Claude Sonnet 5 API Migration Guide (2026): 1M Context, Adaptive Thinking, and Pricing

Anthropic announced Claude Sonnet 5 on July 2, 2026, and the news matters for developers because this isn't just another model name in a dropdown. Sonnet 5 is now available on the Claude API and Claude Code, ships with a native 1M-token context window, turns adaptive thinking on by default, and launches with temporary pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026.

That's attractive. It's also the kind of launch that can quietly break production if you treat it like a pure string replacement. Sonnet 5 removes manual extended thinking, rejects non-default sampling parameters, and uses a newer tokenizer that can produce roughly 30% more tokens for the same text. If you run agents, code review bots, support automation, or long-context retrieval workflows, you should migrate deliberately.

This guide is the practical version: what changed, what code to update, what to measure, and where teams are likely to waste money.

The short version: what changed?

Step 1: change the model ID, but don't stop there

The minimum change is simple:

model = "claude-sonnet-4-6"  # before
model = "claude-sonnet-5"    # after

That part takes five seconds. The migration work is everything around it: gateway routing, request validators, test fixtures, cost dashboards, retry logic, and any code that assumes a smaller context window. If your app has a hardcoded model allowlist, update it in the same pull request. If your billing dashboard groups models by prefix, make sure Sonnet 5 doesn't fall into an "unknown" bucket.

If you use an OpenAI-compatible gateway such as KissAPI to route across Claude, GPT, Gemini, and other models, add the new model as a separate route first. Don't overwrite your Sonnet 4.6 production route until you have real usage data.

Step 2: remove unsupported sampling parameters

This is the migration bug I expect to see most often. Many teams set temperature by habit. On Claude Sonnet 5, setting temperature, top_p, or top_k to non-default values returns a 400 error.

Bad migration request:

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 1200,
    "temperature": 0.2,
    "messages": [{"role": "user", "content": "Review this patch."}]
  }'

Safer request:

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 1200,
    "system": "Be concise. Prefer deterministic, production-safe answers.",
    "messages": [{"role": "user", "content": "Review this patch."}]
  }'

Use the system prompt to guide style and consistency. Don't fight the API.

Step 3: migrate manual extended thinking to adaptive thinking

Sonnet 5 runs with adaptive thinking by default. Manual extended thinking with thinking: {"type":"enabled", "budget_tokens": N} is gone and returns a 400. If your current wrapper injects a thinking budget for every Claude request, update it before switching the model ID.

# Not supported on Claude Sonnet 5
thinking = {"type": "enabled", "budget_tokens": 32000}

# Use adaptive thinking instead
thinking = {"type": "adaptive"}

There's one practical catch: max_tokens is a hard limit for total output, including thinking plus final answer text. If you previously used max_tokens: 1000 for short answers, Sonnet 5 may have less room for visible output than you expect. Test with real prompts, not toy prompts.

Python migration wrapper

Here's a small wrapper that strips the risky fields and logs usage so you can compare Sonnet 4.6 and Sonnet 5 before a full cutover.

import os
from anthropic import Anthropic

client = Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

UNSUPPORTED_ON_SONNET_5 = {"temperature", "top_p", "top_k"}

def ask_sonnet5(messages, system=None, max_tokens=2000):
    kwargs = {
        "model": "claude-sonnet-5",
        "max_tokens": max_tokens,
        "messages": messages,
    }
    if system:
        kwargs["system"] = system

    # Optional: be explicit. Omitting thinking also enables adaptive thinking.
    kwargs["thinking"] = {"type": "adaptive"}

    resp = client.messages.create(**kwargs)
    usage = getattr(resp, "usage", None)
    if usage:
        print({
            "input_tokens": getattr(usage, "input_tokens", None),
            "output_tokens": getattr(usage, "output_tokens", None),
            "cache_read_input_tokens": getattr(usage, "cache_read_input_tokens", 0),
        })
    return resp.content[0].text

result = ask_sonnet5(
    messages=[{"role": "user", "content": "Find risks in this deployment plan."}],
    system="You are a pragmatic senior engineer. Return blockers first.",
    max_tokens=2500,
)
print(result)

Node.js example for a coding agent route

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

const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });

export async function reviewWithSonnet5(diff) {
  const response = await client.messages.create({
    model: "claude-sonnet-5",
    max_tokens: 3000,
    thinking: { type: "adaptive" },
    system: [
      {
        type: "text",
        text: "You are a senior code reviewer. Prioritize security, data loss, and backwards compatibility."
      }
    ],
    messages: [
      { role: "user", content: `Review this diff:\n\n${diff}` }
    ]
  });

  return {
    text: response.content[0].text,
    usage: response.usage
  };
}

Cost check: the promotional price is not the whole story

The launch price looks great: $2 per million input tokens and $10 per million output tokens until August 31, then $3/$15. But the tokenizer change matters. If the same codebase prompt becomes 30% more tokens, your real request cost may not fall as much as the pricing table suggests.

For teams running multi-model gateways, this is a perfect place to use a cost calculator instead of guessing. Long-context requests might still belong on Sonnet 5. Small deterministic extraction tasks may belong on a cheaper model. That's not a downgrade; it's routing discipline.

A practical rollout plan

1. Create a new route: Add claude-sonnet-5 beside your existing Sonnet 4.6 route.
2. Run shadow tests: Replay recent prompts without sending Sonnet 5 output to users.
3. Remove invalid parameters: Strip non-default sampling fields and old extended-thinking budgets.
4. Recount tokens: Pay special attention to code, JSON, logs, and non-English text.
5. Raise output limits selectively: Don't blindly set 128k. That's how bills get weird.
6. Canary by endpoint: Start with code review, research, or internal agent workflows before customer-facing automation.
7. Keep fallback routing: If Sonnet 5 refuses a category of cyber-adjacent task or hits limits, route safely instead of failing hard.

KissAPI can help here if you want one OpenAI-compatible endpoint for model routing, fallback, and quick A/B tests across providers. Use it as an operational layer, not a magic wand: you still need good evals and sane token budgets.

When should you switch?

My take: switch early for coding agents, long-context document analysis, and multi-step automation where Sonnet 4.6 was getting stuck. Wait and test for highly controlled JSON extraction, support classifiers, and workloads with tight cost ceilings. Sonnet 5 is stronger, but stronger models are not automatically cheaper in production.

The best migration is boring: a new route, a canary, a dashboard, and a rollback button. Do that, and Sonnet 5 is a clean upgrade rather than a Friday incident.

FAQ

What is the Claude Sonnet 5 API model ID?

The model ID is claude-sonnet-5. Treat it as a new production route first, then migrate traffic after testing token counts and behavior changes.

Does Claude Sonnet 5 support 1M context?

Yes. Anthropic says Sonnet 5 has a 1M token context window by default and supports up to 128k output tokens.

Is Claude Sonnet 5 cheaper than Sonnet 4.6?

During the introductory period, yes on per-token pricing: $2/$10 per million input/output tokens through August 31, 2026. After that it moves to $3/$15, the same standard rate as Sonnet 4.6. But the new tokenizer can produce more tokens for the same text, so measure your workload.