Claude Series 5 & Hybrid Reasoning Architecture — Models & Architecture

Architecture of the Series 5.5 Adaptive Reasoning Engine

Claude Series 5.5 (Opus 5.5, Sonnet 5.5, Haiku 5.5) and Claude Fable 5.1 evolve the hybrid reasoning paradigm pioneered by retired Claude 3.7 Sonnet. Rather than splitting models into purely conversational systems versus non-steerable reasoning black boxes, Anthropic unifies rapid response generation and deep internal chain-of-thought within the exact same model weights.

Effort-Steered Adaptive Thinking vs Instant Generation

On modern Claude 5.x models, manual budget_tokens is replaced by Adaptive Thinking and coarse-grained Effort control:

• Instant Execution Mode: Set thinking: { type: "disabled" } on Sonnet 5.5 or Haiku 5.5 for ultra-low latency classification and lightweight routing (note: Fable 5.1 and Opus 5.5 keep adaptive thinking always on).
• Adaptive Reasoning Mode: Set thinking: { type: "adaptive" } paired with output_config: { effort: "low" | "medium" | "high" | "max" }. Claude dynamically calibrates its scratchpad to task complexity—skipping overhead on simple turns and exploring deep proof trees on hard problems.

Streaming Thinking Blocks with the Official TypeScript SDK

Anthropic exposes thinking tokens in real time through SSE using distinct delta types, enabling live "Thinking..." accordions in UIs:

import Anthropic from "@anthropic-ai/sdk";

const anthropic = new Anthropic();

const stream = await anthropic.messages.stream({
  model: "claude-sonnet-5-5",
  max_tokens: 16384,
  thinking: { type: "adaptive" },
  output_config: { effort: "high" },
  messages: [
    { role: "user", content: "Formulate a formal proof that red-black tree insertion maintains O(log n) height." }
  ]
});

// Intercept streaming events by block type
stream.on("streamEvent", (event) => {
  if (event.type === "content_block_start" && event.content_block.type === "thinking") {
    console.log("[START THINKING]");
  } else if (event.type === "content_block_delta" && event.delta.type === "thinking_delta") {
    process.stdout.write(event.delta.thinking); // Live stream of thinking tokens
  } else if (event.type === "content_block_delta" && event.delta.type === "text_delta") {
    process.stdout.write(event.delta.text); // Final verified answer tokens
  }
});

const finalMessage = await stream.finalMessage();
console.log("\nUsage Summary:", finalMessage.usage);

128K Output Ceilings & 1M Context Boundaries

All current Claude 5.x models pair 128,000 max output tokens with 1M token context windows, enabling agents to synthesize entire multi-module refactors and formal verification suites in a single pass.