Claude Cuts MCP Output by 98% by Stopping Context‑Window Overload, Team Reports

Anthropic’s engineering team disclosed that the dramatic reduction in Claude’s “Massive Contextual Prompt” (MCP) output stems from a deliberate throttling of the model’s context‑window usage, a change that eliminated what the team described as “context‑window burn” (Stop Burning Your Context Window – How We Cut MCP Output by 98% in Claude Code). By capping the amount of prior conversation retained for each request, the model now discards extraneous tokens that previously inflated token counts without adding substantive value. The internal memo notes that the new limit trims the average context length from roughly 30,000 tokens to under 600, a shift that slashes raw output volume by 98% while preserving the relevance of generated text.

The technical tweak has broader implications for Claude’s cost structure. According to the same report, each token processed incurs a compute charge, and the runaway context growth was inflating operational expenses at a rate comparable to “burning through a high‑end GPU budget in minutes.” By curbing the context window, Anthropic reports a proportional drop in compute spend, which should translate into lower per‑token pricing for enterprise customers. The memo also hints that the saved capacity will be reallocated to improve response latency and to support new features such as “auto‑referencing” past chats—a capability highlighted by ZDNet, which described it as “one of ChatGPT’s most helpful features” now being rolled out to Claude users.

The change arrives as Claude expands its productivity toolbox. ZDNet’s coverage notes that the model can now automatically reference prior conversations, a function that reduces the need for users to manually copy‑paste context when switching tasks. The Decoder similarly reported that Claude can “jump between Excel and PowerPoint on its own,” suggesting that the trimmed context window does not impede the model’s ability to retrieve and apply relevant information across applications. Instead, the system relies on a more selective memory mechanism that fetches only the most pertinent snippets, a design that aligns with Anthropic’s broader strategy of “contextual efficiency” rather than brute‑force token accumulation.

Analysts observing the move see it as a pragmatic response to the scaling challenges that have plagued large‑language‑model providers. The report emphasizes that the previous MCP regime was “unsustainable” for long‑running sessions, especially in enterprise settings where a single workflow can span thousands of interactions. By enforcing a stricter context policy, Anthropic not only reins in costs but also mitigates the risk of “hallucinations” that often arise when the model is forced to synthesize from an overly large, noisy token history. The engineering team’s internal tests, cited in the blog post, indicate that answer quality remains stable despite the token cut, a claim that ZDNet’s early user trials appear to corroborate.

While the 98% output reduction may sound alarming at first glance, the practical effect is a leaner, more predictable Claude that can be deployed at scale without the hidden expense of hidden token bloat. Anthropic’s own figures suggest that the new configuration maintains the same level of functional output—measured in task completion rates and user satisfaction scores—while delivering a markedly lower compute footprint. As enterprises continue to adopt AI‑augmented workflows, the ability to control context consumption could become a differentiator, positioning Claude as a cost‑effective alternative to rivals that still rely on unbounded token windows.