Claude

Anthropic's family of AI assistants. The current (April 2026) generation is Claude 4.x. Trained with safety/constitutional-ai and optimised for helpfulness, harmlessness, and honesty.

Model Lineup (April 2026)

[Source: Perplexity research / benchlm.ai, 2026-04-29]

Opus 4.7 released April 16 2026. Sonnet 4.6 is the default recommendation for most production workloads. 1.2 percentage points behind Opus 4.6 on SWE-bench at 40% lower cost.

Which Model to Use

• Haiku 4.5 — classification, routing, extraction, summarisation at high volume. Fastest, cheapest.
• Sonnet 4.6 — daily driver. Complex reasoning, code generation, agent tasks. The right default.
• Opus 4.6 / 4.7 — hardest problems where quality > cost. Extended architectural reasoning, research synthesis, complex multi-step code.

Extended Thinking

Claude's deep reasoning mode. The model reasons internally (invisible unless you request the thinking block) before producing its final answer.

response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[...]
)
# response.content includes a ThinkingBlock + TextBlock

Available on Opus and Sonnet. Not available on Haiku. Do not combine with explicit chain-of-thought prompting. The internal reasoning replaces it.

Prompt Caching

Claude's most impactful cost feature. Cache the system prompt, large documents, or few-shot examples. Subsequent calls with the same prefix pay 0.1x read cost.

All Claude 4.x models support both 5-minute and 1-hour cache TTLs. See apis/anthropic-api for the full breakdown.

With a 200K-token document as context:

• Without caching: every call pays for 200K input tokens
• With caching (1-hour TTL): first call pays 2x, subsequent calls pay 0.1x — 95% cost reduction for repeated queries

Claude's Character

Anthropic publishes explicit documentation of Claude's values and character. Key traits:

• Intellectual curiosity across all domains
• Warmth and genuine care for the humans it works with
• Playful wit balanced with substance
• Directness combined with openness to other views
• Deep commitment to honesty and ethics

These are explicitly trained via safety/constitutional-ai and reinforcement learning from human feedback. Claude is not simply following rules. It has internalised values.

Context Window and Long-Context Behaviour

All Claude 4.x models support 1M tokens (Haiku: 200K). For practical reference:

• 1M tokens ≈ 750K words ≈ an entire codebase
• At 1M context, quality degrades for information retrieval tasks if the key information is buried in the middle (the "lost in the middle" problem)
• For RAG tasks, Claude performs better on retrieved chunks than on raw long context injection

Vision and Multimodal

Claude's best-in-class strength is document understanding: PDFs, tables, charts, financial statements, technical diagrams. See multimodal/vision for the full treatment.

Responsible Scaling Policy (RSP)

Anthropic's commitment to safety at capability levels. Each Claude model is evaluated against a defined set of dangerous capability thresholds before deployment. If a model crosses a threshold, deployment is paused until mitigations are in place.

The RSP is public. See safety/alignment for details.

Claude Code

Claude's agentic coding tool. Uses Claude Sonnet 4.6 by default, switchable to Opus. See ai-tools/claude-code for the full tool treatment.

Key Facts

• Claude Opus 4.6: 80.8% SWE-bench Verified, 91.3% GPQA Diamond, $5/$25 per M tokens
• Claude Sonnet 4.6: 79.6% SWE-bench, $3/$15 per M — recommended production default
• Claude Haiku 4.5: 73.3% SWE-bench, $1/$5 per M — high-volume classification/routing
• Extended thinking available on Opus and Sonnet; not on Haiku
• 1M token context (Haiku: 200K); 1M tokens ≈ an entire large codebase
• Prompt caching: 95% cost reduction for repeated 200K-token document queries (1-hour TTL)
• "Lost in the middle" degradation: RAG over retrieved chunks outperforms raw long context injection

Common Failure Cases

Prompt caching miss because the cached prefix is not byte-identical across requests Why: cache hits require that the content before the cache_control breakpoint is exactly the same on every call; any change in whitespace, dynamic timestamps, or user-ID injection into the system prompt invalidates the cache and charges full input cost. Detect: cache hit rate in the Anthropic API usage dashboard is near 0% despite an apparently stable system prompt; cost per call matches uncached pricing. Fix: move all dynamic content (user ID, session context, timestamps) after the last cache_control breakpoint, so the static prefix before it never changes.

Extended thinking enabled on Haiku, causing a 400 API error Why: the thinking parameter is only valid for Opus and Sonnet models; passing thinking={"type": "enabled", ...} to a Haiku model ID returns an invalid_request_error. Detect: the API call raises a BadRequestError with a message about unsupported parameters; the model string in the request is claude-haiku-*. Fix: check the model string before enabling thinking; gate thinking on model family — only enable for claude-opus-* or claude-sonnet-*.

"Lost in the middle" degrades retrieval quality at 1M-token context, but the failure is silent Why: when the key document or fact is placed in the middle of a 1M-token context, Claude's retrieval accuracy degrades significantly compared to placing it at the beginning or end; the model still produces a confident, fluent answer using peripheral context. Detect: accuracy drops in eval runs when the target document is placed at positions 30-70% through the context; the model gives plausible but wrong answers with no expressed uncertainty. Fix: place the most important context at the start or end of the prompt; use RAG to surface the top-k chunks and inject them near the beginning rather than injecting the full corpus.

Explicit chain-of-thought prompt combined with extended thinking produces lower-quality output than either alone Why: the model's internal reasoning already performs chain-of-thought; adding an explicit "think step by step" instruction interferes with the extended thinking process and can cause the model to repeat reasoning redundantly or truncate the thinking block. Detect: response quality is lower than without CoT instructions; the thinking block is shorter than expected or the final answer contradicts the thinking. Fix: remove explicit CoT instructions (e.g., "think step by step", "reason through this") when extended thinking is enabled; let the thinking parameter handle reasoning internally.

Connections

• apis/anthropic-api — full API reference including prompt caching and extended thinking
• safety/constitutional-ai — how Claude is trained to be helpful and harmless
• safety/alignment — RSP (Responsible Scaling Policy) and safety evaluations
• prompting/techniques — getting the best out of Claude; do not use explicit CoT with extended thinking
• ai-tools/claude-code — Claude as an autonomous coding agent (Sonnet 4.6 by default)
• llms/model-families — Claude in context of the broader model landscape

Open Questions

• How does Claude Opus 4.7 differ from 4.6 in practice — is the improvement primarily reasoning depth?
• What is the practical quality ceiling for extended thinking on frontier coding tasks?
• How does "lost in the middle" degradation scale as context grows from 200K to 1M tokens?