Claude API Ops

Claude API Operations

Building applications and agents on Anthropic's API: the Messages API, tool use, prompt caching, structured outputs, batches, thinking/effort, and the Claude Agent SDK. For developers writing apps against the API — not for using Claude Code itself.

API surfaces move fast. Model IDs, parameters, and betas in this skill were verified against platform.claude.com (2026-06). When in doubt — especially for "latest model" or pricing questions — verify with WebFetch against https://platform.claude.com/docs/en/about-claude/models/overview.md or query the Models API (client.models.list()).

Current Models (verified 2026-06)

Use these alias IDs verbatim. Never append date suffixes (claude-sonnet-4-6-20251114 is wrong → 404). Older actives: claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-4-5. Live capability lookup: client.models.retrieve("claude-opus-4-8") → .max_input_tokens, .max_tokens, .capabilities dict.

Model Selection Decision Tree

What is the workload?
│
├─ Hardest problems, long-horizon agents, deep research, ceiling intelligence
│  └─ claude-fable-5 (premium) or claude-opus-4-8 (default flagship)
│
├─ Agentic coding, tool-heavy workflows, production assistants
│  └─ claude-opus-4-8 (quality) or claude-sonnet-4-6 (speed/cost balance)
│
├─ High-volume production: summarization, RAG answers, extraction
│  └─ claude-sonnet-4-6
│
├─ Classification, routing, simple Q&A, latency-critical
│  └─ claude-haiku-4-5
│
└─ Subagents inside a larger system
   └─ One tier below the orchestrator (Opus loop → Sonnet/Haiku workers)

Tiering rule: route by task difficulty, not by uniform default. An Opus orchestrator dispatching Haiku classifiers is routinely 5-10x cheaper than Opus-everywhere with no quality loss on the simple legs.

Which Surface? (API vs Agent SDK vs Batches)

Rule of thumb: start at the simplest tier. Reach for an agent only when the task is genuinely open-ended (multi-step, hard to fully specify, errors recoverable, value justifies cost).

Messages API Quick Start

Everything goes through POST /v1/messages. Headers: x-api-key, anthropic-version: 2023-06-01, content-type: application/json.

# pip install anthropic
import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=16000,
    system="You are a concise technical assistant.",
    messages=[{"role": "user", "content": "Explain CRDTs in one paragraph."}],
)
for block in response.content:        # content is a list of typed blocks
    if block.type == "text":          # always check .type before .text
        print(block.text)
print(response.stop_reason, response.usage.input_tokens, response.usage.output_tokens)

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

const client = new Anthropic();

const response = await client.messages.create({
  model: "claude-opus-4-8",
  max_tokens: 16000,
  messages: [{ role: "user", content: "Explain CRDTs in one paragraph." }],
});
for (const block of response.content) {
  if (block.type === "text") console.log(block.text);  // narrow the union first
}

Streaming (default to it for long outputs — non-streaming above ~16K max_tokens risks SDK HTTP timeouts):

with client.messages.stream(model="claude-opus-4-8", max_tokens=64000,
                            messages=[{"role": "user", "content": "Write a long report"}]) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)
    final = stream.get_final_message()   # full Message after streaming

Full params, response shape, stop reasons, errors, retries, rate limits: references/messages-api.md

Thinking & Effort (quick reference)

• Claude 4.6+ (Fable 5, Opus 4.8/4.7/4.6, Sonnet 4.6): use thinking: {"type": "adaptive"}. The old fixed budget {"type": "enabled", "budget_tokens": N} is removed on Fable 5 / Opus 4.8 / 4.7 (400 error) and deprecated on Opus 4.6 / Sonnet 4.6.
• Effort (GA): output_config: {"effort": "low" | "medium" | "high" | "xhigh" | "max"} — nested in output_config, not top-level. Default high. xhigh (Opus 4.7+) is best for coding/agentic work; max is Opus-tier + Sonnet 4.6 only.
• Sampling params removed on Fable 5 / Opus 4.8 / 4.7: temperature, top_p, top_k all return 400. Steer with prompting + effort.
• Thinking + forced tool_choice is incompatible: with thinking on, only tool_choice: {"type": "auto"} (default) or "none" is allowed — {"type": "any"} or {"type": "tool", ...} returns a 400.
• Thinking text is omitted by default on Fable 5 / Opus 4.8 / 4.7 — opt in with thinking: {"type": "adaptive", "display": "summarized"} if you surface reasoning to users.

Details and gotchas: references/structured-outputs.md (thinking interplay) and references/messages-api.md.

Tool Use (quick reference)

tools = [{
    "name": "get_weather",
    "description": "Get current weather. Call when the user asks about weather conditions.",
    "input_schema": {
        "type": "object",
        "properties": {"location": {"type": "string", "description": "City, e.g. Paris"}},
        "required": ["location"],
    },
}]
response = client.messages.create(model="claude-opus-4-8", max_tokens=16000,
                                  tools=tools, messages=messages)
if response.stop_reason == "tool_use":
    ...  # execute, send tool_result back, loop

tool_choice: {"type": "auto"} (default) | {"type": "any"} | {"type": "tool", "name": "..."} | {"type": "none"}. Add "disable_parallel_tool_use": true to force at most one call per response.

