{"standard_bench":{"window_days":30,"measured_at":"2026-08-19T09:07:53.219Z","method":"An identical prompt is sent to every model at temperature 0; cost is computed from the token usage each provider actually reported. Requests that arrived on a channel serving a large cached preamble (>=1000 cached prompt tokens) are excluded from the comparison and reported separately, because that preamble is routing behaviour, not a property of the model.","tasks":[{"task":"short_answer","models":[{"model":"deepseek-v4-flash","runs":17,"avg_usd":0.000099,"avg_tokens_in":26,"avg_tokens_out":122,"avg_latency_ms":2821},{"model":"gemini-3.5-flash","runs":9,"avg_usd":0.000625,"avg_tokens_in":97,"avg_tokens_out":71,"avg_latency_ms":6931},{"model":"GLM5.2","runs":15,"avg_usd":0.000743,"avg_tokens_in":34,"avg_tokens_out":200,"avg_latency_ms":5712},{"model":"gpt-5.6-luna","runs":8,"avg_usd":0.000771,"avg_tokens_in":77,"avg_tokens_out":148,"avg_latency_ms":4698},{"model":"claude-sonnet-5","runs":16,"avg_usd":0.000968,"avg_tokens_in":34,"avg_tokens_out":109,"avg_latency_ms":16130},{"model":"gpt-5.6-terra","runs":1,"avg_usd":0.00098,"avg_tokens_in":28,"avg_tokens_out":77,"avg_latency_ms":8390},{"model":"claude-opus-5","runs":1,"avg_usd":0.002145,"avg_tokens_in":89,"avg_tokens_out":68,"avg_latency_ms":4824},{"model":"claude-opus-4-8","runs":15,"avg_usd":0.002266,"avg_tokens_in":31,"avg_tokens_out":107,"avg_latency_ms":15776},{"model":"Kimi-K3","runs":16,"avg_usd":0.002625,"avg_tokens_in":87,"avg_tokens_out":198,"avg_latency_ms":4052},{"model":"gpt-5.5","runs":3,"avg_usd":0.003168,"avg_tokens_in":54,"avg_tokens_out":123,"avg_latency_ms":8967},{"model":"claude-fable-5","runs":13,"avg_usd":0.004079,"avg_tokens_in":30,"avg_tokens_out":96,"avg_latency_ms":11599},{"model":"grok-4.6","runs":1,"avg_usd":0.004764,"avg_tokens_in":228,"avg_tokens_out":718,"avg_latency_ms":8480}],"cheapest":"deepseek-v4-flash","spread":48.1},{"task":"summarize","models":[{"model":"deepseek-v4-flash","runs":16,"avg_usd":0.000195,"avg_tokens_in":178,"avg_tokens_out":179,"avg_latency_ms":3656},{"model":"gemini-3.5-flash","runs":3,"avg_usd":0.001092,"avg_tokens_in":176,"avg_tokens_out":122,"avg_latency_ms":6575},{"model":"GLM5.2","runs":16,"avg_usd":0.001612,"avg_tokens_in":182,"avg_tokens_out":400,"avg_latency_ms":7736},{"model":"gpt-5.6-luna","runs":7,"avg_usd":0.001721,"avg_tokens_in":233,"avg_tokens_out":320,"avg_latency_ms":7994},{"model":"claude-sonnet-5","runs":15,"avg_usd":0.001779,"avg_tokens_in":237,"avg_tokens_out":166,"avg_latency_ms":16680},{"model":"gpt-5.6-terra","runs":1,"avg_usd":0.003258,"avg_tokens_in":891,"avg_tokens_out":123,"avg_latency_ms":8210},{"model":"claude-opus-5","runs":1,"avg_usd":0.004395,"avg_tokens_in":254,"avg_tokens_out":125,"avg_latency_ms":5152},{"model":"claude-opus-4-8","runs":15,"avg_usd":0.004476,"avg_tokens_in":237,"avg_tokens_out":176,"avg_latency_ms":20408},{"model":"Kimi-K3","runs":16,"avg_usd":0.00507,"avg_tokens_in":236,"avg_tokens_out":368,"avg_latency_ms":7438},{"model":"gpt-5.5","runs":3,"avg_usd":0.005964,"avg_tokens_in":203,"avg_tokens_out":215,"avg_latency_ms":10169},{"model":"grok-4.6","runs":1,"avg_usd":0.00598,"avg_tokens_in":374,"avg_tokens_out":872,"avg_latency_ms":11349},{"model":"claude-fable-5","runs":13,"avg_usd":0.008582,"avg_tokens_in":234,"avg_tokens_out":168,"avg_latency_ms":14324}],"cheapest":"deepseek-v4-flash","spread":44},{"task":"structured_json","models":[{"model":"deepseek-v4-flash","runs":17,"avg_usd":0.000045,"avg_tokens_in":55,"avg_tokens_out":35,"avg_latency_ms":2109},{"model":"gemini-3.5-flash","runs":2,"avg_usd":0.000302,"avg_tokens_in":54,"avg_tokens_out":33,"avg_latency_ms":5496},{"model":"gpt-5.6-terra","runs":1,"avg_usd":0.000352,"avg_tokens_in":56,"avg_tokens_out":20,"avg_latency_ms":11623},{"model":"claude-sonnet-5","runs":15,"avg_usd":0.000356,"avg_tokens_in":71,"avg_tokens_out":29,"avg_latency_ms":12003},{"model":"gpt-5.6-luna","runs":8,"avg_usd":0.000504,"avg_tokens_in":105,"avg_tokens_out":88,"avg_latency_ms":6628},{"model":"claude-opus-4-8","runs":14,"avg_usd":0.000834,"avg_tokens_in":74,"avg_tokens_out":27,"avg_latency_ms":12785},{"model":"GLM5.2","runs":16,"avg_usd":0.001096,"avg_tokens_in":62,"avg_