0Pricing
Claude Architect · 课时

Claude 模型系列

根据成本、速度和能力选择合适的模型。

Claude 模型系列 是 CoddyKit 上的免费 Claude Architect 课时。 这是第 1 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 Claude Architect 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Claude Architect 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Meet the Model Family

Every Claude API request begins with one decision: which model? That single field sets your cost, your latency, and your intelligence ceiling for the whole call.

The current family spans three tiers:

  • Opus — the most capable, for the hardest, long-horizon work.
  • Sonnet — the best balance of speed and intelligence.
  • Haiku — the fastest and cheapest, for simple high-volume tasks.

As a Claude Certified Architect, picking the right tier per workload is a core skill. Let's build the decision framework.

The model field is just a string

You select a model by passing its model ID as a string. The request also keeps no state — the model is stateless, so you send the full message history on every turn.

The exact ID strings matter. A typo (for example claude-sonnet-4.6 instead of claude-sonnet-4-6) returns a 404 not_found_error, not a silent fallback.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)
print(response.content[0].text)

Opus — top of the family

Claude Opus 4.8 (claude-opus-4-8) is the most capable model. Reach for it when correctness on a hard problem outweighs cost and speed:

  • Large multi-step refactors and overnight coding runs.
  • Deep research and long-horizon agentic loops.
  • Complex reasoning where a cheaper model would degrade quality.

It offers a 1M-token context window at standard pricing and up to 128K output tokens. Pricing: $5 / 1M input, $25 / 1M output.

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=64000,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "Refactor this service end to end."}],
)

Sonnet — the balanced workhorse

Claude Sonnet 4.6 (claude-sonnet-4-6) is the best combination of speed and intelligence. It's the default choice for most production traffic:

  • Chat, classification, content generation, extraction.
  • Tool-heavy workflows that need fast turnaround.
  • Anything where Opus would be overkill but Haiku might miss nuance.

It has a 1M-token context window and up to 64K output tokens. Pricing: $3 / 1M input, $15 / 1M output — cheaper than Opus on both ends.

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=8192,
    messages=[{"role": "user", "content": "Classify this support ticket."}],
)

Haiku — fast and cheap at scale

Claude Haiku 4.5 (claude-haiku-4-5) is the fastest and most cost-effective model. Use it for simple, high-volume tasks where latency and price dominate:

  • Lightweight classification and routing.
  • Short summaries and formatting fixes.
  • Sub-agents doing narrow, well-scoped work.

Its context window is 200K tokens (smaller than the 1M of Opus and Sonnet), with up to 64K output tokens. Pricing: $1 / 1M input, $5 / 1M output — the cheapest tier.

response = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=256,
    messages=[{"role": "user", "content": "Is this email spam? Answer yes or no."}],
)

The three-way tradeoff

Picking a model is always a tradeoff across three axes:

  • Capability — Opus > Sonnet > Haiku for hard reasoning.
  • Speed — Haiku is fastest, Opus is most deliberate.
  • Cost — Haiku ($1/$5) < Sonnet ($3/$15) < Opus ($5/$25) per 1M tokens.

The architect's instinct: start at the cheapest tier that meets the quality bar, then move up only where evals show you need it. Don't reflexively default everything to Opus — you'll burn budget on tasks Sonnet or Haiku handle just as well.

Match the tier to the task

A practical routing heuristic for a multi-stage system:

  • Coordinator / planner doing complex decomposition → Opus.
  • Main worker handling tool calls and synthesis → Sonnet.
  • Narrow sub-agents (classify, grep-and-summarize) → Haiku.

This mirrors how production agents stay affordable: spend Opus-level intelligence only where the decision is hard, and push routine, parallel work down to cheaper models.

Don't switch models mid-conversation

Models are cache-scoped. Prompt caching is a prefix match, and the cache is keyed per model — so switching the model string mid-conversation invalidates the entire cache and you reprocess everything at full price.

The architect-grade pattern: keep one model on the main loop, and if a sub-task needs a cheaper model, spawn a sub-agent for it rather than swapping the model on the live thread.

