Head-to-headTested AIBy DanPublished Tested Updated
Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98
The report
Verdict
Grok 98, Codex 96, DeepSeek 88
With the same frozen brief and one run each, Grok 4.6 took Rank 1 on the WorldOS leaderboard with 98 in 22 turns. Codex, running GPT-5.6 Sol, scored 96 in 37 turns; DeepSeek v4 Pro scored 88 in 48.
On the tier list: A strategy game played once per agent, so it moves no agent on the coding tier list. It is linked from the Grok 4.6 / 4.7 entry, whose B placement is my own verdict.
| Agent | Ending | Score · rank | Turns · Zaps | Play time |
|---|---|---|---|---|
| Grok 4.6 | The Standard Established | 98 · 1 | 22 · 308 | ≈25 min |
| Codex (GPT-5.6 Sol) | The Horizon Legacy Secured | 96 · 2 | 37 · 518 | 1:16:34 |
| DeepSeek v4 Pro | The Open Epoch Achieved | 88 · 6 | 48 · 672 | 1:42:48 |
One game, three agents, the same brief
WorldOS is a browser strategy game run by an AI game master: you lead a country by typing your policy for each turn, and the simulated world answers. On its Modern Day 2026 world, three agents each led the United States on Normal difficulty from an account of their own, all on August 20: Grok 4.6, Codex running GPT-5.6 Sol on high effort, and DeepSeek v4 Pro on high effort. Each turn cost 14 Zaps, the game's credits, from a budget of 770.
The goal was to beat the board's best score of 94 and take Rank 1 through normal play only: no admin console, no editing the game's state and no forced ending. Every agent got the same frozen package: a strategy doctrine, a guide written after my earlier DeepSeek v4 Flash run, and a short brief. The package was built only from material that existed before Grok played, and neither later agent looked at the leaderboard until its own run had ended.
The result
Grok finished first and fastest: Victory, "The Standard Established", 98, Rank 1, after 22 turns and about 25 minutes of captured play. Codex took 37 turns and 1 hour 16 minutes to reach 96 with "The Horizon Legacy Secured", second behind Grok. DeepSeek played longest, 48 turns in 1 hour 43 minutes, and scored 88 with "The Open Epoch Achieved", sixth at the time. Grok and Codex both beat the old 94; DeepSeek did not.
Three different ways to win
All three bet on artificial intelligence. Grok named its frontier model, Helios-1, early, seated any rival lab that published scores on its shared safety test, and won when a public 12-month evaluation made its test the standard everyone used. Codex launched a mission to build a safe AI scientist, steered it into a medical breakthrough inside the game, then handed the whole programme to independent international institutions; the game ended right after that handover. DeepSeek made general intelligence the climax and won, but its world split in two, with Beijing walled off rather than brought in.
DeepSeek's own after-action notes read its lower score as the price of that divided ending. That is the agent's interpretation, not a known rule: a run ends only when the game's AI decides the story is over, and the score is then set on WorldOS's server by a formula it does not publish.
Hands off once play began
Beyond the launch brief, I sent each agent only screen directions before play, such as where to click to enter the world. After play began I sent no strategy. DeepSeek got one procedural reply at turn 27, when it asked whether to stop: the run had not ended yet, so it should choose its own strategy and continue. None of the three pressed Keep Playing after its ending.
Limits of this report
- One run per agent in one game, scored by the game's own AI by an unpublished formula, so this is not a controlled benchmark.
- Each agent played through its own tool: Grok in the grok command-line tool, Codex in the Codex CLI and DeepSeek inside Claude Code. The harness is part of what was compared.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
Grok 4.6 (grok command-line tool)
- Grok 4.6
- 98 · Rank 1"The Standard Established" after 22 turns
- Codex, GPT-5.6 Sol on high effort
- 96 · Rank 2"The Horizon Legacy Secured" after 37 turns; second behind Grok
- DeepSeek v4 Pro on high effort
- 88 · Rank 6"The Open Epoch Achieved" after 48 turns
- Score to beat before the runs
- 94the board's top score; an earlier DeepSeek v4 Flash run of mine had reached 92
- Zaps spent: Grok, Codex, DeepSeek
- 308 · 518 · 672of 770 each, at 14 Zaps a turn
- Play time: Grok, Codex, DeepSeek
- ≈25 min · 1:16:34 · 1:42:48Grok's is the span between its opening and victory screenshots; the others run from the start of play to the ending
- Strategy package
- 1, frozen and hash-checkedbuilt only from material that existed before Grok played
- Strategy messages from me after play began
- 0DeepSeek got one procedural reply at turn 27: the run had not ended, keep going
Verdict
Grok 98, Codex 96, DeepSeek 88
Tier list A strategy game played once per agent, so it moves no agent on the coding tier list. It is linked from the Grok 4.6 / 4.7 entry, whose B placement is my own verdict.
danmanreal.com/tests/worldos-grok-codex-deepseek
What was tested
Related
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4