April 2026 monthly summary for UKGovernmentBEIS/inspect_ai: Token counting accuracy improvement for Google Gemini provider; fixed double-counting of cached tokens and aligned token usage metrics with OpenAI/Anthropic conventions. Major bug fix: Gemini's prompt_token_count included the cached prefix, causing input_tokens to double-count cache hits. Implemented subtracting cached_content_token_count from input_tokens and surfacing it as input_tokens_cache_read. This change reduces telemetry discrepancies, improves reporting accuracy for usage dashboards, and enhances billing precision. Impact: more reliable analytics, better decision-making for provider usage and cost; Maintained code quality with conventional commits and clear documentation. Technologies/skills demonstrated: Python code changes, metrics instrumentation, provider integration, git workflows, and cross-provider metric normalization.