Claude Skills, Files, and Browser Use Are GA: Recalculate Agent Cost per Task

Anthropic has made computer use, the Skills API, and the Files API generally available on the Claude Platform, while adding a browser-use tool and multi-action turns. The economic effect is not one simple discount: fewer round trips and reusable assets can lower tokens, but richer tools can also expand the number of tasks an agent attempts.

The August 20 Anthropic announcement says computer use can now take several actions per turn, browser use combines screenshots with page structure, skills can be uploaded and versioned, and files can be referenced by ID rather than resent. Anthropic also says Files API limits are five times higher and organizations receive 1 TB of storage.

Three Different Cost Levers

Multi-action turns reduce orchestration round trips. Instead of model call, click, screenshot, and another model call for each field, one turn can produce several actions. Files reduce repeated upload and prompt assembly, although the referenced content can still contribute tokens when read. Skills move stable instructions, scripts, and templates into versioned packages that load when relevant instead of bloating every system prompt.

Treat these as separate experiment variables. A lower call count does not guarantee lower output tokens, and a stored file does not mean free model context. Export calls, input tokens, cached tokens, output tokens, tool actions, browser minutes, and completion status. Compare the same workflow before and after each feature.

A Browser Workflow Example

Imagine a coding agent that deploys a preview, opens the application, completes six form steps, downloads a report, and checks the result. A one-action loop might require 18 model calls after navigation and recovery. If every call processes 35,000 input tokens and 2,000 output tokens, the run uses 630,000 input and 36,000 output tokens.

If multi-action turns cut calls to eight while average input rises to 45,000 because each turn carries more state, total input falls to 360,000 and output might fall to 24,000. On Claude Sonnet 5 at $2/M input and $10/M output, inference drops from $1.62 to $0.96. The actual result depends on retries and cache hits, so measure rather than applying the call-count reduction directly.

Make Skills and Files Reusable

• Version skills with code. Record the skill version beside each run so cost or quality changes can be traced.
• Separate stable from dynamic content. Stable procedures and templates belong in skills; task facts belong in the request.
• Reuse file IDs. Upload common specifications once and manage expiration intentionally instead of attaching duplicates.
• Track storage lifecycle. Automatic expiration prevents silent accumulation, but applications must handle missing or renewed files.

Reliability Changes the Denominator

Anthropic cites a customer whose longest claims workflow reportedly fell from 32 minutes to 13, with cost per task down about 30% and completion reaching 100% in its testing. That is a vendor-provided customer example, not a universal benchmark. Still, it highlights the correct denominator: cost per completed workflow. A tool that costs more per turn can be cheaper if it eliminates recovery loops.

Build an evaluation set containing normal pages, slow pages, changed layouts, authentication prompts, and partial failures. Require evidence for the final state, not merely a successful tool response. Browser agents can confidently act on the wrong control, and retries can erase the savings from multi-action turns.

Watch for Scope Expansion

Lower friction can increase total spend because product teams automate workflows that were previously uneconomical. Separate unit economics from volume. Report cost per verified completion, total completions, and total monthly cost. A 30% unit-cost reduction paired with ten times as many valuable tasks is a success even though the invoice grows.

Set task-class budgets rather than one universal cap. A short file transformation should stop quickly; a browser workflow crossing several legacy applications may need more actions and time. Define maximum model tokens, browser minutes, files read, and side effects for each class, then require an explicit escalation path when the budget is exhausted.

Control Version and Retention Drift

A skill update can change cost and behavior across every attached workflow. Release skills through evaluation and staged rollout, keep a rollback version, and record the exact version on each run. The same discipline applies to browser-tool and model snapshots. Without it, an invoice change cannot be separated from a workload change.

Files can contain code, customer data, and generated artifacts. Give them owners and expiry policies, and avoid using a generous storage allowance as a retention target. Deleting unused files reduces governance surface even when storage is bundled. Verify that a workflow can recreate or re-upload required inputs after expiration without silently falling back to stale local copies.

Bottom Line

Claude's newly GA tools can reduce repeated context and browser round trips, but only a task-level ledger will show the benefit. Version skills, reuse files, measure actions and tokens separately, and include browser runtime and failure recovery. Optimize for finished, verified work rather than the fewest model calls.