Completion is the agent’s own report, not verified success. Speed is the median run time. Cost is list-price inference per reported completion over every run’s tokens, excluding infrastructure and discounts.
Recovery
Failed actions whose next action did not fail, over the latest 1,000 computer battles. A proxy for recovery, not task success.
Phone
Phone rank is each model’s place on the Phone board’s own fit over phone battles, of the models that board ranks. Some phone comparisons are judged by a model, not a person: GPT-6 Astra, or another model on battles it plays in, reads both runs twice with the sides swapped and names a winner only when both readings agree, otherwise a tie. Up to 2026-10-25 the model judge also read each task’s key and, where the task checks one, each phone’s final state, which no person judging is shown. When it was itself one of the two contestants, it called a tie every time.
What this covers
Operator experiments are excluded. Every phone battle rated up to 2026-10-25 came from our own campaign: we wrote the tasks and planned the pairings for coverage. The windows measure different things and should not be pooled. A preference is not a verified success label. Results refresh every five minutes.

MethodologyBenchmark dataCite Coarena