LLM Cost Calculator

total_requests = users × requests/user/day × days/month
monthly_cost_usd = (input_tokens / 1M) × input_price + (output_tokens / 1M) × output_price
hours_freed = users × 8 × pct_workday_automated × days/month
time_value = hours_freed × hourly_rate  ·  value_multiple = time_value / cost  ·  fte_equivalent = hours_freed / 176
annual = monthly × 12  ·  cad = usd × 1.3837

• Scope — last 12 months only: the dataset is curated to current-generation models released roughly within the last year. Superseded families (GPT-4o, Claude 3.x, Gemini 2.0, Qwen 2.x) are intentionally excluded — they would compute correctly but aren't what you'd plan a new deployment around in 2026.
• Row order: grouped by model owner (OpenAI → Anthropic → Google → Qwen), sorted alphabetically by model name within each group. The cheapest-visible row is highlighted regardless of group position.
• Negotiated rates: the "Apply your negotiated rates" disclosure lets you enter an enterprise discount per hosting (e.g. 15% off Azure OpenAI). The discount is applied to both input and output token rates after the tier (on-demand / batch / cached) is resolved, so it stacks with batch and cached discounts. Discounts are encoded in the share-link hash so a colleague opening your URL sees the same negotiated view.
• Scenario presets draw workload assumptions from public benchmarks: Microsoft Copilot for M365 impact study (knowledge-work uplift), GitHub Copilot productivity research, and McKinsey "AI in the workplace" 2024. Adjust against your own deployment metrics before treating any output as a budget commitment.
• Business value mode compares inference cost against the dollar value of capacity freed up — users × workday hours × % automated × workdays × hourly rate. The % of workday automated slider is the lever: 10–25% is the range supported by Microsoft Copilot for M365 (~14% knowledge-worker uplift), GitHub Copilot studies (~26% engineer uplift), and McKinsey "AI in the workplace" 2024 (which projects up to 30% achievable by 2030). The default hourly rate ($100/hr) approximates $200K fully-loaded salary for US enterprise mid-level. Adjust both to your geography and role mix.
• Capacity freed vs replacement: the value mode frames the output as hours reinvested (time given back to each user) rather than headcount displaced. The FTE-equivalent figure is a capacity benchmark — "this much extra bandwidth" — not a layoff projection.
• Batch API tier: uses each provider's published batch price (typically ~50% of on-demand) where available. Models without a published batch tier fall back to on-demand at the same rate.
• Prompt-cached tier: only this tier consumes the cache slider. Input cost becomes fresh × on-demand_input + cached × cached_input_price. Cache write cost is ignored at v1 (small relative to read at high hit rates).
• Reasoning models (GPT-5.4 Pro, GPT-5.5 Pro, Claude Opus 4.8): output tokens include hidden chain-of-thought billed at the output rate. Pad output token estimates ~3–5× for reasoning-heavy tasks.
• Provisioned throughput (PTUs / Bedrock PT) tiers from Azure OpenAI and AWS Bedrock aren't modelled — those are reserved-capacity quotes negotiated per workload, not a single per-token rate.
• Gemini 3.1 Pro (preview) has a tiered price: $2/$12 for prompts ≤200K tokens; $4/$18 above. Calculator uses the ≤200K rate.

Token pricing source: LiteLLM model_prices_and_context_window.json, cross-checked against each provider's official pricing page (click any model name to open the citation). Snapshot date 2026-06-07.
USD → CAD rate: 1.3837 as published by Bank of Canada daily exchange rates on 2026-06-01. FX moves daily — for budget commitments above $100K, re-verify the rate.
Provider prices change without notice — re-verify before any financial decision.