{
 "as_of": "2026-10-28",
 "screen": [
  1280,
  800
 ],
 "steps_per_trajectory": 89,
 "catalog": {
  "procedures": 52,
  "episodes": 156,
  "verified": 156,
  "frames_mean": 90.58,
  "steps_mean": 89.33
 },
 "cloud_cpu": {
  "label": "AWS c8g (Graviton4), on-demand, us-east-1",
  "usd_per_vcpu_hour": 0.03988,
  "sources": [
   {
    "title": "Amazon EC2 On-Demand pricing",
    "publisher": "AWS",
    "date": "price file of 2026-10-28",
    "url": "https://aws.amazon.com/ec2/pricing/on-demand/",
    "says": "c8g.large, 2 vCPU, $0.07976 an hour in us-east-1 (Linux)",
    "key": "aws-ec2"
   }
  ]
 },
 "token_model": {
  "screen": [
   1280,
   800
  ],
  "fixed_prompt_tokens": 1500,
  "images_in_context": 3,
  "history_per_step": 100,
  "output_per_step": 150,
  "environment_seconds_per_step": 0,
  "steps": "the catalog's measured mean actions per trajectory, taken as the agent's steps (1x: no wasted steps)",
  "latency": "the measured mean DeepSeek call latency, given to every agent as its seconds per step",
  "notes": [
   "Every step re-reads a fixed prompt (system prompt, tool definitions, task) of 1,500 tokens plus the provider's own computer-use overhead, the last three screenshots, and 100 tokens of history for every earlier step; it writes 150 tokens (the DeepSeek screenshot row instead writes the measured sample's mean, the same model). What was already sent the step before is priced at the provider's cached-input rate.",
   "Agents are given the expert's own action count. OSWorld-Human measured the best agents taking 1.4-2.7x (v1) and 2.7-4.3x (v2) more steps than necessary; using 1x understates their cost and time.",
   "Seconds per step is the mean latency measured for DeepSeek's flash model on text observations, with no time charged for the environment to take a screenshot or act. Larger models reading images are slower, so the agent times are lower bounds.",
   "Success rates are the agents' published OSWorld figures. A verified trajectory also needs a verifier; the agent rows assume one exists at no cost."
  ]
 },
 "headline": {
  "cw_usd_per_1k": 0.08015,
  "cw_seconds": 7.24,
  "cheapest_agent": "UI-TARS-1.5-7B, hosted",
  "cheapest_agent_usd_per_1k": 305.4,
  "cost_ratio": 3810.7
 },
 "rows": [
  {
   "id": "cw-native",
   "name": "ComputerWorld, native",
   "kind": "measured",
   "family": "computerworld",
   "attempts": 156,
   "verified": 156,
   "success_rate": 1.0,
   "steps_mean": 89.33,
   "frames_mean": 90.58,
   "seconds_per_trajectory": 7.24,
   "wall_mean": 7.24,
   "wall_median": 6.35,
   "cpu_seconds_per_verified": 7.24,
   "phases_cpu_mean": {
    "boot": 0.006442,
    "project": 0.4578,
    "render": 5.5,
    "encode": 1.27
   },
   "usd_per_attempt": 8.015e-05,
   "usd_per_1k_verified": 0.08015,
   "note": "Python binding, one core, frames rendered and PNG-encoded",
   "procedures": 52,
   "host": {
    "machine": "aarch64",
    "processor": "aarch64",
    "python": "3.12.3",
    "node": "promaxgb10-4dfb"
   },
   "engine": "0.3.0",
   "runtime": null,
   "sources": [
    {
     "title": "Amazon EC2 On-Demand pricing",
     "publisher": "AWS",
     "date": "price file of 2026-10-28",
     "url": "https://aws.amazon.com/ec2/pricing/on-demand/",
     "says": "c8g.large, 2 vCPU, $0.07976 an hour in us-east-1 (Linux)",
     "key": "aws-ec2"
    }
   ]
  },
  {
   "id": "cw-wasm",
   "name": "ComputerWorld, Wasm",
   "kind": "measured",
   "family": "computerworld",
   "attempts": 156,
   "verified": 156,
   "success_rate": 1.0,
   "steps_mean": 89.33,
   "frames_mean": 90.58,
   "seconds_per_trajectory": 19.09,
   "wall_mean": 19.09,
   "wall_median": 16.0,
   "cpu_seconds_per_verified": 19.09,
   "phases_cpu_mean": {
    "boot": 0.1097,
    "project": 1.88,
    "render": 17.1
   },
   "usd_per_attempt": 0.0002115,
