{ "_": "Empirically-validated coefficients for the cost simulator inside calc.ajinkya.ai. Each value is derived from running agent-cost-bench scenarios and measuring actual provider behavior. Versioned — the calc fetches this file at startup and uses it to override its hand-tuned defaults.", "schema_version": "1.1", "generated_at": "2026-05-06", "agent_cost_bench_version": "0.2.0", "coefficients": { "cache_hit_rate": { "value": 0.91, "previous_value": 0.84, "unit": "fraction", "description": "Median fraction of input tokens served from prompt cache on a typical multi-turn workload with a >1024-token shared system prompt.", "source_scenario": "cached-pipeline", "sample_size": 18, "provider": "openai", "model": "gpt-4o-mini", "measured_range": {"p10": 0.61, "median": 0.94, "p90": 0.97}, "by_provider": { "openai_auto_prefix": {"value": 0.91, "model": "gpt-4o-mini", "sample": 18, "scenario": "cached-pipeline"}, "anthropic_explicit": {"value": 0.77, "model": "claude-sonnet-4-5-20250929", "sample": 18, "scenario": "cached-pipeline-anthropic"}, "parallel_fan_out": {"value": 0.60, "model": "gpt-4o-mini", "sample": 40, "scenario": "parallel-fan-out"} }, "notes": "Cold-start (turn 0) caches only ~61% (system prompt). Warm turns 1+ hold 94-97%. Provider behavior differs materially: OpenAI's automatic prefix-matching beats Anthropic's explicit cache_control on multi-turn chat patterns where conversation history grows. the simulator should model cache as a curve, not a flat coefficient — phase-2 enhancement." }, "input_output_ratio": { "value": 73, "previous_value": 6, "unit": "ratio", "description": "Median ratio of input tokens to output tokens for confirmation-gated tool-orchestration agents (e.g., EIE-shape data-discovery pipelines).", "source_scenario": "data-discovery", "sample_size": 33, "provider": "openai", "model": "gpt-4o-mini", "measured_range": {"observed": 73, "gpt5_2_observed": 88}, "notes": "the simulator's 6x default applies to chat-style agents. For agents with output-suppression rules + tool-state bypass + long sysprompts, the ratio jumps an order of magnitude. Single coefficient is wrong; calc should branch on agent topology." }, "sequential_handoff_overhead_tokens": { "value": 700, "previous_value": 200, "unit": "tokens_per_stage", "description": "Per-stage cumulative-context growth in a sequential multi-agent pipeline. Each downstream stage sees prior stages' outputs concatenated.", "source_scenario": "multi-stage-research", "sample_size": 15, "provider": "openai", "model": "gpt-4o-mini", "measured_range": {"min": 600, "median": 700, "max": 800}, "notes": "Flat overhead constant is wrong; real growth is roughly linear in number of upstream stages times their average output size." }, "parallel_fan_out_synthesizer_input_tokens": { "value": 1900, "previous_value": null, "unit": "tokens", "description": "Input to a fan-in synthesizer that merges outputs from N parallel specialists. Grows with N.", "source_scenario": "parallel-fan-out", "sample_size": 40, "provider": "openai", "model": "gpt-4o-mini", "measured_range": {"specialist_input": 250, "synthesizer_input": 1900, "ratio": "7.6x"}, "notes": "Parallel topology has different cost shape from sequential: specialists each see only the orchestrator's output (~250 tok); the synthesizer at the end sees all specialist outputs concatenated (~1900 tok with 3 specialists)." }, "median_latency_ms": { "value": 1700, "previous_value": null, "unit": "milliseconds", "description": "Median per-call latency on gpt-4o-mini (blocking, non-streaming).", "source_scenario": "multiple", "sample_size": 116, "provider": "openai", "model": "gpt-4o-mini", "measured_range": {"p50": 1224, "p90": 3437, "max": 26326}, "notes": "the simulator doesn't currently model latency — bench introduces this as a new tracked coefficient. p90 latency is what production teams should plan for." }, "time_to_first_token_ms": { "value": 873, "previous_value": null, "unit": "milliseconds", "description": "Median time-to-first-token on streaming gpt-4o-mini calls.", "source_scenario": "streaming-pipeline", "sample_size": 9, "provider": "openai", "model": "gpt-4o-mini", "measured_range": {"p50": 873, "p90": 4956}, "notes": "Production-relevant for agent UX — users perceive responsiveness from TTFT, not total latency." }, "output_rate_tokens_per_sec": { "value": 47, "previous_value": null, "unit": "tokens/sec", "description": "Median streaming output rate on gpt-4o-mini.", "source_scenario": "streaming-pipeline", "sample_size": 9, "provider": "openai", "model": "gpt-4o-mini", "measured_range": {"p50": 47, "max": 77}, "notes": "Token throughput from streaming — useful for capacity-planning queue depth." } }, "scenarios_run": [ {"name": "smoke-test", "calls": 9, "input_tokens": 489, "output_tokens": 99, "cost_usd": 0.0001, "provider": "openai"}, {"name": "multi-stage-research", "calls": 15, "input_tokens": 19016, "output_tokens": 8601, "cost_usd": 0.005, "provider": "openai"}, {"name": "streaming-pipeline", "calls": 9, "input_tokens": 2536, "output_tokens": 1579, "cost_usd": 0.001, "provider": "openai"}, {"name": "tool-chain", "calls": 21, "input_tokens": 54082, "output_tokens": 5718, "cost_usd": 0.011, "provider": "openai"}, {"name": "data-discovery", "calls": 33, "input_tokens": 56409, "output_tokens": 772, "cost_usd": 0.009, "provider": "openai"}, {"name": "data-discovery-gpt52", "calls": 11, "input_tokens": 19024, "output_tokens": 215, "cost_usd": 0.098, "provider": "openai"}, {"name": "cached-pipeline", "calls": 18, "input_tokens": 39375, "output_tokens": 3287, "cost_usd": 0.006, "provider": "openai", "cached_tokens": 35840, "cache_hit_rate": 0.910}, {"name": "cached-pipeline-anthropic", "calls": 18, "input_tokens": 44069, "output_tokens": 3600, "cost_usd": 0.063, "provider": "anthropic", "cached_tokens": 34102, "cache_hit_rate": 0.774}, {"name": "parallel-fan-out", "calls": 40, "input_tokens": 142855, "output_tokens": 16251, "cost_usd": 0.030, "provider": "openai", "cached_tokens": 85248, "cache_hit_rate": 0.597} ], "totals": { "calls": 174, "input_tokens": 377855, "output_tokens": 40122, "cost_usd": 0.224, "providers": ["openai", "anthropic"], "models": ["gpt-4o-mini", "gpt-5.2", "claude-sonnet-4-5-20250929"] } }