Anthropic 承认 Claude Code 用户用量超预期提前触顶

AI/ML

Users of Claude Code, Anthropic's AI-powered coding assistant, are experiencing high token usage and early quota exhaustion, disrupting their work.

Anthropic has acknowledged the issue, stating that "people are hitting usage limits in Claude Code way faster than expected. We're actively investigating... it's the top priority for the team."

A user on the Claude Pro subscription ($200 annually) said on the company's Discord forum that "it's maxed out every Monday and resets at Saturday and it's been like that for a couple of weeks... out of 30 days I get to use Claude 12."

The Anthropic forum on Reddit is buzzing with complaints. "I used up Max 5 in 1 hour of working, before I could work 8 hours," said one developer today. The Max 5 plan costs $100 per month.

There are several possible factors in the change. Last week, Anthropic said it was reducing quotas during peak hours, a change that engineer Thariq Shihipar said would affect around 7 percent of users, while also claiming that "we've landed a lot of efficiency wins to offset this."

March 28 was also the last day of a Claude promotion that doubled usage limits outside a six-hour peak window.

A third factor is that Claude Code may have bugs that increase token usage. A user claimed that after reverse engineering the Claude Code binary, they "found two independent bugs that cause prompt cache to break, silently inflating costs by 10-20x." Some users confirmed that downgrading to an older version helped. "Downgrading to 2.1.34 made a very noticeable difference," said one.

The documentation on prompt caching says that the cache "significantly reduces processing time and costs for repetitive tasks or prompts with consistent elements." That said, the cache has only a five-minute lifetime, which means stopping for a short break, or not using Claude Code for a few minutes, results in higher costs on resumption.

Developers can upgrade the cache lifetime to one hour but "1-hour cache write tokens are 2 times the base input tokens price," the documentation states. A cache read token is 0.1 times the base price, so this is a key area for optimization.

Anthropic does not state the exact usage limits for its plans. For example, the Pro plan promises only "at least five times the usage per session compared to our free service." The Standard Team plan promises "1.25x more usage per session than the Pro plan." This makes it hard for developers to know what their usage limits are, other than by examining their dashboard showing how much quota they have consumed.

Problems like this are not unusual. Earlier this month, users of Google Antigravity were protesting about similar issues.

Bugs aside, what we are seeing is an implicit negotiation between users and providers over what is an acceptable pricing and usage model for AI development. Users want to control costs and providers need to make a profit. There is also a disconnect between vendor marketing that urges developers to insert AI into every process, including in some cases automated workflows, and a quota system that can cause AI tools to stop responding.

"For folks running Claude Code in automated workflows: rate-limit errors need to be caught explicitly – they look like generic failures and will silently trigger retries. One session in a loop can drain your daily budget in minutes," observed one user.

When you just need to build a tiny desktop app: TinyJS

October 13, 2026

Valen creator drives 'Golden Spike' to connect new languages with Rust

September 26, 2026

Microsoft agentically ports Copilot runtime to Rust for $120K

September 19, 2026

Apple iPhone Duo makes developers think in folds

September 16, 2026

JavaScript installer pnpm recast in Rust because ECMAScript can't keep up

September 5, 2026

Kubernetes cleans house, bins legacy kube-dns, IPVS, and cgroup v1

August 27, 2026

How Cursor beat Git's scalability shortcomings

August 24, 2026

Go updates may delight diehard gophers but displease AI overlords

August 21, 2026

SvelteKit 3 puts heat on Next.js with radical approach to RPCs

August 20, 2026

Almost nobody pays attention to web standards anymore

August 18, 2026

ChainDrop worm crawls into npm supply chain, evades standard defenses

August 16, 2026

Node.js creator liberates Durable Objects from Cloudflare

August 13, 2026

AI titans to tidy agent frontier with plugin prescription

August 8, 2026

Platform Engineering 2.0: your platform was built for a different era. AI just exposed it

August 6, 2026

Next.js 16.3 aims to reduce dreaded FATAL ERROR messages

August 5, 2026

Dev who gave HashiCorp its name returns with a faster terminal multiplexer

August 1, 2026

How AI drove Shopify back to clean code

Turns out, agents just want the same things as humans: easily-readable code, explicit contracts, and helpful feedback

July 26, 2026

Zig creator calls Bun's Claude Rust rewrite 'unreviewed slop'

July 15, 2026

HTTP gets a QUERY method so complex searches can stop pretending to be POST

July 14, 2026

Speedier type checks in TypeScript 7.0 as first stable Go release ships

July 10, 2026

GitHub cuts short offer to burn repos on CD after mockery ensues

A purported jab of Sony's physical media phase-out blows up on GitHub itself

July 7, 2026

.NET's long term support is not long term enough, dev complains

June 30, 2026