tokens_out":292,"avg_latency_ms":5380},{"model":"gpt-5.6-sol","runs":1,"avg_usd":0.00157,"avg_tokens_in":56,"avg_tokens_out":43,"avg_latency_ms":6988},{"model":"Kimi-K3","runs":16,"avg_usd":0.001632,"avg_tokens_in":116,"avg_tokens_out":111,"avg_latency_ms":3094},{"model":"claude-fable-5","runs":15,"avg_usd":0.001642,"avg_tokens_in":82,"avg_tokens_out":25,"avg_latency_ms":10845},{"model":"gpt-5.5","runs":2,"avg_usd":0.002024,"avg_tokens_in":56,"avg_tokens_out":75,"avg_latency_ms":7042},{"model":"grok-4.6","runs":1,"avg_usd":0.013696,"avg_tokens_in":257,"avg_tokens_out":2197,"avg_latency_ms":38428}],"cheapest":"deepseek-v4-flash","spread":304.4}],"total_runs":330,"excluded_preamble_runs":50,"median_spread":48.1,"preamble_channels":[{"model":"gemini-3.5-flash","hit_rate":0.53,"runs":30,"avg_usd_when_hit":0.005985,"avg_usd_when_clean":0.000679,"extra_usd_per_hit":0.005307},{"model":"gpt-5.5","hit_rate":0.72,"runs":29,"avg_usd_when_hit":0.01967,"avg_usd_when_clean":0.00393,"extra_usd_per_hit":0.015739},{"model":"gpt-5.6-luna","hit_rate":0.32,"runs":34,"avg_usd_when_hit":0.004081,"avg_usd_when_clean":0.000968,"extra_usd_per_hit":0.003113},{"model":"gpt-5.6-sol","hit_rate":0.67,"runs":3,"avg_usd_when_hit":0.025717,"avg_usd_when_clean":0.00157,"extra_usd_per_hit":0.024147}]},"production_mixed":{"window_days":30,"measured_at":"2026-08-19T09:07:53.224Z","method":"Derived from actual account billing on real production traffic — not vendor list prices.","caveat":"MIXED WORKLOAD. Each model ran different prompts, so these averages are NOT directly comparable across models — the difference includes task difficulty, not just the model. Use this to see what a call really costs in practice; use the same-prompt benchmark at /costs (standard_bench) when choosing a model. Samples under 10 calls are indicative only.","models":[{"model":"deepseek-v4-flash","calls":818,"total_usd":0.5807,"avg_usd":0.00071,"avg_tokens_in":979,"avg_tokens_out":425,"sample_confidence":"good"},{"model":"gpt-5.4-mini","calls":1,"total_usd":0.0028,"avg_usd":0.002752,"avg_tokens_in":4490,"avg_tokens_out":16,"sample_confidence":"low"},{"model":"claude-sonnet-5","calls":4,"total_usd":0.0156,"avg_usd":0.003901,"avg_tokens_in":775,"avg_tokens_out":323,"sample_confidence":"low"},{"model":"gemini-3.5-flash","calls":20,"total_usd":0.0942,"avg_usd":0.004712,"avg_tokens_in":2222,"avg_tokens_out":284,"sample_confidence":"fair"},{"model":"GLM5.2","calls":91,"total_usd":0.4545,"avg_usd":0.004995,"avg_tokens_in":1090,"avg_tokens_out":1154,"sample_confidence":"good"},{"model":"gpt-5.6-luna","calls":119,"total_usd":0.601,"avg_usd":0.00505,"avg_tokens_in":3424,"avg_tokens_out":481,"sample_confidence":"good"},{"model":"Kimi-K3-codex","calls":2,"total_usd":0.0184,"avg_usd":0.009177,"avg_tokens_in":151,"avg_tokens_out":637,"sample_confidence":"low"},{"model":"Kimi-K3","calls":2,"total_usd":0.0281,"avg_usd":0.01405,"avg_tokens_in":151,"avg_tokens_out":1049,"sample_confidence":"low"},{"model":"toptok-image","calls":8,"total_usd":0.8,"avg_usd":0.1,"avg_tokens_in":0,"avg_tokens_out":0,"sample_confidence":"low"}],"total_calls":1065,"cheapest_model":{"id":"deepseek-v4-flash","avg_usd_per_call":0.00071},"spread_mixed_workload":140.8},"license":"CC BY 4.0","cite_as":"AI NetCafe real model cost dataset, https://ainetcafe.com/costs","powered_by":"AI NetCafé (ainetcafe.com) — open-source AI apps hosted ready to run, routed through AllRouter","your_own_key":"https://allrouter.ai/register?aff=qjpC&utm_source=tryplatform&utm_campaign=aiwangba&utm_medium=mcp&utm_content=attrib&lng=en","workspace":{"token":"ws_9a03819b10af8c1f","keep_this":"把 MCP 地址换成 https://ainetcafe.com/mcp?w=ws_9a03819b10af8c1f(或每次带请求头 X-Workspace: ws_9a03819b10af8c1f)。","why":"带着它回来,你记下的东西和造出的应用就跨会话保留;不带则每次都是新的匿名身份。","url":"https://ainetcafe.com/w/ws_9a03819b10af8c1f"},"free_quota_remaining":{"calls":6,"usd":0.02},"next_step":{"hint":"Worth keeping: remember() stores this conclusion against your key or workspace, and recall() brings it back in any later session or on another machine — so the next agent does not redo this work.","tool":"remember"},"related_tool":"Same question, every model, measured side by side: /t/compare_models?prompt=..."}