Claude 5.5 Is Complete: Opus 5.5, Sonnet 5.5 and Haiku 5.5 in Two Weeks

In short:

• Anthropic released three models in sixteen days: Claude Opus 5.5 on September 22, Claude Sonnet 5.5 on September 28 and Claude Haiku 5.5 on October 7, 2026.
• Opus 5.5 is cheaper than Opus 5 at list price ($4 input / $20 output per million tokens). Sonnet 5.5 keeps Sonnet 5’s $2 / $10. Haiku 5.5 starts at $0.10 / $0.50, with a higher rate for prompts over 100,000 tokens.
• All three have a 1M-token context window and up to 128K output tokens.
• The newer models change request rules. Forced tool use, thinking: disabled, manual thinking budgets and non-default sampling parameters can now return a 400 depending on the model. Read the checklist below before you swap a model ID in production.

Claude Opus 5.5 (September 22)

Anthropic introduced Opus 5.5 as the first model in its Claude 5.5 family. The model ID is claude-opus-5-5. Anthropic’s published list price is $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. Anthropic says the model costs about 40% less than Opus 5 on typical workloads and generates output more than 30% faster.

Two behaviour changes matter for API users:

• Thinking cannot be switched off. Anthropic’s announcement states the model is no longer available with thinking turned off. Adaptive thinking is always on, steered by the effort parameter, whose default on the Claude API is medium.
• Some cybersecurity work is routed to an older model. Outside Anthropic’s verification programmes, most cybersecurity tasks are handled by Opus 4.8 instead. If you run security tooling through Opus, expect different behaviour than on Opus 5.

The same announcement said Sonnet 5.5 and Haiku 5.5 would follow “in the coming weeks”. Both shipped within sixteen days.

Claude Sonnet 5.5 (September 28)

Sonnet 5.5 (claude-sonnet-5-5) has the same input and output prices as Sonnet 5. On October 7 Anthropic halved its prompt-cache read price to $0.10 per million tokens. It uses the same tokenizer as Sonnet 5, so your token counts do not change. Its effort default on the Claude API is high.

This is the release with the most breaking changes. Anthropic’s migration notes list five:

1. thinking: {"type": "disabled"} returns a 400. To turn off up-front thinking, send thinking: {"type": "between_tools"}. That setting is accepted at low, medium and high effort only.
2. Forced tool use returns a 400. tool_choice of type any or tool is rejected. Keep auto and use strict tool schemas, or say in the prompt when the tool should be called.
3. Thinking blocks are bound to the model and the conversation. If you edit earlier turns and replay a Sonnet 5.5 thinking block, the request can fail. Keep histories append-only.
4. The older computer_20251124 tool is not accepted on the Claude API and Google Cloud. Use the computer_toolset_20260801 toolset.
5. Some advisor-tool pairings are rejected. Opus 4.8, Opus 4.7 and Sonnet 5 cannot advise a Sonnet 5.5 executor.

There is also a quieter change: text the model writes between tool calls now comes back inside thinking blocks. With the default display setting that text is empty. If your app streams those progress notes to users, it will go silent between tool calls until you change the thinking display setting.

Claude Haiku 5.5 (October 7)

Haiku 5.5 (claude-haiku-5-5) is the first new Haiku in about a year and the biggest price change of the three. Anthropic’s model page lists two price bands:

Context grows from 200K to 1M tokens and maximum output from 64K to 128K. Haiku 5.5 is also the first Haiku with adaptive thinking and an effort control, defaulting to medium.

Read the fine print before you budget:

• Same text, more tokens. Haiku 5.5 uses the newer tokenizer introduced with Claude 4.7. Anthropic says the same text counts as roughly 30% more tokens than on Haiku 4.5. Per-token price comparisons with Haiku 4.5 overstate the saving, so recount real prompts.
• Sampling parameters are out. A non-default temperature, top_p or top_k returns a 400.
• Assistant prefill is out. Ending messages with an assistant turn to “start” the reply returns a 400. End with a user turn.
• Manual thinking budgets are out. budget_tokens returns a 400. Use effort instead.
• Responses can begin with thinking blocks even when you never asked for thinking. Thinking tokens also count toward max_tokens, so a very small limit can end the response before any text.

Anthropic lists Haiku 5.5’s retirement as “not sooner than October 7, 2027”.

Also in the release notes

• Claude Sonnet 4.5 retires on November 30, 2026. Anthropic announced the deprecation of claude-sonnet-4-5-20250929 on September 30 and recommends Sonnet 5.5 as the replacement.
• Refusal billing changed on September 24. Refusals that happen before any output are billed again at normal rates when the category is bio, frontier_llm or reasoning_extraction. Other refusal categories are still not billed.
• The Python SDK reached v1.0 on August 20. It requires Python 3.10 or later and no longer accepts temperature, top_p or top_k. Pin your SDK version before upgrading model IDs, so you change one thing at a time.

A migration checklist that covers all three

Most of the breakage comes from code written for older models. These five changes make your code safe on all three new models and still work on older ones:

1. Read content blocks by type, never by position. content[0] is no longer guaranteed to be text.
2. Remove temperature, top_p and top_k from Claude requests unless you know the target model accepts them.
3. Stop forcing tools. Replace tool_choice: any or tool with auto and a clear instruction.
4. Handle stop_reason: "refusal" as a normal outcome, not an exception. It arrives with HTTP 200.
5. Re-measure tokens and max_tokens on real traffic before trusting old cost estimates.

from anthropic import Anthropic

client = Anthropic(base_url="https://api.oneaiapikey.com", api_key="sk-your-key")

msg = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=2048,
    messages=[{"role": "user", "content": "Summarise this diff in one line: ..."}],
)

# Newer Claude models can return thinking blocks before the answer.
# Select blocks by type instead of reading msg.content[0].
text = "".join(b.text for b in msg.content if b.type == "text")

if msg.stop_reason == "refusal":
    # Declines arrive as HTTP 200 with stop_reason "refusal".
    print("Declined:", getattr(msg, "stop_details", None))
else:
    print(text)

What this means if you use Relaykit

The Claude models in our catalogue today are:

• Claude Opus 5.5 (claude-opus-5-5): $4.40 input / $22 output per million tokens at our published rate.
• Claude Sonnet 5 (claude-sonnet-5): $2.20 input / $11 output per million tokens at our published rate.
• Claude Haiku 4.5 (claude-haiku-4-5): $1.10 input / $5.50 output per million tokens at our published rate.
• Claude Sonnet 5.5 (claude-sonnet-5-5): $2.20 input / $11 output per million tokens at our published rate.

Claude Opus 5.5 is available on our endpoint under its standard ID. The always-on thinking described above applies, so check your code against the checklist before you switch from an older Opus. Sonnet 5.5 and Haiku 5.5 are not in our catalogue at the time of writing. We add a model only after we have tested it through the gateway, and the models page is the place to check.

Our rates are our own published prices, not Anthropic’s list prices. The models page always shows the current rate per model, and the usage log in the console shows what each request actually cost. If you use Claude Code, pin your model aliases to IDs we serve, as described in our Claude Code base URL guide. That way a new default in Claude Code does not send requests for a model we do not offer.

Relaykit is an independent service and is not affiliated with Anthropic.