   "usd_per_1k_verified": 0.2115,
   "note": "Wasm engine in Node, one core, frames rendered (PNG encode not included)",
   "procedures": 52,
   "host": {
    "machine": "arm64",
    "cpu": "unknown",
    "node": "promaxgb10-4dfb"
   },
   "engine": "0.3.0",
   "runtime": "node v22.22.2 (wasm32, pkg/node)",
   "sources": [
    {
     "title": "Amazon EC2 On-Demand pricing",
     "publisher": "AWS",
     "date": "price file of 2026-10-28",
     "url": "https://aws.amazon.com/ec2/pricing/on-demand/",
     "says": "c8g.large, 2 vCPU, $0.07976 an hour in us-east-1 (Linux)",
     "key": "aws-ec2"
    }
   ]
  },
  {
   "id": "deepseek-measured",
   "name": "DeepSeek flash, text screen",
   "kind": "measured",
   "family": "agent",
   "model": "deepseek-flash (DeepSeek-V4.1-Flash)",
   "attempts": 16,
   "verified": 1,
   "success_rate": 0.0625,
   "calls": 1619,
   "latency_per_step": 3.34,
   "latency_median": 2.45,
   "tokens_in_per_call": 1437.8,
   "tokens_out_per_call": 366.2,
   "usd_total": 1.33,
   "usd_per_attempt": 0.08302,
   "usd_per_1k_verified": 1328.3,
   "usd_per_1k_verified_at_least": null,
   "seconds_per_attempt": 339.5,
   "seconds_per_trajectory": 5432.1,
   "seconds_per_trajectory_at_least": null,
   "checks_passed_share": 0.1032,
   "per_procedure": [
    {
     "procedure": "accounting-card-reconciliation",
     "seed": 1,
     "expert_actions": 24,
     "budget": 72,
     "agent_turns": 22,
     "ended": "done",
     "passed": true,
     "checks_passed": 6,
     "checks": 6,
     "wall_seconds": 65.74
    },
    {
     "procedure": "accounting-card-reconciliation",
     "seed": 2,
     "expert_actions": 24,
     "budget": 72,
     "agent_turns": 72,
     "ended": "budget",
     "passed": false,
     "checks_passed": 5,
     "checks": 6,
     "wall_seconds": 250.2
    },
    {
     "procedure": "cad-mounting-plate",
     "seed": 1,
     "expert_actions": 86,
     "budget": 150,
     "agent_turns": 145,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 16,
     "wall_seconds": 694.9
    },
    {
     "procedure": "cad-mounting-plate",
     "seed": 2,
     "expert_actions": 86,
     "budget": 150,
     "agent_turns": 89,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 16,
     "wall_seconds": 313.8
    },
    {
     "procedure": "eda-voltage-divider",
     "seed": 1,
     "expert_actions": 99,
     "budget": 150,
     "agent_turns": 147,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 34,
     "wall_seconds": 433.0
    },
    {
     "procedure": "eda-voltage-divider",
     "seed": 2,
     "expert_actions": 99,
     "budget": 150,
     "agent_turns": 147,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 34,
     "wall_seconds": 520.6
    },
    {
     "procedure": "legal-engagement",
     "seed": 1,
     "expert_actions": 34,
     "budget": 102,
     "agent_turns": 101,
     "ended": "budget",
     "passed": false,
     "checks_passed": 1,
     "checks": 3,
     "wall_seconds": 265.6
    },
    {
     "procedure": "legal-engagement",
     "seed": 2,
     "expert_actions": 34,
     "budget": 102,
     "agent_turns": 57,
     "ended": "budget",
     "passed": false,
     "checks_passed": 1,
     "checks": 3,
     "wall_seconds": 196.1
    },
    {
     "procedure": "media-image-web-export",
     "seed": 1,
     "expert_actions": 37,
     "budget": 111,
     "agent_turns": 111,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 421.8
    },
    {
     "procedure": "media-image-web-export",
     "seed": 2,
     "expert_actions": 37,
     "budget": 111,
     "agent_turns": 111,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 302.7
    },
    {
     "procedure": "office-calc-opex-variance",
     "seed": 1,
     "expert_actions": 53,
     "budget": 150,
     "agent_turns": 149,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 347.8
