Back to Blog
Token Insights March 25, 2026 · 4 min read

7 Things Anthropic Doesn't Tell You About Your Claude Code Usage

You're paying $20-200/month for Claude Code but Anthropic gives you almost no usage data. Here are the 7 critical metrics you're missing — and why they matter.

By Champlin Enterprises

You're Flying Blind

Claude Code is arguably the most powerful AI coding tool available today. Opus 4.6 can reason about complex architectures, refactor entire codebases, and write production-quality code. But Anthropic has a blind spot: they give you almost no visibility into your own usage.

Visit claude.ai/settings/usage and you'll see a single progress bar showing what percentage of your plan you've consumed. That's it. No breakdown by model, project, session, or time period. No cost estimates. No trends.

Here are the 7 things you're missing — and why each one matters.

1. Per-Session Token Counts

Every Claude Code session consumes a different number of tokens depending on complexity, context size, and model. A quick code review might use 15K tokens. A multi-file refactoring session can burn 300K+. But Anthropic doesn't show you session-level data.

Why it matters: Without session-level data, you can't identify which types of tasks are the most expensive. You might be burning through your plan on tasks that don't need Opus when Sonnet would produce identical results at 3x lower token cost.

2. Per-Project Cost Breakdown

If you work on multiple projects — which most developers do — you have no idea which project is consuming the most tokens. Anthropic treats all usage as one undifferentiated pool.

Why it matters: For freelancers billing clients, this is a deal-breaker. For teams, it means you can't allocate AI budgets by project. For anyone, it means you can't optimize where AI adds the most value.

3. Model-by-Model Analytics

Claude Code uses different models — Opus 4.6 for complex reasoning, Sonnet 4.6 for faster routine tasks, and Haiku 4.5 for lightweight operations. Each has dramatically different token economics. Opus costs roughly 3x more per token than Sonnet in effective plan usage.

Why it matters: If 60% of your usage is Opus but half those tasks could be done with Sonnet, you're burning through your plan allocation 1.5x faster than necessary. But you'll never know this without model-level tracking.

4. API-Equivalent Cost Comparison

Every Pro ($20/month) or Max ($100-200/month) subscriber should know the answer to one question: am I saving money with a subscription vs. paying per-token at API rates? The answer varies dramatically by usage pattern.

Why it matters: Some heavy users save 10x or more vs. API pricing. Light users might be overpaying. Without knowing your actual API-equivalent cost, you can't evaluate whether your subscription is a good deal.

5. Daily and Weekly Usage Trends

Your Claude Code usage isn't constant. You probably use more on coding-heavy days and less during meetings or planning. But Anthropic shows no historical trends — just a current snapshot that resets with your billing cycle.

Why it matters: Trends reveal patterns. If your usage is climbing week-over-week, you might be approaching your plan limits and should consider upgrading before you start hitting rate limits during critical work. If it's declining, you might be overpaying.

6. Cache Hit Rates

Claude Code uses prompt caching aggressively — when you send similar context across turns in a conversation, cached tokens are dramatically cheaper (essentially free on subscription plans). But Anthropic doesn't show you how much of your usage comes from cache hits vs. fresh tokens.

Why it matters: High cache hit rates mean your effective cost per session is much lower than the raw token count suggests. Understanding your cache utilization helps you structure prompts for maximum efficiency — reusing context rather than starting fresh conversations for related tasks.

7. Rate Limit Proximity

The most frustrating Claude Code experience is hitting a rate limit mid-task. You're deep in a complex refactoring session, Claude is making great progress, and suddenly — nothing. Wait and retry. Your flow state is destroyed.

Why it matters: If you could see how close you are to your rate limit in real time, you could pace your usage. Save the Opus-heavy session for when you have headroom. Do the quick Sonnet tasks when you're approaching the limit. Instead, Anthropic gives you no warning until you're already blocked.

The Solution: Track Everything

MyTokenTracker fills every one of these gaps. After a 10-second install, your dashboard shows:

  • Per-session token counts with model, project, and duration
  • Per-project cost breakdown with visual donut chart
  • Model distribution analytics (Opus vs Sonnet vs Haiku)
  • API-equivalent cost comparison (subscription savings calculator)
  • Week-over-week and month-over-month trends
  • Cache read/write breakdowns per session
  • Real-time usage tracking via background daemon

The free plan includes all of these features with 100 logs per day and 7-day retention. Most individual developers never need more.

You wouldn't run a business without financial tracking. You wouldn't deploy code without monitoring. Why are you using the most expensive AI coding tool on the market without usage analytics?