# Claude Opus 5.5 cuts API prices and brings migration changes

> Anthropic's Opus 5.5 lowers token prices and targets supervised enterprise work below Fable 5.1, while API clients face breaking thinking and tool changes, plus conditional fallback.

By BIG CHANGE Editorial

Published: 2026-09-22T17:10:37.939Z
Updated: 2026-09-23T03:16:01.519Z
Canonical: https://bigchange.ai/blog/claude-opus-5-5-cost-performance-access

![A large abstract work capsule travels on a broad supported track while a separate side track ends at an unlabelled gate.](https://bigchange.ai/api/media/file/claude-opus-5-5-access-boundary-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE. More affordable capable AI can put longer jobs within reach, while routing and safeguards still determine what work a service handles. This is a conceptual mechanism, not an Anthropic product, benchmark or measured customer result.

Anthropic [launched Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) on September 22. It claims near-Fable 5.1 performance on most work, with lower prices than Opus 5.

For teams already using Opus 5, the update is more than a cheaper model ID. Opus 5.5 always uses adaptive thinking, defaults to medium effort and rejects several request shapes that Opus 5 accepted. It also sits below Fable 5.1 as Anthropic's recommended daily model for supervised work, with Fable reserved for the hardest and longest autonomous jobs.

## The big change

- **What changed:** Opus 5.5 lowers input and output rates by 20% and cache-read rates by 60%. Always-on adaptive thinking also changes how existing API clients must request and budget work.
- **Why it matters:** Long agent jobs repeatedly read earlier instructions and tool results. The largest price cut applies to that reused context, so cache usage becomes central to understanding the whole job’s bill.
- **What to watch:** Sonnet 5.5 and Haiku 5.5 are announced for the coming weeks; Opus is the first available. Choosing a default means comparing completed-task costs and the review each tier needs.

## Opus 5.5 is aimed at production work below the Fable tier

Anthropic's partner positioning recommends Opus 5.5 as the default enterprise choice for production coding, financial and legal knowledge work, and agents. Its [public model-selection guide](https://platform.claude.com/docs/en/about-claude/models/choosing-a-model) says most workloads start with Opus 5.5 and pairs it with complex agentic coding and enterprise work. A separate [cost guide](https://claude.com/blog/what-a-task-costs-on-opus-5-5) calls it a “daily driver” for supervised feature work, debugging and code review. Both guides place Fable 5.1 above it when the highest capability or a long, difficult run matters more than token price.

Anthropic's [prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5) supports that breadth with company tests across root-cause analysis, code checking, parallel subagents, research, spreadsheets, reports and presentations. It also describes clearer progress messages and final summaries, along with vision and computer-use improvements. These are Anthropic's evaluations, not tests performed by BIG CHANGE.

The distinction between Opus and Fable matters because Fable 5.1 lists at $10 per million input tokens and $50 per million output tokens, versus $4 and $20 for Opus 5.5. Cache reads are much closer at $0.25 and $0.20. For long cached sessions, the price difference can therefore be narrower than the headline input and output rates suggest.

## Claude Opus 5.5 pricing changes four API rates

The standard API prices below are per million tokens. Anthropic's [model reference](https://platform.claude.com/docs/en/models/opus-5-5/overview) lists a 1 million-token context window and a 128,000-token maximum output for Opus 5.5. Its [pricing documentation](https://platform.claude.com/docs/en/about-claude/pricing) says the full context window uses standard rates.

| API charge | Opus 5.5 | Opus 5 | Change |
| --- | --- | --- | --- |
| Input | $4 | $5 | 20% lower |
| Output | $20 | $25 | 20% lower |
| 5-minute cache write | $5 | $6.25 | 20% lower |
| Cache read | $0.20 | $0.50 | 60% lower |

Anthropic reports 40% lower typical-workload cost at default settings, combining rates with token efficiency. Its [task-cost explanation](https://claude.com/blog/what-a-task-costs-on-opus-5-5) shows why the mix matters: long agent sessions repeatedly read growing context, output includes thinking tokens and a task can make many turns. A 60% cache-read reduction can dominate a cache-heavy session, while a short uncached request with a long answer is closer to the 20% output-rate cut.

The reported improvement of more than 30% is output generation speed, according to the prompting guide. It does not measure time spent waiting for tools, tests, external services or a reviewer. Anthropic also offers a faster API mode at $8 input and $40 output per million tokens, in research preview on the Claude API rather than partner cloud platforms.

Anthropic also increased five-hour limits on Pro, Max, Team and seat-based Enterprise plans and provided a bankable reset. Subscription capacity is separate from API billing.

## Company benchmarks show strengths and boundaries

Anthropic reports Terminal-Bench 4.0 scores of 66.4% for Opus 5.5 at xhigh, 52.3% for Opus 5 and 55.8% for Fable 5.1. Most other headline results use max effort.

GPT-6 Astra leads the displayed AutomationBench and Terminal-Bench-Science rows. Safeguarded cyber tasks used Opus 4.8; biology and frontier-model-development tasks used Opus 5. AutomationBench used no fallback.