    },
    {
     "procedure": "office-calc-opex-variance",
     "seed": 2,
     "expert_actions": 53,
     "budget": 150,
     "agent_turns": 44,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 268.7
    },
    {
     "procedure": "office-code-change-control",
     "seed": 1,
     "expert_actions": 90,
     "budget": 150,
     "agent_turns": 145,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 494.4
    },
    {
     "procedure": "office-code-change-control",
     "seed": 2,
     "expert_actions": 90,
     "budget": 150,
     "agent_turns": 150,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 397.8
    },
    {
     "procedure": "office-sql-key-accounts",
     "seed": 1,
     "expert_actions": 41,
     "budget": 123,
     "agent_turns": 22,
     "ended": "done",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 69.81
    },
    {
     "procedure": "office-sql-key-accounts",
     "seed": 2,
     "expert_actions": 41,
     "budget": 123,
     "agent_turns": 107,
     "ended": "budget",
     "passed": false,
     "checks_passed": 0,
     "checks": 1,
     "wall_seconds": 389.2
    }
   ],
   "note": "Measured: real API calls on 8 one-person catalog procedures x 2 seeds, the screen given as text (windows, numbered controls, painted text) and clicks grounded on control numbers, a budget of 3x the expert's actions, judged by each procedure's checks. Priced at DeepSeek's peak rates.",
   "sources": [
    {
     "title": "Models and pricing",
     "publisher": "DeepSeek",
     "date": "accessed 2026-10-28",
     "url": "https://api-docs.deepseek.com/quick_start/pricing",
     "says": "deepseek-flash peak $0.30 input (cache miss), $0.006 (cache hit), $1.20 output",
     "key": "deepseek-pricing"
    }
   ]
  },
  {
   "id": "claude-opus",
   "name": "Claude Opus 5.5",
   "kind": "estimate",
   "family": "agent",
   "model": "claude-opus-5-5, computer_toolset_20260801",
   "steps_mean": 89,
   "tokens_in_per_attempt": 1277776,
   "tokens_out_per_attempt": 13350,
   "image_tokens": 1334,
   "usd_per_attempt": 1.03,
   "success_rate": 0.818,
   "success_basis": "OSWorld 2.1 (partial), Anthropic launch post",
   "usd_per_1k_verified": 1259.6,
   "seconds_per_trajectory": 363.5,
   "seconds_per_attempt": 297.3,
   "latency_per_step": 3.34,
   "note": "",
   "sources": [
    {
     "title": "Claude pricing",
     "publisher": "Anthropic",
     "date": "accessed 2026-10-28",
     "url": "https://platform.claude.com/docs/en/about-claude/pricing",
     "says": "Opus 5.5 $4 / $20, cache read $0.20; Sonnet 5.5 $2 / $10, cache read $0.10; computer_toolset_20260801 adds about 4,500 input tokens a request",
     "key": "anthropic-pricing"
    },
    {
     "title": "Vision",
     "publisher": "Anthropic",
     "date": "accessed 2026-10-28",
     "url": "https://platform.claude.com/docs/en/build-with-claude/vision",
     "says": "an image costs ceil(w/28) x ceil(h/28) tokens: 1,334 at 1280x800",
     "key": "anthropic-vision"
    },
    {
     "title": "Claude Opus 5.5",
     "publisher": "Anthropic",
     "date": "2026-09-22",
     "url": "https://www.anthropic.com/news/claude-opus-5-5",
     "says": "81.8% on OSWorld 2.1 (partial)",
     "key": "anthropic-opus-55"
    },
    {
     "title": "OSWorld-Human: benchmarking the efficiency of computer-use agents",
     "publisher": "arXiv 2506.16042",
     "date": "v2 2026-05-18",
     "url": "https://arxiv.org/abs/2506.16042",
     "says": "the best agents take 1.4-2.7x (v1), 2.7-4.3x (v2) more steps than necessary; agents take tens of minutes a task",
     "key": "osworld-human"
    }
   ]
  },
  {
   "id": "claude-sonnet",
   "name": "Claude Sonnet 5.5",
   "kind": "estimate",
   "family": "agent",
   "model": "claude-sonnet-5-5, computer_toolset_20260801",
