Claude Sonnet 5.5 API: measured against Sonnet 5 on 41 prompts

Claude Sonnet 5.5 answered our 41 test prompts a little faster than Sonnet 5 and wrote about half as much again. It went live on FastMetal on September 30, 2026, as anthropic-claude-sonnet-5-5, at the same per-token rates as Sonnet 5, so existing code only needs a new model value.

Anthropic announced it on September 28 as the second model in the Claude 5.5 family: the faster, lower-cost partner to Opus 5.5, aimed at well-scoped everyday tasks, bug fixes and polished documents.

The measurements

We sent the same 41 prompts in Japanese and English to both models with our comparison runner on September 30, 2026. Figures are medians per answer.

• The median latency is lower for Sonnet 5.5, but the gap is narrow. Sonnet 5.5 was faster on 22 of the 41 prompts.
• Sonnet 5.5 wrote more on 33 of the 41 prompts. Anthropic says it uses fewer tokens than Sonnet 5; on our prompts we saw the opposite. Your prompts and settings may differ, so measure on your own workload.
• Since it writes more in less time, its throughput is higher: about 129 output tokens per second against about 75, dividing one median by the other.

The extra length is often extra content. Asked to rewrite a paragraph on the CAP theorem to be clear and concise while staying technically accurate, Sonnet 5.5 gave the rewrite, a list of the main changes, and a note that the original's "you can only have two of the three" framing is a popular simplification. Sonnet 5 gave the rewrite and a short list of changes. Which you want depends on the job. All 41 answers are on the Claude Sonnet 5.5 model page.

What Anthropic says

• More than 30% faster output generation than Sonnet 5, with fewer tokens and tool calls.
• 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5.

These are the vendor's own figures. Ours are above.

What it supports

• A 1,000,000-token context and up to 128,000 output tokens.
• Image and file input alongside text.
• Tool calling, structured output and reasoning controls.
• No training on your prompts. We confirmed that it answers with training opted out.

How it is billed

Sonnet 5.5 is billed per million input and output tokens, with separate rates for cache reads and cache writes, including a 1-hour cache write. The rates are the same as Claude Sonnet 5. Current rates are on the model page and the pricing page.

Billing is prepaid: you load a balance by card, from $5, and requests stop when it runs out. There is no bill after the fact, so a trial has a fixed ceiling.

Calling it

It is OpenAI-compatible, so change the model value:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic-claude-sonnet-5-5",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

With the OpenAI SDK, point base_url at https://api.fastmetal.ai/v1:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.fastmetal.ai/v1",
    api_key="sk-...",
)

res = client.chat.completions.create(
    model="anthropic-claude-sonnet-5-5",
    messages=[{"role": "user", "content": "Explain the time complexity of this function."}],
)
print(res.choices[0].message.content)

Moving from Sonnet 5

With the rates equal, the deciding question is how output length and latency change on your prompts. Output tokens are billed, and on our run Sonnet 5.5 wrote more, so set a max_tokens limit and watch your spend for a few days before you move production traffic. For prompt-by-prompt quality against another model, see GPT-6.1 Sol vs Claude Sonnet 5.5 and the comparison page.

Everything else we serve is in the model catalog.

Source: Introducing Claude Sonnet 5.5 (Anthropic)