Anthropic reports fewer irreversible or out-of-bounds actions in its behavioral audit. Its [prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5) also reports stronger resistance to indirect prompt injection than earlier Opus models. Both are company-run test results rather than a general safety certification.

## Existing API clients need an explicit migration

API customers must request `claude-opus-5-5`; existing Opus 5 calls do not silently change models. The [migration reference](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5) lists several request changes that can otherwise return errors or alter an application's output stream.

| Migration point | Opus 5.5 behavior |
| --- | --- |
| Model selection | Use `claude-opus-5-5` explicitly. |
| Thinking | Adaptive thinking is always on. Disabled thinking and manual token budgets return HTTP 400. |
| Effort | Use `low`, `medium`, `high`, `xhigh` or `max`; the default is `medium`. |
| Forced tools | `tool_choice` values `any` and `tool` return HTTP 400; `auto` and `none` remain supported. |
| Computer use | The older `computer_20251124` tool is rejected on the Claude API and Google Cloud; Amazon Bedrock keeps supporting it. |
| Progress text | Text between tool calls arrives in thinking blocks, so a client using the default omitted display can appear silent between calls. |

Preserved thinking adds another boundary for long-running clients. Thinking blocks are tied to a model and conversation prefix. On API and cloud-platform accounts created on or after August 31, 2026, changing earlier instructions, tools or messages before replaying a block returns a 400 error by default. [Anthropic's preserved-thinking documentation](https://platform.claude.com/docs/en/build-with-claude/preserved-thinking) explains the append-only path and the controlled option to drop an incompatible block.

## Safety routing can change the model that answers

The [prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5) describes Fable 5.1-level biology safeguards and classifiers for cybersecurity and reasoning extraction. Source-code vulnerability finding remains allowed, and organizations can apply for life-sciences verification. Expanded cyber verification is announced for the coming weeks.

A refusal can still arrive with HTTP 200 and `stop_reason: "refusal"`. The [fallback documentation](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback) describes server-side fallback, SDK middleware and client-managed retries. Fallback occurs only when one of those supported paths is configured; it is not an automatic promise attached to every Opus 5.5 request. Categories without a recommended fallback can still be refused.

That makes the serving model part of the result. An application moving to Opus 5.5 needs to record refusals and fallback-served responses separately, especially when its work touches safeguarded domains. The lower rates and broader capability claims apply to Opus 5.5; a routed request follows the behavior and limits of the model that actually answered.

## Sources

- [Anthropic: Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) — Primary launch announcement for the model, Anthropic's typical-workload cost and speed claims, benchmarks, subscriber limit changes and safety approach. Published tests and early-user examples are company reports, not independent results across all tasks.
- [Claude Platform: What's new in Claude Opus 5.5](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5) — Live model and migration reference for API ID, provider availability, prices, effort defaults, unsupported thinking/tool settings, fast-mode scope and computer-use tool migration. Check a chosen provider's support before deployment.
- [Claude Platform: Preserved thinking](https://platform.claude.com/docs/en/build-with-claude/preserved-thinking) — Explains model binding and append-only conversation-prefix rules for returned thinking blocks. The default prefix check applies to accounts created on or after August 31, 2026; Anthropic-managed interfaces do not require direct client changes for this feature.
- [Claude Platform: Refusals and fallback](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback) — Documents classifier refusals as HTTP 200 responses with stop_reason refusal and category details, plus fallback options. A successful HTTP status is not proof the requested model completed the task.
- [Claude Opus 5.5 model reference](https://platform.claude.com/docs/en/models/opus-5-5/overview) — Official specification for the 1M-token context window, 128K maximum output, standard and cache prices, default medium effort, model IDs and supported platforms. The article does not infer performance from these specifications.
- [What a task costs on Opus 5.5](https://claude.com/blog/what-a-task-costs-on-opus-5-5) — Anthropic product guidance explaining how turns, cache reads, output and effort contribute to task cost, and positioning Opus 5.5 for supervised daily work below Fable 5.1. Its examples are illustrative company guidance, not BIG CHANGE tests.
- [Prompting Claude Opus 5.5](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5) — Official guidance for company-tested coding, knowledge-work, communication, visual, computer-use and multi-agent capabilities, plus the output-generation speed claim and progress-update behavior. These are Anthropic evaluations, not BIG CHANGE product tests.
- [Choosing the right Claude model](https://platform.claude.com/docs/en/about-claude/models/choosing-a-model) — Official selection matrix says most workloads start with Opus 5.5, maps it to complex agentic coding and enterprise work, and reserves Fable 5.1 for the highest available capability. This supports the model-tier positioning without relying only on partner material.
- [Claude Platform pricing](https://platform.claude.com/docs/en/about-claude/pricing) — Official pricing reference confirms that Claude 4.6 and later models include the full 1M-token context window at standard per-token rates. It also confirms prompt-cache and fast-mode rates used in the comparison.