Claude Code News Roundup — Anthropic 캐시 축소와 토큰 예산 위기
The biggest story this week is Anthropic quietly shrinking Claude Code's 1-hour cache without warning, leaving users reporting skyrocketing token consumption. On the release front, v2.1.92 fixed fullscreen scroll duplicate message bugs, while the community is rallying around CodeBurn, an open-source tool that breaks down token costs by task. If you're building with Claude Code, auditing your cache policy and token budget right now isn't optional—it's survival.
Claude Code News Roundup — 2026-04-21
Anthropic Quietly Kills Claude Code's 1-Hour Cache — Token Budget Exploding Across User Base
Anthropic has silently pulled back (or removed entirely) Claude Code's 1-hour prompt cache with zero announcement, and users are now reporting their daily token budgets vanishing way faster than before. According to XDA Developers, this change happened with no official release notes or heads-up—the real killer is that costs have noticeably spiked compared to when the cache was working. Long sessions and power users grinding repetitive tasks feel the pain most. Anthropic hasn't said a word publicly about it yet.
📢 Official Anthropic Updates (3+ Stories)
Limited new official Anthropic announcements dropped in the last 24 hours, but here's what's confirmed live since 2026-04-20:
Claude API Models Overview Page Refresh (platform.claude.com)
- Date: 2026-04-21 (4 hours ago)
- What: The
platform.claude.com/docs/en/about-claude/models/overviewpage just got refreshed with updated performance comparison data across the Claude model lineup. - Dev impact: If you're integrating via API, double-check the latest docs to see if your model selection criteria has shifted.
Bedrock/Vertex/Foundry 429 Error Reference Fix
- Date: 2026-04-21 (GitHub Releases, 10 hours ago)
- What: Release notes confirm a bug where Bedrock, Vertex, and Foundry users were getting 429 rate-limit errors pointing to the wrong status page (
status.claude.com—which only applies to Anthropic's direct cloud offerings). - Dev impact: Teams using Claude Code via AWS Bedrock, Google Vertex AI, or other cloud platforms need to review error-handling logic and status check URLs.
claude-code-action v1 Launch with Interactive & Automation Modes
- Date: 2026-04-14–21 (1 week ago, GitHub release)
- What:
anthropics/claude-code-action@v1went live with two modes: Interactive (responds to@claudementions) and Automation (instant-execute prompt-based). Supports CLI flags like--model claude-4-0-sonnet-20250805and--max-turns 10for CI/CD pipelines. - Dev impact: You can now plug Claude Code straight into GitHub Actions for PR reviews, security audits, and other repetitive workflows.
🛠 Claude Code Releases & GitHub Moves (2+ Stories)
v2.1.92 — Fixed Fullscreen Scroll Duplicate Message Bug
- Version/Date: v2.1.92 / 2026-04-21 (≈10 hours ago)
- What fixed:
- Fullscreen mode was showing the same message twice when scrolling up (hit iTerm2, Ghostty, and all DEC 2026-compliant terminals).
- Bedrock/Vertex/Foundry 429 errors were pointing to
status.claude.comby mistake—now fixed.
- How to upgrade: Run
npm update -g @anthropic-ai/claude-codeor let the app auto-update. iTerm2 and Ghostty users especially should grab this one now.
Show HN: CodeBurn — Task-Level Token Usage Analyzer from Session JSONL
- Release Date: 2026-04-18 (3 days ago, Hacker News)
- What:
- Reads JSONL session transcripts stored in
~/.claude/projects/and sorts each turn into 13 cost categories. - Unlike the older
ccusagetool (which only shows model/date costs), CodeBurn breaks it down by task. - No API key needed; runs entirely local.
- Reads JSONL session transcripts stored in
- Why it matters: The tool hit HN right as the cache nuke story broke, and suddenly everyone needs visibility into where their tokens actually go. Community saw a gap and filled it fast.
📰 Tech Media Coverage (2+ Stories)
"Anthropic quietly nerfed Claude Code's 1-hour cache, and your token budget is paying the price"
- Outlet / Date: XDA Developers / 2026-04-21 (14 hours ago)
- Gist: Anthropic killed Claude Code's 1-hour prompt cache without notice, and users are now hitting token limits way sooner than expected. The story traces back to community reports from the last few days.
