Anthropic Performance Tuning

Overview

Claude latency has two components: time to first token (TTFT) and tokens per second (TPS). Different strategies target each.

Latency Benchmarks (approximate)

Model	TTFT (p50)	TTFT (p95)	Output TPS
Claude Haiku 4.5	200ms	600ms	~150
Claude Sonnet 4	400ms	1.2s	~90
Claude Opus 4	800ms	2.5s	~40

Optimization Strategies

Instructions

Step 1: Always Stream

// Streaming delivers the first token ASAP — user sees response instantly
// instead of waiting for the full response to generate

const stream = client.messages.stream({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  messages,
});

// First token arrives in ~400ms (Sonnet)
// Full response may take 5-10s, but user sees progress immediately
for await (const event of stream) {
  if (event.type === 'content_block_delta') {
    yield event.delta.text;
  }
}

Step 2: Prompt Caching — Faster TTFT

// Cached prompts skip re-processing — dramatically lower TTFT for large system prompts
const message = await client.messages.create({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  system: [{
    type: 'text',
    text: largeSystemPrompt, // 10K+ tokens
    cache_control: { type: 'ephemeral' },
  }],
  messages,
}, {
  headers: { 'claude-beta': 'prompt-caching-2024-07-31' },
});
// TTFT drops from ~2s to ~500ms on cache hit with large prompts

Step 3: Use Haiku for Speed-Critical Paths

// Haiku is 2-4x faster than Sonnet with 80% quality for many tasks
// Use for: classification, extraction, simple Q&A, routing decisions

const route = await client.messages.create({
  model: 'claude-haiku-4-5-20251001', // 200ms TTFT
  max_tokens: 10,
  system: 'Classify the intent. Reply with exactly one word: search, create, update, delete.',
  messages: [{ role: 'user', content: userInput }],
});

// Then use Sonnet/Opus for the actual task

Step 4: Reuse Client Instance

// BAD — creates new connection pool per request
app.get('/api/chat', async (req, res) => {
  const client = new Anthropic(); // DON'T
  // ...
});

// GOOD — single client shared across requests
const client = new Anthropic(); // Module-level singleton

app.get('/api/chat', async (req, res) => {
  const message = await client.messages.create({ ... });
  // ...
});

Step 5: Parallel Requests

// When you need multiple independent Claude calls, fire them in parallel
const [summary, sentiment, entities] = await Promise.all([
  client.messages.create({ model: 'claude-haiku-4-5-20251001', max_tokens: 200,
    messages: [{ role: 'user', content: `Summarize: ${text}` }] }),
  client.messages.create({ model: 'claude-haiku-4-5-20251001', max_tokens: 20,
    messages: [{ role: 'user', content: `Sentiment (positive/negative/neutral): ${text}` }] }),
  client.messages.create({ model: 'claude-haiku-4-5-20251001', max_tokens: 200,
    messages: [{ role: 'user', content: `Extract named entities from: ${text}` }] }),
]);

Step 6: Minimize Output Tokens

// Fewer output tokens = faster response
system: 'Be extremely concise. Use bullet points, not paragraphs.',

// Set tight max_tokens
max_tokens: 256, // Don't use 4096 for short answers

Output

• Streaming enabled for all user-facing responses (first token in ~400ms with Sonnet)
• Prompt caching reducing TTFT for large system prompts
• Model routing to Haiku for speed-critical classification/routing tasks
• Client instance reused across requests (no per-request connection overhead)
• Parallel requests firing independent Claude calls concurrently

Error Handling

Issue	Cause	Fix
TTFT > 3s	Large uncached prompt	Enable prompt caching
Slow output	Using Opus for simple tasks	Downgrade to Haiku/Sonnet
Timeouts	Long generation + default timeout	new Anthropic({ timeout: 120_000 })
529 overloaded	API capacity	SDK auto-retries; add fallback model

Examples

See Latency Benchmarks table and six numbered strategy sections above, each with complete TypeScript code examples.

Resources

• Streaming Docs
• Prompt Caching
• Models Comparison

Next Steps

See clade-deploy-integration for production deployment patterns.

Prerequisites

• Completed clade-install-auth
• User-facing application where latency matters
• Understanding of streaming and async patterns