   "steps_mean": 89,
   "tokens_in_per_attempt": 1277776,
   "tokens_out_per_attempt": 13350,
   "image_tokens": 1334,
   "usd_per_attempt": 0.5152,
   "success_rate": 0.801,
   "success_basis": "OSWorld 2.1 (partial), Anthropic launch post",
   "usd_per_1k_verified": 643.2,
   "seconds_per_trajectory": 371.2,
   "seconds_per_attempt": 297.3,
   "latency_per_step": 3.34,
   "note": "",
   "sources": [
    {
     "title": "Claude pricing",
     "publisher": "Anthropic",
     "date": "accessed 2026-10-28",
     "url": "https://platform.claude.com/docs/en/about-claude/pricing",
     "says": "Opus 5.5 $4 / $20, cache read $0.20; Sonnet 5.5 $2 / $10, cache read $0.10; computer_toolset_20260801 adds about 4,500 input tokens a request",
     "key": "anthropic-pricing"
    },
    {
     "title": "Vision",
     "publisher": "Anthropic",
     "date": "accessed 2026-10-28",
     "url": "https://platform.claude.com/docs/en/build-with-claude/vision",
     "says": "an image costs ceil(w/28) x ceil(h/28) tokens: 1,334 at 1280x800",
     "key": "anthropic-vision"
    },
    {
     "title": "Claude Sonnet 5.5",
     "publisher": "Anthropic",
     "date": "2026-09-28",
     "url": "https://www.anthropic.com/news/claude-sonnet-5-5",
     "says": "80.1% on OSWorld 2.1 (partial)",
     "key": "anthropic-sonnet-55"
    },
    {
     "title": "OSWorld-Human: benchmarking the efficiency of computer-use agents",
     "publisher": "arXiv 2506.16042",
     "date": "v2 2026-05-18",
     "url": "https://arxiv.org/abs/2506.16042",
     "says": "the best agents take 1.4-2.7x (v1), 2.7-4.3x (v2) more steps than necessary; agents take tens of minutes a task",
     "key": "osworld-human"
    }
   ]
  },
  {
   "id": "openai-sol",
   "name": "OpenAI GPT-6.1 Sol",
   "kind": "estimate",
   "family": "agent",
   "model": "gpt-6.1-sol, computer_use tool",
   "steps_mean": 89,
   "tokens_in_per_attempt": 841900,
   "tokens_out_per_attempt": 13350,
   "image_tokens": 1200,
   "usd_per_attempt": 0.4404,
   "success_rate": 0.714,
   "success_basis": "OSWorld 2.0, third-party reports of OpenAI's launch figure (unverified); not on OSWorld-Verified",
   "usd_per_1k_verified": 616.8,
   "seconds_per_trajectory": 416.4,
   "seconds_per_attempt": 297.3,
   "latency_per_step": 3.34,
   "note": "",
   "sources": [
    {
     "title": "gpt-6.1-sol model",
     "publisher": "OpenAI",
     "date": "accessed 2026-10-28",
     "url": "https://developers.openai.com/api/docs/models/gpt-6.1-sol",
     "says": "$2 / $10, cached $0.10; supports the computer_use tool",
     "key": "openai-sol"
    },
    {
     "title": "Images and vision",
     "publisher": "OpenAI",
     "date": "accessed 2026-10-28",
     "url": "https://developers.openai.com/api/docs/guides/images-vision",
     "says": "ceil(w/32) x ceil(h/32) patches x 1.2 for gpt-6: 1,200 tokens at 1280x800",
     "key": "openai-vision"
    },
    {
     "title": "OSWorld-Human: benchmarking the efficiency of computer-use agents",
     "publisher": "arXiv 2506.16042",
     "date": "v2 2026-05-18",
     "url": "https://arxiv.org/abs/2506.16042",
     "says": "the best agents take 1.4-2.7x (v1), 2.7-4.3x (v2) more steps than necessary; agents take tens of minutes a task",
     "key": "osworld-human"
    }
   ]
  },
  {
   "id": "gemini-flash",
   "name": "Gemini 3.8 Flash",
   "kind": "estimate",
   "family": "agent",
   "model": "gemini-3.8-flash, computer use",
   "steps_mean": 89,
   "tokens_in_per_attempt": 820780,
   "tokens_out_per_attempt": 13350,
   "image_tokens": 1120,
   "usd_per_attempt": 0.1859,
   "success_rate": 0.59,
   "success_basis": "OSWorld 2.0, one third-party report (unverified); not on OSWorld-Verified",
   "usd_per_1k_verified": 315.1,
   "seconds_per_trajectory": 503.9,
   "seconds_per_attempt": 297.3,
