Guide détaillé à venir
Nous préparons un guide éducatif complet pour le LLM Latency Cost Calculator. Revenez bientôt pour des explications étape par étape, des formules, des exemples concrets et des conseils d'experts.
The LLM Latency Cost Calculator helps developers quantify the hidden costs of response time in AI applications by modeling time-to-first-token (TTFT), tokens-per-second throughput, and total response time across different models and configurations. While most cost discussions focus on token pricing, latency has its own economic impact: slower responses increase user abandonment, reduce throughput capacity, and degrade the perceived quality of AI-powered features. Latency varies dramatically across models and providers. GPT-4o typically delivers time-to-first-token in 200 to 500 milliseconds and generates 80 to 120 tokens per second. GPT-4o-mini is faster at 100 to 300ms TTFT and 100 to 150 tokens per second. Claude Sonnet 4 ranges from 300 to 700ms TTFT with 70 to 100 tokens per second. These differences mean a 500-token response takes 3 to 7 seconds depending on model choice, directly impacting user experience and application design. This calculator models the total cost of latency including direct API costs, infrastructure costs of holding connections open, user drop-off rates correlated with response time, and the throughput implications of slower models requiring more concurrent connections to serve the same request volume. For real-time applications like chatbots and search, latency optimization can be as impactful as token cost optimization for overall system economics.
Total Response Time = Time to First Token + (Output Tokens / Tokens per Second). Effective Cost per Request = API Token Cost + (Response Time / 3600) x Server Connection Cost per Hour + Drop-off Probability x Lost Revenue per User. For example: 400ms TTFT + 300 tokens at 100 tok/s = 400ms + 3,000ms = 3.4 seconds total response time.
- 1Measure or estimate the time-to-first-token (TTFT) for your chosen model and configuration. TTFT is the delay between sending the API request and receiving the first token of the response. It depends on model complexity, input prompt length, server load, and geographic distance to the API endpoint. GPT-4o TTFT ranges from 200 to 500ms, while reasoning models like o1 can take 2 to 10 seconds for the initial thinking phase.
- 2Determine the tokens-per-second generation rate for your model. This is the speed at which the model produces output tokens after the first token arrives. Standard models generate 80 to 150 tokens per second. Longer outputs take proportionally longer: a 500-token response at 100 tokens per second takes 5 seconds after TTFT. Streaming the response to users reduces perceived latency by showing tokens as they arrive.
- 3Calculate total response time for your typical output lengths. For a chatbot with 200-token responses on GPT-4o: TTFT (350ms) + generation (200 tokens / 100 tok/s = 2,000ms) = 2.35 seconds total. For a content generation feature with 1,000-token outputs: TTFT (350ms) + generation (10,000ms) = 10.35 seconds. These times determine whether the feature feels responsive or sluggish to users.
- 4Model the user experience impact of latency. Research shows that user satisfaction drops significantly above 3-second response times. For chatbots, responses over 5 seconds cause 20 to 30 percent of users to abandon the conversation. For search features, results taking over 2 seconds see 10 to 15 percent lower engagement. The calculator assigns a dollar value to this lost engagement based on your conversion rates and user lifetime value.
- 5Calculate throughput capacity and its cost implications. A server handling streaming responses must hold connections open for the full response duration. If each response takes 5 seconds, one server thread handles 12 requests per minute. Switching to a faster model that responds in 2 seconds increases throughput to 30 requests per minute, requiring 60 percent fewer server resources for the same traffic. This infrastructure savings can exceed the API cost difference between models.
- 6Compare the total cost across model options including both token pricing and latency costs. A cheaper-per-token model that is slower might actually cost more when accounting for infrastructure, user drop-off, and throughput constraints. The calculator produces a total economic comparison that includes API cost, server cost, and estimated revenue impact from latency.
- 7Optimize latency through configuration changes. Reducing output token limits with max_tokens, using streaming to improve perceived responsiveness, implementing prompt caching to reduce TTFT, and choosing geographically closer API endpoints can each reduce latency by 20 to 50 percent without changing models. The calculator models the cost impact of each optimization.
For a chatbot targeting under 3-second responses, GPT-4o and GPT-4o-mini both meet the threshold while Claude Sonnet 4 is borderline. GPT-4o-mini is 28 percent faster than GPT-4o and 94 percent cheaper, making it the optimal choice for most chatbot applications.
Each 1,000-token generation takes 10.4 seconds. To serve 50 concurrent users, you need approximately 9 server threads holding connections open. At $5 per hour for server infrastructure, the throughput cost adds $0.0024 per request on top of the API token cost.
After 200ms for retrieval and 150ms TTFT, 1,450ms remain for token generation. At 130 tokens per second, the maximum output is 188 tokens. If the feature needs longer responses, either the latency budget must increase or a faster model or reduced retrieval time is needed.
Search engines with AI-powered answer generation must deliver results within 2 to 3 seconds to match user expectations set by traditional search. A search platform using GPT-4o for answer synthesis budgets 500ms for retrieval and 2,000ms for LLM generation. At 100 tokens per second, they can generate approximately 170 tokens (about 130 words) within the latency budget. This constraint dictates the maximum answer length and drives the choice of the fastest available model.
Real-time translation services must minimize latency for conversational flow. A live translation feature using GPT-4o-mini achieves 150ms TTFT and 130 tokens per second, translating a 50-word sentence (approximately 80 tokens output) in 0.77 seconds total. This sub-second latency enables natural conversation pacing. Using GPT-4o instead would add 200ms TTFT and reduce throughput, creating noticeable pauses that break conversational flow.
Trading and financial analysis platforms use LLMs for real-time market commentary and alert generation. Latency directly impacts the value of market-moving information. A financial platform using GPT-4o-mini for 100-token market alerts achieves delivery in under 1 second, meeting the requirement for time-sensitive financial information. The platform routes longer analytical pieces to GPT-4o in the background where latency is less critical.