The agentic loop, parallel tool results, pause_turn, is_error, server-side tools, and SDK tool runners: references/tool-use.md

Cost Optimization Checklist

Work top-down; each item is independent:

• Right-size the model. Haiku for classification/routing, Sonnet for volume work, Opus/Fable for the hard 10%. Largest single lever.
• Prompt caching on stable prefixes (system prompt, tool defs, big docs): cache_control: {"type": "ephemeral"}. Reads cost ~0.1x; up to 90% savings. Verify with usage.cache_read_input_tokens > 0 — zero means a silent invalidator (timestamp in system prompt, unsorted JSON, varying tools).
• Batches API for anything that can wait ≤24h: flat 50% off all tokens, stacks with caching.
• Cap output: set max_tokens to what you need (256 for classification); stream + generous cap for long generation.
• Tune effort down where quality allows: medium is often the sweet spot; low for subagents and simple tasks.
• Count before sending: client.messages.count_tokens(...) (never tiktoken — it's OpenAI's tokenizer and undercounts Claude by 15-20%).
• Keep prefixes stable: order requests tools → system → messages, volatile content last; don't swap tool sets or models mid-conversation.

Mechanics, breakpoints, TTLs, batch lifecycle, tiering math: references/caching-and-cost.md

Claude Agent SDK (quick reference)

# pip install claude-agent-sdk   (Python >= 3.10)
import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions

async def main():
    async for message in query(
        prompt="Find and fix the bug in auth.py",
        options=ClaudeAgentOptions(allowed_tools=["Read", "Edit", "Bash"]),
    ):
        if hasattr(message, "result"):
            print(message.result)

asyncio.run(main())

// npm install @anthropic-ai/claude-agent-sdk
import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: "Find and fix the bug in auth.ts",
  options: { allowedTools: ["Read", "Edit", "Bash"] },
})) {
  if ("result" in message) console.log(message.result);
}

Built-in tools (Read/Write/Edit/Bash/Glob/Grep/WebSearch/WebFetch/...), hooks (PreToolUse, PostToolUse, ...), subagents, MCP servers, sessions (resume/fork), permission modes, and the SDK-vs-raw-API decision: references/agent-sdk.md

Common Pitfalls

Resources & Verification

This skill ships a staleness verifier and two copy-and-adapt starter assets. The model table and pricing above are the facts most likely to drift — run the verifier when you suspect they're stale.

scripts/check-model-table.py — guards the Current Models table (this file) and the per-model prompt-cache minimum table (references/caching-and-cost.md) against drift. Two modes per the resource protocol §7:

# Structural (default, no network): every row well-formed, ids carry no date
# suffix, prices numeric, the two files agree on the model lineup. Exit 4 on a
# malformed/contradictory row.
python skills/claude-api-ops/scripts/check-model-table.py --offline
python skills/claude-api-ops/scripts/check-model-table.py --offline --json | python -m json.tool

# Live (advisory, needs ANTHROPIC_API_KEY): curls the Models API and compares
# its id set against the documented ids. Exit 10 if a documented id is gone or a
# newer alias id is missing from the table; exit 7 (not a failure) if the key is
# unset or the API is unreachable. Live mode checks model-ID coverage ONLY — the
# API returns no pricing, so pricing/context drift stays an --offline + docs concern.
ANTHROPIC_API_KEY=sk-... python skills/claude-api-ops/scripts/check-model-table.py --live

assets/agentic-loop.py — a minimal, runnable tool-use loop (define a tool, call messages.create, loop while stop_reason == "tool_use", append tool_result, re-request until end_turn). Copy it as the starting point when building a manual agent loop; the >>> ADAPT marks show what to change.

assets/output-schema.json — a known-good structured-outputs request body in the canonical output_config.format shape (with additionalProperties: false and a required array). Copy and reshape schema.properties when adding JSON outputs; see references/structured-outputs.md for the rules. (Not supported on Fable 5 — that model uses system-prompt instructions instead.)

Reference Files

Live Documentation

When cached facts may be stale, WebFetch (append .md for clean markdown):

• Models/pricing: https://platform.claude.com/docs/en/about-claude/models/overview.md
• Messages API: https://platform.claude.com/docs/en/api/messages
• Tool use: https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview.md
• Prompt caching: https://platform.claude.com/docs/en/build-with-claude/prompt-caching.md
• Structured outputs: https://platform.claude.com/docs/en/build-with-claude/structured-outputs.md
• Batches: https://platform.claude.com/docs/en/build-with-claude/batch-processing.md
• Agent SDK: https://code.claude.com/docs/en/agent-sdk/overview