   "latency_per_step": 3.34,
   "note": "",
   "sources": [
    {
     "title": "Gemini API pricing",
     "publisher": "Google",
     "date": "accessed 2026-10-28",
     "url": "https://ai.google.dev/gemini-api/docs/pricing",
     "says": "gemini-3.8-flash $0.75 / $3.75, cached $0.075, until 2026-12-31; computer use charged as regular tokens",
     "key": "gemini-pricing"
    },
    {
     "title": "Media resolution",
     "publisher": "Google",
     "date": "accessed 2026-10-28",
     "url": "https://ai.google.dev/gemini-api/docs/media-resolution",
     "says": "1,120 tokens an image at the default resolution",
     "key": "gemini-media-resolution"
    },
    {
     "title": "Computer use",
     "publisher": "Google",
     "date": "accessed 2026-10-28",
     "url": "https://ai.google.dev/gemini-api/docs/computer-use",
     "says": "Gemini 3.8 Flash is the recommended model for computer use",
     "key": "gemini-computer-use"
    },
    {
     "title": "OSWorld-Human: benchmarking the efficiency of computer-use agents",
     "publisher": "arXiv 2506.16042",
     "date": "v2 2026-05-18",
     "url": "https://arxiv.org/abs/2506.16042",
     "says": "the best agents take 1.4-2.7x (v1), 2.7-4.3x (v2) more steps than necessary; agents take tens of minutes a task",
     "key": "osworld-human"
    }
   ]
  },
  {
   "id": "ui-tars",
   "name": "UI-TARS-1.5-7B, hosted",
   "kind": "estimate",
   "family": "agent",
   "model": "bytedance/ui-tars-1.5-7b on OpenRouter",
   "steps_mean": 89,
   "tokens_in_per_attempt": 877276,
   "tokens_out_per_attempt": 13350,
   "image_tokens": 1334,
   "usd_per_attempt": 0.09041,
   "success_rate": 0.296,
   "success_basis": "OSWorld-Verified, 100 steps, best reported run",
   "usd_per_1k_verified": 305.4,
   "seconds_per_trajectory": 1004.4,
   "seconds_per_attempt": 297.3,
   "latency_per_step": 3.34,
   "note": "",
   "sources": [
    {
     "title": "OpenRouter models API",
     "publisher": "OpenRouter",
     "date": "accessed 2026-10-28",
     "url": "https://openrouter.ai/api/v1/models",
     "says": "bytedance/ui-tars-1.5-7b $0.10 / $0.20 per million tokens",
     "key": "openrouter-models"
    },
    {
     "title": "OSWorld-Verified results",
     "publisher": "XLANG Lab",
     "date": "accessed 2026-10-28",
     "url": "https://os-world.github.io/",
     "says": "UI-TARS-1.5-7B 25.2-29.6% at 100 steps; human 72.36%",
     "key": "osworld-verified"
    },
    {
     "title": "OSWorld-Human: benchmarking the efficiency of computer-use agents",
     "publisher": "arXiv 2506.16042",
     "date": "v2 2026-05-18",
     "url": "https://arxiv.org/abs/2506.16042",
     "says": "the best agents take 1.4-2.7x (v1), 2.7-4.3x (v2) more steps than necessary; agents take tens of minutes a task",
     "key": "osworld-human"
    }
   ]
  },
  {
   "id": "deepseek-vision",
   "name": "DeepSeek flash, screenshots",
   "kind": "estimate",
   "family": "agent",
   "model": "deepseek-flash with vision",
   "steps_mean": 89,
   "tokens_in_per_attempt": 795436,
   "tokens_out_per_attempt": 32574,
   "image_tokens": 1024,
   "usd_per_attempt": 0.07371,
   "success_rate": 0.0625,
   "success_basis": "the measured text-screen sample's pass rate on this catalog (no published OSWorld figure)",
   "usd_per_1k_verified": 1179.4,
   "seconds_per_trajectory": 4757.0,
   "seconds_per_attempt": 297.3,
   "latency_per_step": 3.34,
   "note": "Output tokens per step are the measured sample's mean, since it is the same model.",
   "sources": [
    {
     "title": "Models and pricing",
     "publisher": "DeepSeek",
     "date": "accessed 2026-10-28",
     "url": "https://api-docs.deepseek.com/quick_start/pricing",
     "says": "deepseek-flash peak $0.30 input (cache miss), $0.006 (cache hit), $1.20 output",
     "key": "deepseek-pricing"
    },
