Not affiliated with
Buried in Fable 5's 319-page system card, a silent rule dumbs down answers on pretraining, training infrastructure, or chip design. After 48 hours of backlash, Anthropic apologized, swapped it for a visible reroute to Opus 4.8, and admitted the replacement will overblock.
Anthropic shoved Agent SDK, `claude -p`, Claude Code GitHub Actions, and third-party Agent SDK apps into a separate monthly credit bucket. Subscription caps stayed the same; programmatic Claude usage now burns API-priced credits that vanish fast or turn into a second bill.
Anthropic's status page said Claude Code v2.1.120 crashed on --resume or --continue, forcing a rollback to v2.1.119. Right after the postmortem, the product served another tiny face-plant.
Claude Code's abuse filter flagged the exact-case string HERMES.md in recent commits as shady and routed Max requests to extra billing instead of quota. One Max 20x user burned $200.98 while 86% weekly capacity remained, then had to binary-search git history to find the magic word. HERMES.md is a legit AI-agent convention, not spam.
A second postmortem blamed the Claude Code slump on three tweaks: lower reasoning effort, a cache bug that made the model forgetful, and a brevity prompt that broke coding. Anthropic denied intentional downgrades and reset subscriber limits.
In the Pentagon fight, Anthropic admitted that once Claude sits inside classified environments it cannot monitor, update, or shut itself off. The safety lever everyone imagined does not exist.
Dev spotted Claude Code vanishing from the $20 Pro tier on Anthropic pricing pages, hinting at a jump to the $100 Max plan. Anthropic called it a 2% newbie test and reverted the docs, but the stunt handed OpenAI free ammo and made pricing look like quicksand.
Anthropic suspended over 60 Belo employees' Claude accounts for a vague policy breach, cutting off workflows and chat history. Access returned after about 15 hours as a false positive, but the only appeal was a generic Google Form.
OX Security said Anthropic's MCP STDIO design could let attackers run commands across millions of users and hundreds of thousands of servers. Researchers claim Anthropic refused a protocol fix, calling the behavior expected and shoving sanitization onto devs.
Anthropic's status page said Claude.ai and Platform were down, Claude Code login failed, and API, Console, and Code were all hit before recovery. Another day, another 'is Claude broken?' meme.
For 33 days in February Claude Code cached every prompt for an hour. On March 6 it silently dropped to 5 minutes, burning cash as cache expired mid-session and rebuilt 12.5 times faster. Three months later Sonnet users were out $949 each, Opus users $1,582. Anthropic admitted "bug" only after 120,000 API calls proved it.
TechCrunch reported Anthropic briefly banned OpenClaw maker Peter Steinberger even after the new API-payment option arrived. The company later unbanned him, but the optics screamed walled-garden chaos.
Anthropic told subscribers that Claude limits no longer cover third-party wrappers like OpenClaw; users had to burn API keys or pay separate usage bills. Build on Claude, then pay again when your tool gets hot.
An AMD AI leader filed a GitHub report saying Claude Code had regressed so badly it could not be trusted for complex tasks, backing it with thousands of sessions. The post helped turn gripes into a data-driven revolt before Anthropic admitted several product tweaks had broken things.
Trying to stuff the leak back in, Anthropic's takedown spree nuked thousands of GitHub repos, including some that had only forked the official Claude repo. The company later called the collateral damage accidental and reversed much of it, but bogus DMCA claims are risky and probably illegal.
A source-map snafu exposed roughly 500K lines of Claude Code guts, including architecture and unreleased features. Anthropic swears no customer data spilled, which is not the same as 'no harm done.'
Anthropic rolled out Claude Code Review and told teams to expect about $15-25 per review, billed by the token. Devs quickly compared it with CodeRabbit at $24/month or Greptile at $30 for 50 reviews plus $1 each after that.
Anthropic said it got a March 4 letter branding it a supply-chain risk, then sued for emergency relief. A model-policy spat morphed into a defense-contracting death match.
