# CheckAPICost AI API Cost & Architecture Simulator (Full Specification 2026)

## 1. Executive Summary & Problem Formulation
The transition toward agentic workflows and multi-step reasoning models has rendered static token calculation obsolete. Modern LLM economics are governed by three non-linear variables:
1. **Temporal Prompt Caching Dynamics**: Caching represents a 50% to 90% discount on input tokens, but diverges sharply across providers. Anthropic imposes a 1.25x (5-min TTL) or 2.0x (1-hour TTL) cache creation write penalty, whereas OpenAI utilizes automatic prefix matching (>= 1,024 tokens) without write penalties.
2. **Payload Hierarchy & Cache Invalidation**: Placing dynamic variables (such as timestamps, UUIDs, or un-anchored session data) above static prompt segments invalidates all subsequent cache blocks. The optimal hierarchy is: System Instructions -> JSON Tool Schemas -> Static Knowledge Corpus -> Conversation History -> Dynamic User Message.
3. **Redundancy & Gateway Arbitrage**: Enterprise high-availability necessitates multi-provider failover. Routing through gateways (OpenRouter at 5.5% vs LLM Gateway at 5% / 0% BYOK) introduces blended financial variables that must be modeled mathematically.

## 2. Temporal Prompt Caching Mechanics
### Anthropic Claude Architecture
- Cache Read Discount: 90% (0.1x of standard input price).
- Cache Creation Surcharge: 1.25x base input price for 5-minute TTL; 2.0x for 1-hour TTL.
- Activation: Explicit `cache_control: {"type": "ephemeral"}` markers on up to 4 blocks.
- Min Threshold: 1,024 tokens (Claude Sonnet / Opus).

### OpenAI Architecture
- Cache Read Discount: 50% on legacy tiers, up to 90% on GPT-5.5 / GPT-5.6 / GPT-6 Astra tiers.
- Cache Creation Surcharge: None (first call billed at standard base input rate).
- Activation: Automatic prefix detection on payloads exceeding 1,024 tokens.
- Expiration: Inactivity window of 5-10 minutes, persisting up to 24 hours under continuous traffic.

### DeepSeek Architecture
- Cache Read Discount: 90% (0.1x of base rate, e.g. $0.014 / 1M tokens on DeepSeek V3).
- Cache Write: Handled in 64-token chunks automatically.

## 3. Dynamic Cost Models & Formulas
Let `T_in` be input tokens, `T_out` be output tokens, `P_in` be base input price per 1M, `P_out` be base output price per 1M.
Let `h` be the empirical cache hit rate `(0 <= h <= 1)` and `m = 1 - h` be the miss rate.

### Standard Linear Calculation (Legacy / Baseline):
`Cost_Base = (T_in * P_in / 1,000,000) + (T_out * P_out / 1,000,000)`

### Anthropic Dynamic Caching Calculation:
`Cost_Anthropic_Input = (T_in / 1,000,000) * [ m * (1.25 * P_in) + h * (0.10 * P_in) ]`
`Total_Cost = Cost_Anthropic_Input + (T_out * P_out / 1,000,000)`

### OpenAI Dynamic Caching Calculation:
`Cost_OpenAI_Input = (T_in / 1,000,000) * [ m * (1.00 * P_in) + h * (0.10 * P_in) ]`
`Total_Cost = Cost_OpenAI_Input + (T_out * P_out / 1,000,000)`

### Gateway Fallback Blended Model:
Let `U` be primary provider uptime `(0 <= U <= 1)`, and `F = 1 - U` be failover rate.
Let `G_fee` be the gateway markup (e.g. 0.055 for OpenRouter, 0.05 for LLM Gateway standard, 0.00 for BYOK).
`Blended_Cost_Per_1M = (U * Cost_Primary) + (F * Cost_Fallback * (1 + G_fee))`

