Key Takeaways
- Anthropic identified tighter limits during peak hours and sessions with million-token-plus contexts as the two main reasons Claude Code users burn through their allocations faster than expected.
- Anthropic’s Lydia Hallie recommends using Sonnet 4.6 instead of Opus, as Opus burns through limits roughly twice as fast.
- Anthropic fixed some bugs but confirmed none of them led to incorrect billing.
- Practical tips include turning off Extended Thinking when not needed, starting fresh sessions, and limiting the context window.
What Happened
Anthropic, the AI safety company, publicly addressed complaints from Claude Code users who were hitting their usage limits far faster than expected. Lydia Hallie, an Anthropic employee, published an explanation on April 3, 2026, identifying the root causes and offering practical advice to reduce token consumption, as reported by The Decoder.
According to Hallie, two factors drive the rapid usage drain: tighter rate limits during peak hours and sessions where context windows balloon past one million tokens. Both consume a user’s allocation faster than the steady-state usage most subscribers expect.
Why It Matters
Claude Code has become one of the most popular AI coding tools since its launch, competing directly with GitHub Copilot, Cursor, and OpenAI’s Codex. Usage complaints threaten Anthropic’s position in this market. Developers who hit limits mid-session experience frustration and productivity loss, which pushes them toward competitors with more predictable usage models.
The problem also highlights a broader tension in AI product design. AI coding assistants encourage long, complex sessions where the model maintains context across an entire codebase. Those sessions generate enormous token counts, often exceeding one million tokens, which is expensive for the provider and unpredictable for the user. Every AI coding tool company is wrestling with this same cost-versus-experience tradeoff.
Technical Details
Hallie’s explanation identified two primary causes of unexpectedly fast usage drain. First, Anthropic applies tighter rate limits during peak hours, which means the same amount of work consumes a larger share of a user’s allocation when demand is highest. Second, context windows that grow past one million tokens consume dramatically more compute per interaction, as each new message must process the entire accumulated context.
Anthropic also confirmed it fixed some bugs, though Hallie stated none of these bugs resulted in incorrect billing. The company has shipped efficiency improvements and added in-product pop-ups to alert users when their consumption is elevated.
Hallie offered several specific recommendations to reduce usage:
- Use Sonnet 4.6 instead of Opus. Opus burns through limits roughly twice as fast as Sonnet, making it the single most impactful change users can make.
- Turn off Extended Thinking when deep reasoning is not required for the task at hand.
- Start fresh sessions instead of continuing old ones, which prevents context windows from ballooning.
- Limit the context window manually to keep sessions focused and efficient.
Users who still notice unusually high usage after applying these tips should report it through Anthropic’s in-product feedback function.
Who’s Affected
All Claude Code subscribers are affected, particularly those on Pro plans who use Opus as their default model. Power users who run long coding sessions with large codebases are the most heavily impacted, as their sessions are most likely to accumulate million-token-plus contexts.
Developers working during US business hours, which coincide with Claude Code’s peak usage periods, face tighter limits than those working during off-peak times. This creates an uneven experience depending on geography and work schedule.
The issue also affects Anthropic’s competitive positioning. Cursor and GitHub Copilot offer more predictable usage models, and developers frustrated by unpredictable Claude Code limits may switch. OpenAI’s recent move to usage-based pricing for Codex adds another alternative.
What’s Next
Anthropic is likely to continue shipping efficiency improvements and better usage transparency tools for Claude Code. The company’s decision to address the complaints publicly suggests it recognizes the issue as a retention risk.
Longer term, Anthropic may need to restructure Claude Code’s pricing to better match how developers actually use the tool. The current flat-rate model with variable effective limits based on peak hours and context sizes creates unpredictability that frustrates users. A move toward more transparent, usage-based pricing, similar to OpenAI’s recent Codex pricing shift, may become necessary.
