Claude Fable 5.1 ships in Claude Code with cache reads 75% cheaper, and drains plan limits fast. Plus: Anthropic walks back 30-day retention, ChatGPT's app bundles LibreOffice, one git flaw hijacks seven agents, a Claude Code bug erased heritage records, and Copilot can approve PRs.
Anthropic shipped its biggest update yet: Claude Fable 5.1, a model that more than doubled its predecessor's score on agentic science tasks and cut its cache-read price 75% doing it, though not its per-token rate. Within hours the conversation had moved to a different number: how fast the new model empties a subscriber's usage window, and why Pro subscribers don't get it at all. The same announcement walked back a data-retention rule enterprise customers had pushed back on since June. Simon Willison, digging through his own cache folder, found OpenAI's ChatGPT desktop app carrying a full copy of LibreOffice. A security firm showed one git-config trick that hijacks seven different coding agents, four findings still unpatched at publication. A Bengaluru heritage project told a newspaper how a Claude Code quoting bug erased five years of records. And GitHub let Copilot's own code review start approving pull requests outright. Seven stories, one 21-hour window.
In this issue•
Claude Fable 5.1 ships in Claude Code, cache reads drop 75%
•
Fable 5.1 drains plan limits fast, and Pro doesn't get it
•
Anthropic walks back the 30-day data-retention rule
•
OpenAI's ChatGPT app bundles a full LibreOffice install
•
One git flaw lets repos hijack seven coding agents
•
A Claude Code quoting bug erased five years of heritage records
•
GitHub lets Copilot's reviews approve pull requests
A new flagship, and what it costs you #
Claude Fable 5.1 ships in Claude Code, cuts cache-read prices 75%
Anthropic shipped Claude Fable 5.1 on Monday, alongside a more restricted sibling, Claude Mythos 5.1. On Terminal-Bench-Science, a test of open-ended agentic science tasks, Fable 5.1 scored 52.6%, more than double Fable 5's 24.7%. On Terminal-Bench 4.0, the general coding benchmark, it lands at 55.8%. Per-token prices hold at $10 in and $50 out per million, but cached reads drop 75% to $0.25, which Anthropic says cuts typical bills about 25% and heavily agentic work up to 45%. Testing lab Artificial Analysis, given pre-release access, put it atop its Intelligence Index, then added a caveat: at max effort on that index it cost 20% more per task than Fable 5, because it used about 1.7 times the output tokens. Simon Willison called the max-effort result "the best pelican I've seen from any of Anthropic's models," at $3.30. It's live in Claude Code and Cursor. Mythos 5.1 stays restricted to vetted cybersecurity and life-sciences users.
Fable 5.1 drains plan limits fast, and Pro subscribers don't get it at all
Anthropic says it reset every subscriber's 5-hour and weekly limits at launch. Within hours users on 5x and 20x Max plans were reporting Fable 5.1 emptying a 5-hour window in under 30 minutes. Anthropic's own help page confirms the shape: on Max, Fable models draw from the same weekly limit as everything else, capped at 50%, and "use them faster than other Claude models." On Pro, Fable 5.1 isn't in the plan; it runs on pay-as-you-go usage credits only.
Anthropic walks back the 30-day data-retention rule enterprises pushed back on since June
The same day, Anthropic reworked the rule enterprises had fought since June: 30-day retention of all Fable and Mythos traffic on Anthropic's servers. Enterprise Frontier Safeguards, built with more than 100 customers including eight consortium banks, stores that monitoring data in the customer's own cloud instead. Kate Jensen, Anthropic's head of Americas, told CNBC the company "spent hundreds of hours" on it. Phased rollout starts this fall, at no charge; eligible customers get zero data retention until then.
What ships quietly inside the app #
OpenAI's ChatGPT desktop app bundles a full LibreOffice install in a 1.7GB cache folder
Simon Willison found that OpenAI's Codex desktop app, since rebranded as ChatGPT, keeps 1.7GB of bundled software there: a full Python install, a full Node.js install, and native binaries for Poppler, git, and LibreOffice. A documents plugin folder holds skills telling Codex how to find and use those binaries. Willison's post stops at the inventory. What the runtime does with a headless office suite is the question the 441-point Hacker News thread is chewing on.
Whose repo is it, really #
A single git flaw lets malicious repos hijack seven coding agents, some still unpatched
Security firm Manifold calls it GitSpawn: agents run git commands for repo context, and several never sanitize the config values they read. A poisoned .git/config
naming a program in a setting like core.fsmonitor
runs that program with the developer's full permissions, before any trust prompt. Cursor, Codex, and Goose are patched. Claude Code closed one vector but left a second open. Grok Build, Hermes, and Qwen Code remain unpatched. Responses ranged from a quick fix to no reply.
A Claude Code quoting bug erased five years of Bengaluru heritage records, and its safety layer blocked the kill
Deccan Herald reports the Mythic Society's inscription-digitization project lost 15% of its records on July 19 and is spending Rs 15 lakh on backups. The project's honorary director, Udaya Kumar P L, filed the mechanism on Claude Code's issue tracker: a cache-clear command whose $1
expanded to nothing, so it ran as rm -rf /*
for four minutes, while the safety classifier twice refused the agent's own attempts to kill it. Filed July 29, still open, no vendor reply recorded.
Letting the agent sign off #
GitHub lets Copilot's code review approve pull requests
GitHub Copilot's code review can now do more than comment: enabled by an admin, its approval counts toward a repository's required-review rules, the same as a human reviewer's. Every review still prints an assessment inline. Acting on it takes an explicit opt-in, at the enterprise, org, or repo level. New commits dismiss a standing approval, just like a person's would. It's off by default, in public preview, and limited to paid Copilot tiers, Pro and up.
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
Also worth your time #
• Anthropic banned my account for "suspicious signals" — a Claude Max subscriber's account was suspended on a template notice citing "suspicious signals," no clause or example given; reinstated the same day the post hit Hacker News (39 points), still with no explanation of what triggered it. Single-sourced, so a link rather than an item, but the opacity is the story and it's checkable: the account visibly changed state.
• Alibaba's Qwen3.8-Max-0902 snapshot claims first place on Code Arena — post-trained "on Coding & Cowork," per Alibaba; TechNode reports the score rose 22 points to 1,691, ahead of Claude Opus 5 at 1,687. Arena's own table marks the result preliminary at 1,390 votes with a rank spread of 1 to 4, the same spread it gives Opus 5, so on Arena's own numbers this is a tie, not a lead.
Ones to watch #
Early and unverified. Engagement, not endorsement.
• ihavebeenclawed.com — a new, community-submitted index of coding-agent incidents (data deletions, leaked secrets, runaway bills), built by developer nezhar and framed by its own creator as "a curated sample, not a census." Show HN, 12 points. Three of today's stories are exactly the class of incident it's built to catalog.
The New Way is human-curated — a person picks every story. The summaries are written with AI (Claude) and reviewed before we hit send.