Voice assistants and voice-enabled AI applications have strict latency budgets because users expect immediate verbal responses. The total pipeline from speech-to-text (300 to 500ms) to LLM generation to text-to-speech (200 to 400ms) must complete within 2 to 3 seconds. This leaves only 1 to 2 seconds for LLM generation, constraining model choice and response length. Many voice applications use GPT-4o-mini or Claude Haiku specifically for their faster TTFT.
For reasoning models like o1 and o3, the TTFT includes an extended thinking
For reasoning models like o1 and o3, the TTFT includes an extended thinking phase that can last 2 to 30 seconds depending on problem complexity. This thinking time is charged at the output token rate but is not visible in the streamed response. A request that produces 200 visible output tokens might have consumed 2,000 to 5,000 thinking tokens, creating both a latency penalty and a hidden cost multiplier. Reasoning models should only be used for tasks where the thinking time produces measurably better outcomes.
When deploying LLMs behind a global CDN or API gateway, the added network hops
When deploying LLMs behind a global CDN or API gateway, the added network hops introduce 10 to 50ms of additional latency per request. While individually small, this overhead compounds in agent applications that make 5 to 15 sequential LLM calls. An agent pipeline with 10 sequential calls accumulates 100 to 500ms of gateway overhead alone. For latency-sensitive agent applications, minimize network hops between the orchestrator and the LLM API endpoint.
Function calling and tool use add latency because the model must generate
Function calling and tool use add latency because the model must generate structured JSON output (which is slower than natural language) and then wait for the tool result before continuing. Each tool call round-trip adds the full TTFT plus tool execution time. An agent making 3 tool calls adds approximately 1 to 3 seconds of LLM latency plus the external tool response times. Design tool interfaces to minimize round-trips by batching multiple queries into single tool calls where possible.
| Model | TTFT (median) | Tokens/Second | 200-Token Response | 500-Token Response |
|---|---|---|---|---|
| GPT-4o | 350ms | 100 tok/s | 2.35s | 5.35s |
| GPT-4o-mini | 150ms | 130 tok/s | 1.69s | 4.00s |
| Claude Sonnet 4 | 500ms | 80 tok/s | 3.00s | 6.75s |
| Claude Haiku | 200ms | 120 tok/s | 1.87s | 4.37s |
| Gemini 1.5 Flash | 200ms | 140 tok/s | 1.63s | 3.77s |
| o1 (reasoning) | 3,000ms | 50 tok/s | 7.00s | 13.00s |
| Llama 3 70B (H100) | 100ms | 90 tok/s | 2.32s | 5.66s |
Which LLM has the lowest latency?
Among major commercial models, GPT-4o-mini consistently delivers the lowest latency with 100 to 200ms TTFT and 120 to 150 tokens per second. Claude Haiku is similarly fast. Among flagship models, GPT-4o is slightly faster than Claude Sonnet 4. Reasoning models like o1 are significantly slower with TTFT of 2 to 10 seconds due to internal thinking. Self-hosted models on H100 GPUs can achieve sub-100ms TTFT but require significant infrastructure investment.
How does prompt length affect latency?
Longer input prompts increase TTFT because the model must process all input tokens before generating the first output token. Processing 1,000 input tokens typically adds 100 to 300ms compared to a minimal prompt. Processing 10,000 input tokens can add 500 to 1,500ms. This is why RAG applications with large retrieved context have higher latency than simple chatbot interactions. Prompt caching (available from Anthropic) eliminates this processing time for repeated prompt prefixes.
Should I use streaming for all API calls?
Streaming should be used for any user-facing response that takes more than 1 second to generate. For programmatic API calls where the output is processed by code rather than displayed to users, non-streaming is simpler and has negligible latency benefit. Streaming adds minimal code complexity with most SDKs and is supported by all major providers at no additional cost. The perceived latency improvement from streaming is substantial: a 10-second response feels like 1 second when streamed.
How do I reduce TTFT for my application?
Key TTFT optimizations include: using prompt caching to skip processing of repeated prompt prefixes (saves 200 to 500ms), choosing geographically closer API endpoints (saves 50 to 200ms of network round-trip), reducing input prompt length (saves 100 to 500ms), and using faster models like GPT-4o-mini (saves 100 to 300ms vs GPT-4o). For self-hosted models, GPU-accelerated inference with optimized serving frameworks like vLLM can achieve sub-100ms TTFT.
What is the acceptable latency for different application types?
Search and autocomplete: under 500ms. Chatbot responses: under 3 seconds (with streaming). Content generation: under 10 seconds (with streaming progress indicator). Batch processing: minutes to hours (no latency requirement). Voice assistants: under 2 seconds total pipeline. Code completion: under 500ms for inline suggestions. These thresholds are based on user experience research and competitive benchmarks.
Conseil Pro
Implement a latency budget for your entire request pipeline and allocate it across components. For a 3-second chatbot budget: 200ms for network and preprocessing, 200ms for RAG retrieval, 300ms for TTFT, and 2,300ms for token generation (allowing approximately 300 tokens at 130 tok/s on GPT-4o-mini). This budget approach prevents individual components from consuming more than their share and highlights when a component needs optimization or a faster model is required.
Le saviez-vous?
Human conversational turn-taking has a natural gap of about 200 milliseconds between one person finishing and another starting to speak. When AI chatbot response times exceed 3 seconds, users unconsciously adopt a 'web search' mental model instead of a 'conversation' mental model, becoming less engaged and more likely to abandon. Achieving sub-2-second responses keeps users in the conversational mindset, increasing both engagement and satisfaction scores by 25 to 40 percent.