- Key insight: The cache policy change never hit any public roadmap or release notes—reviving concerns about Anthropic's "silent change" playbook and breaking trust on cost predictability.
"Issue: Claude Code is unusable for complex engineering tasks with Feb updates" (HN Thread)
- Outlet / Date: Hacker News / 2026-04-14–21 (1 week ago, still hot)
- Gist: An engineer posted detailed analysis showing Claude Code's quality tanked for complex engineering work after the Feb update. An Anthropic staffer actually jumped into the thread to clarify and name two root causes.
- Key insight: Anthropic engineers responding publicly to criticism in the wild is rare. It signals community feedback is genuinely moving the needle on product direction.
💬 Developer Community Reaction (3+ Stories)
"Claude Code and the Great Productivity Panic of 2026" — HN Thread
- Source / Vibe: Hacker News / 2026-04-07 (2 weeks ago, still quoted heavily)
- Debate: Commenters are torn between 10-hour flow-state deep coding vs. 3-hour agent-assisted sessions. One thread argues that agent coding—moving faster, making more decisions per minute—actually exhausts you faster mentally. Is Claude Code a productivity boost or just a new flavor of cognitive overload?
"CodeBurn — Task-Level Token Analysis" Show HN Thread
- Source / Vibe: Hacker News / 2026-04-18 (3 days ago)
- Debate: A dev who was frustrated that
ccusageonly showed model/date breakdowns built a tool to analyze by task instead. With the cache nuke dropping days later, demand for token visibility exploded. Timing is everything.
"Claude Code is unusable for complex engineering tasks" — Anthropic Staffer Jumps In
- Source / Vibe: Hacker News / 2026-04-14–21 (1 week ago)
- Debate: When an Anthropic engineer showed up to say "we hear you, there are two factors at play," the response was mixed—some saw it as refreshing transparency, others saw it as band-aid on a bullet wound. Fundamental performance fix still pending.
🔍 Comparison & Analysis
The elephant in the room this cycle is transparency. A silent 1-hour cache deletion undercuts Claude Code's whole pitch against GitHub Copilot or Cursor—predictable costs. Cursor is doubling down on editor integration; Claude Code is playing a two-track game of terminal-first agents plus desktop GUI expansion (sidebar, routines). GitHub Copilot leans on VS Code depth and enterprise licensing. Claude Code's edge? Multi-session flexibility, routine automation, and CLI power. But if silent cost tweaks keep happening, enterprises that live and die by budget predictability will bail. The speed at which CodeBurn got built and upvoted proves the community sees a real gap in official token visibility. If Anthropic doesn't plug that hole soon with first-party dashboards or CLI tools, open-source community tools will own the narrative.
🧭 Practical Tips for Developers (3+ Tips)
-
Audit your token spend right now: Cache shrinkage means per-session token burn is spiking. Fire up
ccusageor grab CodeBurn (parses~/.claude/projects/JSONL files) and find out which task categories eat the most tokens. Then optimize. -
Update to v2.1.92 immediately: Run
npm update -g @anthropic-ai/claude-code. Fixes the iTerm2/Ghostty fullscreen duplicate message mess and the Bedrock/Vertex 429 error reference bug in one go. -
Wire up GitHub Actions with claude-code-action v1: Drop this into your Actions workflow to auto-run reviews and security checks when someone mentions
@claudeon a PR:yaml- uses: anthropics/claude-code-action@v1 with: anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
👀 What to Watch Next Cycle
- Official Anthropic statement on the cache cut: Whether they explain the 1-hour cache removal or drop new caching guidelines is the #1 question. If token complaints keep piling up, expect a pricing policy response.
- Performance fix for "complex engineering tasks": The HN thread had an Anthropic engineer mention two root causes. Which version ships the actual fix?
- CodeBurn or a similar tool going official: Community demand for task-level token breakdown is already huge. Anthropic might integrate this into the CLI or a web dashboard sooner than you'd think.
📌 Action Items This Week
- Right now: Run
npm update -g @anthropic-ai/claude-codeto grab v2.1.92. If you're on Bedrock/Vertex/Foundry, audit your error-handling code for hardcodedstatus.claude.comreferences. - Today: Install CodeBurn (search GitHub: "CodeBurn claude-code"), point it at your
~/.claude/projects/logs, and map out which workflows burn the most tokens. Adapt your patterns to survive the cache-less future.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
