Anthropic API (Claude)

LLM API · by Anthropic · official site

What it actually does

Anthropic’s API provides programmatic access to the Claude family of large language models. As of mid-2026 the available tiers are Opus 4, Sonnet 4.5, and Haiku (with occasional minor version bumps). All models accept text inputs and emit text; Opus 4 and Sonnet 4.5 also accept images and PDFs as inputs. The API supports streaming, tool use (function calling), system prompts, and a retrieval‑augmented generation helper (document grounding via uploaded files). Opus 4 additionally exposes an “extended thinking” mode that allows the model to generate internal reasoning tokens before producing its final answer, effectively enabling chain‑of‑thought without visibility into that thinking. Context windows are 200K tokens for all tiers, though only Opus 4 reliably uses the full context for deep multi‑hop reasoning.

Who it’s for

Founders shipping production products that require a language model to behave predictably, follow instructions precisely, and not produce harmful or off‑brand output. Opus 4 targets applications where reasoning depth is paramount: legal document analysis, research summarization, complex code generation, financial modeling. Sonnet 4.5 is the default for most production workloads: customer‑facing chatbots, content generation, data extraction pipelines. Haiku is for cost‑sensitive, high‑volume tasks: classification, sentiment analysis, simple Q&A, real‑time streaming responses where latency matters more than nuance.

What works

• Instruction adherence and formatting control. Claude is broadly the most consistent model at following explicit formatting instructions (JSON, XML, markdown) without additional parsing work. This reduces the fragility of production systems.
• Tool use (function calling). The JSON‑based tool schema is clean and the model calls tools with high precision. It also interprets tool results (including errors) more gracefully than many alternatives.
• Extended thinking (Opus 4 only). When the complexity budget is high, enabling extended thinking yields noticeable gains in multistep reasoning tasks (e.g., proving mathematical theorems, evaluating logic trees). It also lets you expose the model’s internal reasoning time as a controllable parameter.
• Safety guardrails. Constitutional AI tuning produces a model that rarely produces toxic or unhinged output, even under adversarial prompting. For regulated industries (healthcare, finance, legal) this reduces manual review overhead.
• Latency for Haiku and Sonnet. Haiku returns under 500ms for moderate token counts; Sonnet remains fast enough for conversational use. Both support streaming with negligible overhead.

What breaks

• Extended thinking is unreliable for shallow prompts. Enabling it on simple tasks often adds cost and latency without benefit, and sometimes degrades output quality because the model wastes reasoning tokens on irrelevant paths.
• Rate limits can bite without enterprise negotiation. Out‑of‑box limits are generous for prototyping but insufficient for production traffic above roughly 500 RPM (varies by tier). Reserved throughput requires an annual commitment and a sales conversation.
• Image and PDF understanding is behind GPT‑4o and Gemini 2.5 Pro. Opus 4 handles clean text in images well but struggles with low‑quality scans, handwritten text, and complex charts. For document extraction, you still need a dedicated OCR layer.
• Refusal rate on benign inputs. Claude sometimes refuses to answer harmless prompts that could be interpreted as “sensitive,” particularly when the prompt includes words like “violence” even in a historical context. Tuning system prompt or using the safety parameters only partially mitigates this.
• Context window execution degrades after 120K tokens. While advertised as 200K tokens, random‑access information retrieval (e.g., “find the line that contains X”) becomes noticeably inaccurate past 100–120K tokens. Sonnet degrades faster than Opus 4. For long‑document workloads, you should implement chunking and retrieval rather than relying on the full window.

Pricing reality (what you actually pay at usage X)

Anthropic’s published per‑million‑token rates change frequently. As of mid‑2026 the approximate values are:

• Haiku: 10M * $0.15/1M = $1.50 input + 2M * $0.75/1M = $1.50 output → $3/month.
• Sonnet 4.5: 10M * $3.50/1M = $35 + 2M * $15/1M = $30 → $65/month.
• Opus 4: 10M * $20/1M = $200 + 2M * $100/1M = $200 → $400/month.

The honest comparison (vs 2–3 alternatives)

When to use it

Use Claude API when instruction reliability, safety, and controlled output format matter more than multimodal breadth or raw cost per token.

Last verified: 2026-06-08 by kernel.