Claude Opus 5.5 explained: pricing, API access and what changed

Anthropic released Claude Opus 5.5 on September 22, 2026. The new Opus model brings lower API prices and reported improvements to coding, long-running agents and professional work. It is also available through Requesty.

The pricing change is straightforward: standard input and output tokens cost 20% less than Opus 5, while cache reads cost 60% less. The implementation change deserves just as much attention. Adaptive thinking is now always enabled, the default effort is medium, and API requests can no longer force a tool call.

The question for anyone building with Claude is whether it completes their work more reliably, at a lower total cost, with an integration that still behaves as expected.

When was Claude Opus 5.5 released?

Anthropic publicly launched Claude Opus 5.5 on Tuesday, September 22, 2026, as the first model in the Claude 5.5 family.

The model is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Requesty provides access using the model ID anthropic/claude-opus-5-5.

Claude Opus 5.5 vs Opus 5: what changed?

Here is the practical comparison. Prices are standard global provider rates in US dollars per million tokens, before gateway fees. Specifications and rates come from Anthropic's Opus 5.5, Opus 5 and pricing documentation.

The context window and standard output limit are not new increases over Opus 5. Anthropic separately offers up to 300K output tokens in a Batch API beta; that is not the limit for standard requests. The more significant changes are how the model approaches work, how much that work costs and which controls your application can use.

More capable coding and longer-running work

Anthropic reports improvements in code migrations, debugging and agentic tasks. Its launch comparison puts Opus 5.5 at 66.4% on Terminal-Bench 4.0, versus 52.3% for Opus 5. These are vendor-reported results, not Requesty benchmarks. The published setup uses xhigh for Opus 5.5 on this test, with safeguards enabled and fallback models where applicable.

Anthropic also reports more than 30% faster output generation than Opus 5. That measures how quickly tokens arrive, not how quickly every task finishes. Tool latency, reasoning and the number of steps still matter.

A sensible evaluation would use a task your application already struggles with: a change spanning several modules, a difficult code review, or an investigation requiring multiple tools. Measure whether it finishes correctly, not just whether the first response looks convincing.

Better collaboration and professional work

Anthropic also describes clearer progress reports, stronger handling of charts and screenshots, and fewer mistakes in document-heavy work. Its prompting guide says Opus 5.5 at medium can match or exceed Opus 5 at high in its coding and knowledge-work evaluations.

That does not remove the need to check a source or review generated code. It does suggest testing more than benchmark-style questions. How much correction does an analyst need to make? Can an engineer quickly understand what changed? Those are useful measures of whether an improvement survives contact with a workflow.

Claude Opus 5.5 API pricing

At standard global provider rates, Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20 per million tokens.

Why the cache-read reduction matters

Prompt caching lets an application reuse previously processed context rather than pay the full input rate for that context each time. For Opus 5.5, a cache read costs 5% of the standard input rate, compared with 10% on Opus 5. The lower base price and lower multiplier combine to take the read price from $0.50 to $0.20.

Consider an illustrative workload with identical token counts on both models: 10 million uncached input tokens, 1 million output tokens and 40 million cache-read tokens.

That is a $27 reduction, or approximately 28.4%, for those three line items at identical usage. This calculation excludes cache creation, tools, retries, taxes and gateway charges. It assumes the cached content already exists; it is not a complete monthly bill or a measured workload result.

Anthropic's separate estimate of roughly 40% lower cost on typical workloads at default settings also reflects model efficiency. It should not be read as a flat 40% discount on every token category, or a promise for every application.

Cache writes, batch, fast mode and regional pricing

There are several other rates worth separating from the headline price in Anthropic's pricing documentation:

The Batch API gives a 50% discount on standard input and output tokens. Cache creation is billed separately from cache reads; consult the pricing schedule for those line items when batching. Regional arrangements depend on the platform, so check the endpoint rather than assuming every deployment uses the global rate.

Anthropic's Opus 5.5 release lists fast mode at $8 / $40 per million input / output tokens, with up to 2.5 times the standard output-token throughput. Its model-specific documentation describes it as a research preview on the Claude API. The general fast-mode guide still lists only Opus 5 and Opus 4.8 among supported models, so confirm access with your account team before relying on it. Higher token throughput does not guarantee an equivalent reduction in end-to-end latency. Do not assume a cloud or Requesty endpoint includes fast mode.

How adaptive thinking works in Opus 5.5

Adaptive thinking means Claude decides whether a request needs reasoning and how much to use. Your application steers that behaviour through effort and the prompt, rather than assigning a fixed thinking-token budget.

There is an important distinction here: always-enabled adaptive thinking does not mean the model must produce a long thinking block on every request. It can skip thinking for simple work. What you cannot do is switch the feature off through the API.

Start at medium, then evaluate

Opus 5.5 supports low, medium, high, xhigh and max, with medium as the default. Start there for a general evaluation. If your previous integration disabled thinking for latency-sensitive work, Anthropic suggests also testing low. Move upwards only where the additional work produces a useful quality gain.

Do not assume that high means the same amount of reasoning across model versions. Anthropic warns that carrying over a higher effort setting can lead to longer turns and more output tokens on Opus 5.5. A lower unit price does not protect you from a poorly matched configuration.

For the native Claude Messages API, a minimal request body looks like this:

{
  "model": "claude-opus-5-5",
  "max_tokens": 4096,
  "output_config": { "effort": "medium" },
  "messages": [
    {
      "role": "user",
      "content": "Review this proposed API migration and identify its three main risks: ..."
    }
  ]
}

The thinking field is intentionally omitted. On Opus 5.5, that still uses adaptive thinking. Sending thinking.type: "disabled" or the manual enabled mode with budget_tokens returns a 400 error.

max_tokens remains a hard ceiling for the request's combined thinking and response tokens. It is not a replacement thinking budget, and it does not cap an entire multi-call agent run. Leave enough room for an answer after reasoning.