    {
     "title": "Vision",
     "publisher": "DeepSeek",
     "date": "accessed 2026-10-28",
     "url": "https://api-docs.deepseek.com/guides/vision",
     "says": "at most 1,024 tokens an image",
     "key": "deepseek-vision"
    },
    {
     "title": "OSWorld-Human: benchmarking the efficiency of computer-use agents",
     "publisher": "arXiv 2506.16042",
     "date": "v2 2026-05-18",
     "url": "https://arxiv.org/abs/2506.16042",
     "says": "the best agents take 1.4-2.7x (v1), 2.7-4.3x (v2) more steps than necessary; agents take tens of minutes a task",
     "key": "osworld-human"
    }
   ]
  },
  {
   "id": "human",
   "name": "Human annotators",
   "kind": "estimate",
   "family": "human",
   "steps_mean": 89,
   "usd_per_attempt": 4.23,
   "success_rate": 1.0,
   "success_basis": "every demonstration counted as verified",
   "usd_per_1k_verified": 4229.9,
   "seconds_per_trajectory": 1722.6,
   "seconds_per_attempt": 1722.6,
   "note": "OpenCUA's AgentNet: 22,625 desktop demonstrations for about $20,000 of annotation at about ten tasks an hour, 18.6 steps on average; scaled to the catalog's actions per trajectory.",
   "sources": [
    {
     "title": "OpenCUA: open foundations for computer-use agents",
     "publisher": "arXiv 2508.09123",
     "date": "2025-08",
     "url": "https://arxiv.org/abs/2508.09123",
     "says": "22,625 tasks, annotation about $20,000, about ten tasks an hour, 18.6 steps a demonstration",
     "key": "opencua"
    }
   ]
  }
 ],
 "notes": [
  "ComputerWorld rows are measured: every catalog procedure (52 procedures, seeds 1-3, 156 trajectories) performed through pointer and keyboard at 1280×800, every frame rendered before each action and after the last, the replay required to end on the projected state hash, and the procedure's checks judged; 156 of 156 verified natively and 156 of 156 in Wasm. Each run was pinned to one Cortex-X925 core; dollars are its CPU-seconds at AWS c8g (Graviton4), on-demand, us-east-1, $0.03988 a vCPU-hour. A Graviton core is slower than the X925, so read the dollars as per X925-second; at three times the CPU time they would still be cents per thousand.",
  "The DeepSeek sample is measured: 16 attempts at one-person catalog procedures, 1619 real API calls, tokens as the API reported them, priced at DeepSeek's peak rates. It saw the screen as text with numbered controls, which spares it finding pixels, and had three times the expert's actions (at most 150); 1 of 16 attempts passed every check, and 10% of all checks passed. Its mean call took 3.34 s and wrote 366 tokens.",
  "Every step re-reads a fixed prompt (system prompt, tool definitions, task) of 1,500 tokens plus the provider's own computer-use overhead, the last three screenshots, and 100 tokens of history for every earlier step; it writes 150 tokens (the DeepSeek screenshot row instead writes the measured sample's mean, the same model). What was already sent the step before is priced at the provider's cached-input rate.",
  "Agents are given the expert's own action count. OSWorld-Human measured the best agents taking 1.4-2.7x (v1) and 2.7-4.3x (v2) more steps than necessary; using 1x understates their cost and time.",
  "Seconds per step is the mean latency measured for DeepSeek's flash model on text observations, with no time charged for the environment to take a screenshot or act. Larger models reading images are slower, so the agent times are lower bounds.",
  "Success rates are the agents' published OSWorld figures. A verified trajectory also needs a verifier; the agent rows assume one exists at no cost.",
  "Human annotators: OpenCUA's AgentNet: 22,625 desktop demonstrations for about $20,000 of annotation at about ten tasks an hour, 18.6 steps on average; scaled to the catalog's actions per trajectory."
 ]
}