Anthropic announced that the 1 million token context window for its flagship models Claude Opus 4.6 and Sonnet 4.6 is officially fully open and billed at standard prices on the Claude platform, with no additional long context premium.Under the new price system, Opus 4.6 maintains an input of $5 and an output of $25 per million tokens, while Sonnet 4.6 maintains an input of $3 and an output of $15 per million tokens. Even if it is a request of 900,000 tokens, its unit token fee is exactly the same as a request of 9,000 tokens.
In this comprehensive opening, Anthropic has systematically upgraded its long context capabilities and user experience. First of all, all requests that support 1M windows enjoy the same rate limit as ordinary requests. Developers do not need to implement additional traffic splitting or downgrading strategies in high concurrency or long conversation scenarios. Secondly, the official has canceled the beta tag header previously required when accessing contexts with more than 200K tokens. Now requests with more than 200,000 tokens can be called directly. Existing applications that are still sending beta headers can make a smooth transition without changing their code.
Media processing capabilities are also one of the focuses of this upgrade. The number of images or PDF pages that Claude can receive in a single request has been increased from 100 to 600, covering complex multi-modal scenarios such as code review, legal documents, scientific research papers, operation and maintenance logs, etc., and has been simultaneously launched on the Claude platform, Microsoft Azure Foundry, Google Cloud Vertex AI and other channels. Anthropic said this adjustment is intended to allow developers to contextualize hundreds of pages of material at once without the need for frequent segmentation and compression, thereby reducing engineering complexity and information loss.
In corporate and professional scenarios, the impact of 1M context is particularly prominent. Officials emphasize that long context is only meaningful if it can be "remembered and used well", so special optimizations have been made on Opus 4.6 and Sonnet 4.6 for long-range memory and reasoning capabilities. In third-party benchmarks, Opus 4.6 achieved a score of 78.3% on MRCR v2, and Sonnet 4.6 scored 68.4% on GraphWalks BFS (1M tokens), leading among cutting-edge models of the same context length.
This means that developers can load a complete code repository, tens of thousands of pages of contract text, or the tool call traces, observation records, and intermediate reasoning processes of long-running agents into a single session, without having to rely on complex "window sliding", "multiple rounds of summarization" or frequent cleaning of historical context. Multiple partners pointed out in citations that 1M context allows the agent to run for hours without losing details, which not only reduces the loss of key information caused by compression, but also reduces the cost of repeated supplements and explanations in multiple rounds of debugging, code reviews, and large-scale operation and maintenance analysis.
In the field of code development, the 1M window of Opus 4.6 has been fully integrated into the Claude Code product line and is open to Max, Team and Enterprise users by default without additional configuration. Developers reported that after expanding the context from 200,000 to 500,000 or even 1 million tokens, the agent was able to complete tasks with less total token consumption when dealing with large code changes and cross-file dependencies, reducing the extra rounds and repeated requests caused by splitting the context.
Operations and security teams are also seen as direct beneficiaries of long-context capabilities. Faced with huge production systems and complex alarm events, engineers can retain all entities, signals, and troubleshooting hypotheses in one session, maintaining a complete link of information from initial alarm to problem mitigation. Partners say that with 1M context, they no longer need to frequently "compress memory" when processing large logs, monitoring data, and multi-source system status, significantly reducing the risk of missing subtle anomalies.
For law firms, scientific research institutions, and data-intensive enterprises, long context windows provide a new knowledge processing paradigm. Legal personnel can introduce multiple versions of hundreds of pages of contracts at once to completely review changes in multiple rounds of negotiations; scientific research teams can simultaneously incorporate hundreds of papers, mathematical formulas, and simulation codes into one reasoning process to build a comprehensive analysis across documents and models. Some collaborating institutions say that this "single integration" capability is significantly accelerating the iterative pace of basic and applied research.
In terms of access channels, 1M context has been enabled on Claude’s native platform and provided through cloud services such as Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, and is open to enterprises and developers simultaneously. Anthropic provides corresponding technical documentation and pricing pages to facilitate the team's evaluation of integration costs and deployment strategies, and emphasizes that all current requests supporting the 1M window will continue to use the existing billing standards.
Anthropic said that as the 1M context enters the fully available stage, the team will continue to iterate on long-range reasoning stability, cross-modal understanding and enterprise-level integration capabilities, hoping to help users shift from "making engineering trade-offs around input length" to "designing agents around the business problem itself." Developers and enterprise users can immediately enable this capability through the Claude platform and major cloud partners to smoothly upgrade existing applications to the long context era.