Thinking tokens are billed as output tokens even when their text is not shown. Compare the usage reported by the API, not just the length of the answer you can see.

Tool use and other migration changes

Opus 5.5 still supports tools. It no longer supports forcing a tool call. On the native Claude API, tool_choice values any and tool are rejected; auto and none remain supported. Describe when a tool is required in the prompt, then have your application detect and handle missing calls.

If you previously forced a function call purely to obtain JSON, consider structured outputs instead. Schema-constrained output and a required external action are different problems. Strict tool use validates the arguments of a call that occurs; it is not a guarantee that the model will invoke the tool.

The migration guide flags other changes to test:

• Conversation handling: Pass thinking blocks back unchanged in ordinary tool loops. Preserved-thinking rules bind them to their conversation; editing earlier prompts, tools or messages can invalidate them, with default enforcement for accounts created on or after August 31, 2026.
• Progress updates: Narration between tool calls now arrives in thinking blocks. The default display omits their text, so a progress UI may need to request updates and handle those blocks.
• Computer use: The Claude API and Google Cloud require the newer computer_toolset_20260801 instead of computer_20251124. Bedrock retains the earlier tool.

For unattended agents, check task completion separately from the end of a response. The prompting guide notes that a progress report can end a turn before the overall job is finished.

How to access Claude Opus 5.5 through Requesty

Claude Opus 5.5 is available on Requesty's model page, which lists provider endpoints, regions and pricing. Use anthropic/claude-opus-5-5 with Requesty's OpenAI-compatible endpoint.

With the OpenAI Python package installed and REQUESTY_API_KEY set in your environment:

import os
from openai import OpenAI
 
client = OpenAI(
    api_key=os.environ["REQUESTY_API_KEY"],
    base_url="https://router.requesty.ai/v1",
)
 
response = client.chat.completions.create(
    model="anthropic/claude-opus-5-5",
    messages=[
        {
            "role": "user",
            "content": "Explain the trade-offs in this API migration plan: ...",
        }
    ],
)
 
print(response.choices[0].message.content)

This is a basic access example using the model's default effort, not a production agent loop. Native Claude request fields and beta features should not be assumed to transfer unchanged to an OpenAI-compatible request. Check the selected route before depending on a specific control.

Provider prices versus Requesty pricing

Requesty's pay-as-you-go pricing adds 5% to provider usage. At the standard global rates, that makes input $4.20, output $21, and cache reads $0.21 per million tokens. Its model page lists 0% markup when you bring your own provider keys. Keep those charges separate from Anthropic's base prices when estimating your spend.

Safeguards, fallback and data retention

Opus 5.5 can support legitimate security work, including code review and vulnerability triage, but some requests can trigger safety classifiers. In Claude products, flagged cyber requests may be served by Opus 4.8; some biology and frontier AI-development requests may move to Opus 5. Expanded Opus 5.5 access through the Cyber Verification Program was not available at launch.

On the API, automatic fallback is opt-in, not a default assumption. Anthropic documents a first-party beta using fallbacks: "default" with the server-side-fallback-2026-07-01 header. It handles eligible classifier refusals, not general outages or rate limits, and some refusal categories have no fallback. It is separate from gateway routing and failover.

AWS lists Opus 5.5 on Amazon Bedrock with zero data retention support by default. For Anthropic's first-party API, zero data retention is an arrangement enabled per organization and subject to feature eligibility. Neither statement establishes the retention settings of a Requesty route or your application's own logging, so check the complete route you select.

Anthropic also applies text watermarking to Opus 5.5. It describes the watermark as imperceptible, without changing readability or adding tokens or cost. This is not unique to the new model: its current support table also lists Opus 5. GitHub's launch note also describes watermarking on the model. Anthropic presents it as a transparency measure, not a substitute for a customer's own compliance assessment.

Should you move from Opus 5 to Opus 5.5?

The starting point is attractive: lower token prices, cheaper cache reads and reported improvements on demanding work. But a model upgrade is also a configuration change. A copied effort setting, an obsolete tool-choice parameter or an assumption about response blocks can change the outcome as much as the model itself.

Start with a representative sample of your existing tasks. Compare completion quality, review effort, latency and total billed usage. Include refusals, retries and cache behaviour rather than measuring only the successful first call.

Then choose the lowest effort setting that meets your quality bar. The useful result is a completed task that costs less to trust.

You can review the available endpoints and get started on the Claude Opus 5.5 model page on Requesty.

Sources

Prices and specifications checked September 25, 2026. Performance claims are attributed to their publishers; no independent Requesty benchmark results are presented here.

• Anthropic: Introducing Claude Opus 5.5
• Claude Platform: Opus 5.5 model overview
• Requesty: Claude Opus 5.5 model and endpoints
• Claude Platform: pricing
• Claude Platform: migrating to Opus 5.5
• Claude Platform: Opus 5 model overview
• Claude Platform: prompting Opus 5.5
• Claude Platform: fast mode
• Claude Platform: steering thinking
• Claude Platform: what's new in Opus 5.5
• Claude Platform: effort
• Claude Platform: structured outputs
• Requesty: pricing
• Claude Help Center: why Claude switches models
• Claude Platform: refusals and fallback
• AWS: Opus 5.5 availability and Bedrock data retention
• Claude Platform: API and data retention
• Claude Help Center: how Claude marks AI-generated content
• GitHub: Claude Opus 5.5 availability and watermarking