Anthropic Deploys Prompt Caching & Computer Use API: 90% Cost Reduction for Frontier Agentic Workflows

The Three-Question Framework

What Happened?

Anthropic officially launched two major capabilities across its Claude API: Prompt Caching and Computer Use. Prompt Caching allows developers to checkpoint multi-megabyte codebases, system documentation, and conversational histories in GPU cache with a 5-minute time-to-live, slashing cached input token prices by 90% and reducing latency by up to 80%. Concurrently, the Computer Use API enables Claude 3.5 Sonnet to interact directly with GUI applications via normalized screen coordinates and keystroke events.

Why It Matters

Prior to prompt caching, autonomous software engineering agents faced prohibitive compounding costs, as every iterative tool call required re-transmitting the entire codebase context window. By eliminating redundant token reprocessing, agentic workflows that previously cost $15 per debugging session now run for less than $1.50.

What Could Happen Next

Enterprise IDE vendors including JetBrains and VS Code extensions have integrated prompt-caching headers natively, while cybersecurity teams establish strict egress firewall rules to restrict Computer Use agent sandboxes from corporate networks.

Executive Key Takeaways

Prompt Caching reduces cached prompt token costs by 90% and decreases first-token latency by up to 80%.

Cached tokens maintain a rolling 5-minute time-to-live (TTL), refreshed automatically with each subsequent request.

Computer Use API allows Claude 3.5 Sonnet to take screenshots, move cursors, click buttons, and type in arbitrary desktop apps.

Enterprise compliance: supports granular API permission scopes, enabling administrators to whitelist specific application windows.

On September 7, 2026, Anthropic rolled out two transformative additions to its developer platform: Prompt Caching and direct OS-level Computer Use for Claude 3.5 Sonnet.

For enterprise software teams building autonomous agent loops, prompt caching addresses the single largest cost bottleneck in generative AI. In complex workflows like automated repository refactoring or legal contract review, developers repeatedly send hundreds of thousands of identical tokens—representing API schemas, test harnesses, and documentation—with every iterative question. Anthropic's caching architecture keeps the compiled attention key-value states in memory across API requests. When a prompt hits a cached breakpoint, Anthropic bills those tokens at a 90% discount ($0.30 per million tokens on Sonnet compared to standard $3.00), while reducing Time-To-First-Token (TTFT) by up to 80%.

Simultaneously, the Computer Use API unlocks direct graphical user interface manipulation. Rather than relying on rigid JSON API connectors, Claude can view a computer screen through sequential high-resolution screenshots, calculate precise (x, y) coordinate targets, click UI buttons, enter keyboard input, and navigate legacy software that lacks modern APIs.

On initial developer evaluations, Claude 3.5 Sonnet achieved strong scores on OSWorld and internal computer task benchmarks, outperforming human novices on multi-step spreadsheet and web browser data entry tasks.

Anthropic stressed that Computer Use remains in experimental release, mandating that developers isolate instances within virtual machines (VMs) without production internet connectivity to prevent prompt-injection attacks from malicious web pages.