## 4. Empirical Benchmark Data & Grounded Model Quality Index
Data grounded from LMSYS Chatbot Arena, Artificial Analysis Quality Index, and SWE-bench Verified:
- **GPT-6 Astra Pro**: 1448 Elo | 82.5% SWE-bench | 92.4% MMLU-Pro | 97.8% MATH-500
- **Claude Opus 5**: 1442 Elo | 81.8% SWE-bench | 91.8% MMLU-Pro | 97.2% MATH-500
- **GPT-6 Astra**: 1435 Elo | 80.5% SWE-bench | 90.5% MMLU-Pro | 96.5% MATH-500
- **Claude Sonnet 5**: 1430 Elo | 79.8% SWE-bench | 90.2% MMLU-Pro | 96.1% MATH-500
- **Claude Fable 5**: 1422 Elo | 78.5% SWE-bench | 89.4% MMLU-Pro | 95.5% MATH-500
- **Grok 4.6**: 1418 Elo | 77.5% SWE-bench | 88.6% MMLU-Pro | 94.8% MATH-500
- **Grok 3**: 1402 Elo | 72.4% SWE-bench | 87.0% MMLU-Pro | 93.8% MATH-500
- **Claude 3.7 Sonnet**: 1385 Elo | 70.3% SWE-bench | 85.2% MMLU-Pro | 92.4% MATH-500
- **GPT-4.5 Preview**: 1380 Elo | 69.5% SWE-bench | 84.8% MMLU-Pro | 91.0% MATH-500
- **o1 Heavyweight**: 1375 Elo | 68.2% SWE-bench | 84.0% MMLU-Pro | 96.4% MATH-500
- **Gemini 3.1 Pro**: 1368 Elo | 67.0% SWE-bench | 83.5% MMLU-Pro | 90.2% MATH-500
- **DeepSeek R1**: 1355 Elo | 65.7% SWE-bench | 82.0% MMLU-Pro | 95.8% MATH-500

## 5. B2B Infrastructure Integration Endpoints
- **LLM Gateway**: 1% perpetual compute spend partner program; 200+ models, 0% BYOK.
- **Portkey.ai**: Enterprise prompt governance, OpenTelemetry tracing, fallback routing.
- **OpenRouter**: Unified API routing across 500+ endpoints.

## 6. Enterprise Specialized Calculators Suite
1. **AI Pipeline DAG Composer** (`tools/ai-pipeline-composer.html`): Interactive visual graph orchestration canvas supporting linear, branching, router, evaluator-optimizer, and swarm topologies with instant LiteLLM Python, TypeScript, and LangGraph exports.
2. **LLM ROI & Headcount TCO Calculator** (`tools/llm-roi-calculator.html`): Financial ROI model comparing human engineering/support hours saved against API token operational spend.
3. **AI SaaS Gross Margin Simulator** (`tools/saas-margin-simulator.html`): Tiered user seat economics, token consumption per active user, and gross margin defense calculator.
4. **Google Maps Enterprise SKU Calculator** (`tools/google-maps-sku-calculator.html`): Tiered pricing breakdown across Places Autocomplete, Geocoding, Directions, and Mapbox alternative savings.
5. **AWS S3 Egress & Storage Calculator** (`tools/aws-s3-egress-calculator.html`): Multi-tier data transfer out modeling with Cloudflare R2 zero-egress arbitrage calculations.
6. **Vector DB & RAG Memory Calculator** (`tools/vector-db-rag-calculator.html`): HNSW indexing RAM overhead, vector dimension byte calculation, and serverless vs dedicated pricing.
7. **AI Scraping Credit Calculator** (`tools/ai-scraping-credit-calculator.html`): Proxy bandwidth billing, headless browser concurrency costs, and markdown extraction credit usage.
8. **Realtime Voice AI Agent Calculator** (`tools/voice-agent-cost-calculator.html`): Audio token pricing, STT/TTS latency budgets, and SIP trunking per-minute unit economics.
9. **Twilio & Global SMS Calculator** (`tools/twilio-global-sms-calculator.html`): Country-specific carrier pass-through rates and A2P 10DLC campaign messaging fees.

## 7. Category Intelligence Hubs & Knowledge Graphs
- **AI Agents & Orchestration Hub** (`apis/ai-agents.html`): Autonomous execution pricing, loops, tool-calling tokens, 6-region geo latency benchmarks.
- **AI SaaS Unit Economics Hub** (`apis/ai-saas.html`): Multi-tenant architecture margins, COGS containment, and token allocation tiers.
- **Cloud Storage & Data Egress Hub** (`apis/cloud-storage.html`): S3 vs GCS vs R2 vs Wasabi egress fee matrix and CDN caching strategies.
- **Maps & Geocoding Hub** (`apis/maps-geocoding.html`): Maps SKU volume discount thresholds, session tokens, and migration cost curves.
- **Vector Databases Hub** (`apis/vector-databases.html`): Hybrid search pricing, index memory formulas, and multi-tenant isolation overhead.
- **Web Scraping Hub** (`apis/web-scraping.html`): Anti-bot stealth proxy fees, JavaScript rendering costs, and batch crawling economics.
- **Voice AI & Speech Hub** (`apis/voice-speech.html`): Latency-cost Pareto frontiers, audio codec tokenization, and telephony integration pricing.
- **Messaging & SMS Hub** (`apis/messaging-sms.html`): Carrier surcharges, alphanumeric sender ID registration, and global delivery rates.

## 8. Omni-Query Programmatic Search Directory
- **Universal API Pricing & Intent Index** (`directory.html`): Indexes 120+ targeted developer search queries with structured JSON-LD item lists, live formula previews, and instant calculator links.