Claude Opus 5.5 Is Out: Pricing, Benchmarks and What Breaks When You Migrate from Opus 5

On September 22, 2026, Anthropic released Claude Opus 5.5, the new flagship of its main line and the successor to Opus 5. The model went live on GPTunneL three minutes after the announcement: in our API it is claude-opus-5.5, in Anthropic's API it is claude-opus-5-5.

The short version: it is cheaper per token than Opus 5 and stronger at agentic coding and computer use, but migrating breaks four things in your request. Below are the price, the benchmarks, the list of fixes and what we checked through our own gateway on release day.

Pricing and availability

Anthropic puts the savings on a typical workload at roughly 40% versus Opus 5: on top of the lower price, output is 30%+ faster and responses carry less filler. Measure it on your own requests: at the same effort Opus 5.5 thinks longer than Opus 5, and thinking tokens eat part of the savings.

Benchmarks

Numbers from Anthropic's announcement.

The more interesting rows are not about Opus 5 but about the neighbors. On FrontierCode Opus 5.5 beats GPT-6 Astra, which costs $10 / $50 per million tokens, and on CursorBench it beats Fable 5.1, a model one class up. On prompt injection resistance (the Gray Swan benchmark) it ties Fable 5.1.

Tester cases from the announcement: a 680,000-line code migration in under a day, and an HAProxy port from C to Rust in 9.5 hours versus 12 for Fable 5.1, at 51% lower cost. These are reports from Anthropic's partners, not reproducible measurements — a reference point, not a guarantee.

What breaks when you migrate from Opus 5

Swapping the model name is not enough. Per Anthropic's guide, four changes fail the request with a 400:

Three more changes fail nothing but alter behavior:

• The default effort is medium; on Opus 5 it was high. A request without the parameter now runs at a different level than before, both in quality and in cost. Set the level explicitly and rerun your eval; don't carry the Opus 5 setting over.
• Text between tool calls arrives in thinking blocks, not text, and it is empty at the default display: "omitted". If your UI showed users "searching the database…", it will quietly go silent.
• Thinking blocks are bound to the model and to the conversation prefix. For accounts created on or after August 31, 2026, replaying a block after editing system, tools or an earlier message returns a 400. Keep the conversation append-only and change instructions with mid-conversation system messages.

What we checked through the GPTunneL gateway

On release day we sent three requests through our OpenAI-compatible API:

• reasoning_effort: "low" is accepted; a simple question cost 28 tokens.
• tool_choice: "required" did not return a 400: the gateway returned a normal tool_calls with finish_reason: "tool_calls". The same request fails on Anthropic's API directly, so if you ever move from us to a direct key, this spot needs rewriting.
• Without tool_choice, with "use get_weather" in the prompt, the model called the tool on its own — which is exactly how Anthropic recommends getting a tool call.

A minimal request:

curl https://gptunnel.ru/v1/chat/completions \
  -H "Authorization: $GPTUNNEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5.5",
    "reasoning_effort": "medium",
    "messages": [{"role": "user", "content": "Find the bug in this SQL: ..."}]
  }'

There is no thinking parameter in requests to us at all — depth is controlled by reasoning_effort, as with the other reasoning models in the catalog. Keep the model name and the level in environment variables: switching between Opus 5, Opus 5.5 and Fable 5.1 then becomes one line of config. Long answers at high and above need streaming or you will hit timeouts — an SSE parser that handles the gateway's keepalive lines is in the GPT-6 Astra article.

Refusals and safety

Anthropic calls Opus 5.5 the best-aligned model it has tested: in the automated behavioral audit it tried to bypass environment restrictions about 85% less often than Opus 5. The safeguards mirror Fable 5.1:

• Cybersecurity. Most tasks in this category are rerouted to Opus 4.8. Routine debugging and fixing vulnerabilities in your own code work.
• Biology. A separate classifier runs; extended access goes through a verification program.
• Reasoning extraction. Requests to reproduce internal reasoning in the answer can be declined with the reasoning_extraction category.

On the direct API a refusal comes back as HTTP 200 with stop_reason: "refusal". That is a normal response, not a network error: retrying the same request on the same model won't help, you need a fallback to another model.

What's next: Sonnet 5.5, Haiku 5.5 and OpenAI's answer

Anthropic promises Sonnet 5.5 and Haiku 5.5 "in the coming weeks" with similar improvements. For most production pipelines that matters more than Opus itself: the bulk of the load usually runs on the mid tier.

OpenAI answered the same day: about 90 minutes after Opus 5.5 it shipped GPT-6 Sol and GPT-6 Luna, the mid and light models of the GPT-6 generation alongside Astra, at half the price of their GPT-5.6 predecessors. Both are already in the GPTunneL catalog; what they can do and which tasks to give them is in a separate breakdown.

Claude Opus 5.5 FAQ

Is Claude Opus 5.5 out? Yes, since September 22, 2026: Claude API, AWS, Google Cloud and Microsoft Foundry. On GPTunneL since the same day.

How much does Opus 5.5 cost? At Anthropic, $4 per million input tokens and $20 per million output tokens. On GPTunneL, $0.012 per 1K input tokens and $0.06 per 1K output tokens, 20% less than Opus 5, no subscription. Current prices are on the pricing page.

Can I use it for free? There is no permanent free access; Anthropic's API is paid. On GPTunneL you pay only for the tokens you use.

Opus 5.5 or Fable 5.1? On Anthropic's benchmarks Opus 5.5 beats Fable 5.1 on CursorBench and GDPval-AA and costs less. Fable 5.1 remains the model for the longest and hardest tasks. The easiest way to compare is on your own prompt in one chat.

How do I turn thinking off? You can't — on Opus 5.5 it is always on. The minimum depth is effort: "low" (reasoning_effort: "low" in our API).

Do I need a VPN? No: GPTunneL works without a VPN and accepts local payment methods.

Bottom line

Opus 5.5 is a rare release where the flagship got both cheaper and stronger than its predecessor. The cost of moving is a pass over your integration layer: remove disabled thinking and manual budgets, drop forced tool_choice, set effort explicitly and check that your users still see the agent's progress messages.

Open Claude Opus 5.5 on GPTunneL and run your request set next to Opus 5 and Fable 5.1 — on one balance and one key. Request and limit specs are in the documentation and at docs.gptunnel.ru.