# Main loop stays on Opus; cheap discovery work goes to a Haiku sub-agent
main = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=16000,
    messages=history,  # cached prefix preserved
)

scout = client.messages.create(
    model="claude-haiku-4-5",  # separate call, separate cache
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize these 5 files."}],
)

Control cost with effort, not model swaps

You don't always have to change the model to change cost. On Opus 4.8 and Sonnet 4.6, the effort parameter tunes how much the model thinks and spends, inside output_config:

  • low — terse, fewer tool calls, latency-friendly.
  • high — the default; thorough.
  • max — Opus-tier only, when correctness beats cost.

Pair it with adaptive thinking (thinking={"type": "adaptive"}) and let the model decide how much to reason per request. Note: budget_tokens is removed on these models — use adaptive thinking instead.

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=8192,
    thinking={"type": "adaptive"},
    output_config={"effort": "low"},  # cut cost without dropping a tier
    messages=[{"role": "user", "content": "Draft a release note."}],
)

Mind the context window and output cap

Capability isn't only intelligence — it's also how much the model can hold and produce:

  • Opus 4.8 / Sonnet 4.6: 1M-token context window.
  • Haiku 4.5: 200K-token context window.

If a workload must reason over a very large codebase or document set in a single call, Haiku's smaller window may rule it out regardless of cost. Also stream any request with large max_tokens (above ~16K) — non-streaming calls risk SDK HTTP timeouts.

with client.messages.stream(
    model="claude-opus-4-8",
    max_tokens=128000,  # large output -> must stream
    messages=[{"role": "user", "content": "Generate the full migration plan."}],
) as stream:
    message = stream.get_final_message()

Batch API: a cost lever, not a model tier

Cost optimization isn't only about model choice. The Message Batches API runs requests at 50% off, within a window of up to 24 hours.

The catch — it's not for blocking or time-sensitive work: there is no latency SLA, and multi-turn tool calling isn't supported. Use it for overnight reports and audits; never for a pre-merge check or anything a user is waiting on. For those, pick a fast model (often Haiku or Sonnet) on the standard API instead.

batch = client.messages.batches.create(
    requests=[
        {"custom_id": "row-1", "params": {
            "model": "claude-haiku-4-5",
            "max_tokens": 256,
            "messages": [{"role": "user", "content": "Classify row 1"}],
        }},
    ],
)

Quick Check: choosing a model

A scenario-style question on matching the model to the constraint.

Recap: choosing across the family

Key takeaways for the exam and for production:

  • Opus 4.8 ($5/$25, 1M context) — hardest reasoning, long-horizon agents; use sparingly.
  • Sonnet 4.6 ($3/$15, 1M context) — the balanced default for most traffic.
  • Haiku 4.5 ($1/$5, 200K context) — fast, cheap, high-volume simple tasks.
  • Start at the cheapest tier that meets the quality bar; move up only where evals demand it.
  • Tune cost with effort and adaptive thinking before swapping tiers.
  • Don't switch models mid-conversation — it invalidates the cache; spawn a cheaper sub-agent instead.
  • Use the Batches API (50% off) for non-blocking jobs only — never for time-sensitive checks.

常见问题解答

「Claude 模型系列」课时是免费的吗？

是的 — 「Claude 模型系列」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Claude Architect 课程的其余内容，请升级到 CoddyKit PRO。 Claude Architect 课程共包含 4 节课。

「Claude 模型系列」这节课中我会学到什么？

根据成本、速度和能力选择合适的模型。 你通过在浏览器中直接运行的动手代码来练习 Claude Architect，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 Claude Architect 需要有经验吗？

无需任何先前经验。CoddyKit 上的 Claude Architect 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 1 节课，共 4 节。

「Claude 模型系列」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 Claude Architect 课中编写并运行代码吗？

能。每节 Claude Architect 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

  1. Claude 模型系列
  2. API 请求剖析
  3. 停止原因详解
  4. 令牌、上下文窗口与成本
← 返回 Claude Architect