Anthropic Releases Claude Opus 5.5: 1M Context, 40% Lower Cost, and Agentic Coding

Image source: https://www.anthropic.com/claude-opus-5-5

On September 22, 2026, Anthropic officially announced Claude Opus 5.5 (model ID: claude-opus-5-5), inaugurating its next-generation Claude 5.5 model family. Specifically built for long-running agentic coding and demanding knowledge work, the new model introduces a 1-million-token context window and a 128K maximum output token limit, while reducing typical operational costs by approximately 40% and accelerating text generation speed by more than 30% compared to its predecessor, Claude Opus 5.

Token Pricing and 60% Cache Discount: 40% Lower Cost on Typical Workloads

Under Anthropic's standard API pricing, Claude Opus 5.5 is priced at $4 per million input tokens ($4 / MTok) and $20 per million output tokens ($20 / MTok). This represents a 20% reduction in per-token rates compared to Claude Opus 5 ($5 / $25 per MTok), and roughly 40% of the cost of Anthropic's top-tier Claude Fable 5.1 ($10 / $50 per MTok). For latency-sensitive enterprise environments, a high-throughput Fast Serving Mode is available at $8 per million input tokens and $40 per million output tokens.

The prompt caching tier sees an even steeper reduction: prompt cache reads now cost $0.20 per million tokens, a 60% drop from the previous $0.50 rate. Anthropic reports that combining lower baseline token rates with reduced token consumption per task cuts total spend by roughly 40% across typical production workloads using default settings. Additionally, generation speed is over 30% faster than Opus 5, noticeably cutting latency during long-context codebase generation and refactoring.

Claude Opus 5.5 is available immediately across Claude web and mobile applications for paid subscribers (Pro, Team, Enterprise), the Anthropic API, Amazon Web Services (AWS Bedrock), Google Cloud (Vertex AI), and Microsoft Azure.

Benchmark Performance: 66.4% on Terminal-Bench 4.0 and Agentic Coding

In autonomous development evaluations, Claude Opus 5.5 demonstrated marked gains in complex task execution. On Terminal-Bench 4.0, a benchmark assessing autonomous coding workflows inside real terminal environments, Claude Opus 5.5 achieved a 66.4% success rate, surpassing Claude Fable 5.1 (55.8%) by more than 10 percentage points.

On AutomationBench, which evaluates operating system interaction and browser navigation, Opus 5.5 scored 40.0% in task completion, outperforming Opus 5's 26.9%. Anthropic highlighted that across broader complex reasoning and knowledge-work benchmarks, the model consistently delivers intelligence comparable to Fable 5.1 while substantially trimming execution expenses.

Complementing these gains, the 1-million-token context window and 128K maximum output token limit allow engineering workflows to keep extensive codebases resident in context during multi-step architectural migrations or complex debugging sessions without premature truncation.

Developer Migration Checklist: Four Breaking Changes and Thinking Block Display

Teams running existing production pipelines on Claude Opus 5 must address four breaking changes introduced with Opus 5.5 (the first three of which also apply to Fable 5.1):

First, extended thinking cannot be disabled. Thinking mode is mandatory and permanently enabled on claude-opus-5-5. Second, thinking blocks are tied to the specific model and conversation session; editing, stripping, or modifying thinking blocks in previous conversation turns triggers an API error.

Third, forced tool execution is no longer permitted. Specifying a mandatory tool selection via tool_choice will return an error, requiring developers to let the model select tools autonomously. Fourth, the legacy computer_20251124 tool definition is deprecated and rejected on both the Claude API and Google Cloud, requiring migration to current tool specifications.

In addition, developers must account for a structural shift in streaming tool interactions. Text generated between sequential tool calls is now encapsulated inside thinking blocks rather than delivered as plain content. Under default display settings, this text is blank in streamed responses. Applications that stream interim agent updates to user interfaces must adjust their display configuration to expose this text, preventing UI screens from appearing blank between tool calls.

Sources

• Anthropic Documentation: Claude Opus 5.5 Model Overview
• Anthropic Documentation: Claude API Pricing Overview
• X Signal Source (@AI_Caffeine): Claude Opus 5.5 Release Announcement