Usage spiked and Claude fell over in early March, followed by a run of status incidents. Anthropic's own page later showed sub-99% 90-day uptime for claude.ai, another reliability punchline.
Dario Amodei said Anthropic would not gut safeguards for mass domestic surveillance or fully autonomous weapons, even as the Pentagon threatened to yank contracts and invoke the Defense Production Act. The safety brand finally met procurement brass.
Anthropic accused DeepSeek, Moonshot, and MiniMax of industrial-scale distillation after 24,000 fake accounts pumped out 16M Claude exchanges. Critics noted DeepSeek's slice was only 150K and Theo Browne said 16M is routine traffic for big chatbots.
Anthropic disclosed banning 1.45 million accounts from July-December 2025, plus 52,000 appeals and 1,700 reversals. The machinery is now visible; from the outside it still looks like 'fill the Google form and pray.'
Anthropic griped over Clawdbot's Claude-style branding, forcing Steinberger to rebrand to Moltbot, then OpenClaw. Weeks later he joined OpenAI to build personal agents.
xAI staff reportedly lost Claude access inside Cursor after Anthropic enforced its no-competitor clause. After Windsurf and OpenAI, the ban list looked like standard policy, not a one-off.
Anthropic says a China-tied group tricked Claude Code into probing around 30 targets, letting AI run most of the op. The agentic coding pitch met an agentic break-in.
A judge preliminarily green-lit a $1.5 billion deal covering about 465,000 books at roughly $3,000 each. Anthropic ducked trial on the shadow-library claim, but the receipt is now public record.
Weeks of users yelling later, Anthropic said three infra bugs degraded answers from August into early September and explained why evals missed it. Useful confession, lousy timing.
An outage knocked out Claude.ai, Console, and API. After a year full of rate-limit whining, the reliability saga earned another screenshot.
Anthropic revealed Claude misuse tied to data extortion, North Korean fake job schemes, and AI-built ransomware. The safety-first model was already handy for crooks too.
Anthropic yanked OpenAI's Claude access, claiming OpenAI used Claude Code to benchmark GPT-5 in violation of anti-rival rules. OpenAI called it routine safety work; Anthropic played bouncer.
Anthropic capped weekly usage for Pro and Max tiers, blaming always-on coding loops and account sharing. Fewer than 5% of users were supposed to feel it, but the real message was flat-rate coding had crashed into the compute budget.
Claude Code began as Boris Cherny's hacky terminal toy and stayed a limited preview while Cat Wu said a dedicated subscription was still TBD. Cursor-maker Anysphere lured them away, then Anthropic yanked them back within two weeks. Claude Code became a paid product soon after.
Judge William Alsup said training on legally acquired books may be fair use, but the pirate-library claims stayed alive. Anthropic scored the legal theory and kept the shadow-library stink.
Reddit sued Anthropic, alleging bots kept hammering Reddit after Anthropic promised to stop and that Claude trained on user posts without permission. Unlike the book fights, Reddit framed it as breach of contract and privacy promises.
Windsurf says Anthropic slashed first-party Claude access with almost no warning, right as OpenAI buyout rumors swirled. Jared Kaplan later claimed compute limits, not games, drove the cutoff, which did not make the lockout feel friendlier.
In the music-publisher case, Anthropic's lawyers had to admit Claude hallucinated a citation in an expert report. The model's glitch became part of the evidence against the model.
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson say Anthropic hoovered books from pirate libraries and baked them into Claude's brain. Every 'responsible AI' slide now shares the stage with that allegation.
ClaudeBot slammed iFixit about a million times in a day while other sites screamed about the bot's appetite. Anthropic swears it obeys robots.txt, which only helps after you catch the thing chewing your server.
Universal, Concord, and ABKCO claim Claude scarfed copyrighted lyrics and can spit them back on demand. Constitutional AI's first big courtroom chorus came from the music biz.