SpaceX agrees to acquire Cursor for $60B, Anthropic tests removing Claude Code from its Pro plan, and TypeScript 7.0 Beta arrives rewritten in Go

Simon Willison

Anthropic Tests Removing Claude Code from Its $20/Month Pro Plan

Anthropic silently updated its pricing page to move Claude Code from the $20/month Pro plan to $100–$200/month Max tiers, then reversed course hours later after widespread community backlash. The company's sole official response came via an employee tweet claiming it was a test affecting "~2% of new prosumer signups," but the change was visible to all visitors and archived by the Internet Archive. While the landing page was reverted, Anthropic confirmed the underlying pricing experiment continues invisibly for a subset of users. The incident handed a marketing opportunity to OpenAI, whose Codex team publicly noted that their product remains available on free and $20 plans.

Microsoft DevBlogs

TypeScript 7.0 Beta: ~10x Faster Compilation via Go Rewrite

TypeScript 7.0 Beta delivers a complete rewrite in Go, achieving approximately 10x faster compilation than TypeScript 6.0, with parallel type-checking workers, simultaneous project reference builds, and a native VS Code extension. The release eliminates legacy targets — ES5 output, classic module resolution, AMD/UMD/SystemJS formats — and changes the `types` field default to an empty array, requiring tsconfig.json updates for many existing projects. Fortune 500 companies have already validated the beta across million-line codebases, with a stable release expected within two months.

Simon Willison

OpenAI Launches GPT-Image-2, Claiming Parity Leap from GPT-3 to GPT-5

OpenAI released gpt-image-2 (ChatGPT Images 2.0), with Sam Altman claiming the quality improvement is comparable to jumping from GPT-3 to GPT-5. In independent testing the model outperformed Google's Nano Banana 2, particularly at resolutions up to 3840×2160, and now leads Arena's text-to-image leaderboard with a 242 Elo-point margin over the next competitor. Downstream integrations launched immediately in Figma, Canva, Firefly, and fal, and the model is accessible via API and ChatGPT. One key limitation surfaces: models still cannot reliably verify their own generated outputs.

Mozilla

Claude AI Helps Mozilla Find 271 Vulnerabilities Patched in Firefox 150

Mozilla applied an early version of Anthropic's Claude Mythos model to Firefox's codebase and discovered 271 vulnerabilities, all patched in Firefox 150 — matching elite human security researchers across all vulnerability categories and complexity levels. Firefox CTO Bobby Holley described it as a turning point: "Defenders finally have a chance to win, decisively," framing AI-assisted security as fundamentally shifting the historical advantage away from attackers. Mozilla adds a critical caveat: codebases that grow beyond human comprehension through AI generation risk accelerating vulnerability creation alongside discovery, making readable, maintainable code a prerequisite for effective AI security scanning.

Framework

Framework Laptop 13 Pro: Intel Core Ultra 3, LPCAMM2, and 20-Hour Battery

Framework announced the Laptop 13 Pro featuring Intel Core Ultra Series 3 processors, LPCAMM2 LPDDR5X memory up to 64GB, a 13.5-inch 2880×1920 touchscreen with 30–120Hz variable refresh and 700 nits brightness, and approximately 20 hours of battery life. The design continues Framework's repairability-first philosophy with physical webcam and microphone kill switches, a customizable Expansion Card port system, and open-sourced component designs. Ubuntu certification and Linux-first support make it a strong pick for developers who prioritize hardware longevity and user control over planned obsolescence.

Krebs on Security

Scattered Spider's 'Tylerb' Pleads Guilty to SIM-Swap Crypto Theft

Tyler Robert Buchanan, 24, a Scottish national operating as "Tylerb," pleaded guilty to wire fraud conspiracy and aggravated identity theft for orchestrating tens of thousands of SMS phishing attacks against Twilio, LastPass, DoorDash, and Mailchimp in 2022. Stolen data from those breaches was used to SIM-swap individual cryptocurrency investors, with Buchanan's admitted haul totaling at least $8 million. He faces up to 22 years in prison with sentencing scheduled for August 21, 2026, making him the second Scattered Spider member to plead guilty following co-conspirator Noah Urban's 10-year sentence last year.

Cloudflare Blog

Cloudflare: The 'Bots vs. Humans' Security Model Is Becoming Obsolete

Cloudflare argues the "bots vs. humans" web security distinction is obsolete as AI agents, accessibility tools, and corporate proxies blur the line — and that what actually matters is intent and behavior, not biological classification. The company is advancing Privacy Pass (RFC 9576/9578) alongside newer primitives like Anonymous Rate-Limit Credentials (ARC) and Anonymous Credit Tokens (ACT) that let clients prove trustworthiness without revealing persistent identity. Without privacy-preserving alternatives, Cloudflare warns, the web risks fragmenting into walled gardens requiring device attestation or account authentication, concentrating power among a handful of major platforms.

dpc.pw

"I Don't Want Your PRs Anymore": Rethinking Open-Source Contribution in the LLM Era

Open-source maintainer Dawid Ciężarkiewicz argues the traditional PR model no longer makes sense: LLMs eliminate the productivity advantage of accepting external code, while the overhead — security review, style negotiation, async coordination, CI cycles — remains unchanged. Rather than merging external PRs, the author prefers contributors to file detailed bug reports, propose designs, or share illustrative prototypes with prompts, while the maintainer uses LLMs to generate the actual implementation. The post crystallizes a broader reassessment of open-source collaboration economics now that code generation is cheap and review is the bottleneck.

Martin Fowler

Thoughtworks Technology Radar Vol. 34: AI Drives Return to Engineering Fundamentals

Thoughtworks published Volume 34 of its Technology Radar (118 blips), with AI dominating the content but simultaneously driving a revival of foundational practices — pair programming, zero-trust architecture, mutation testing, and DORA metrics — as a "necessary counterweight to the speed at which AI tools can generate complexity." The radar introduces "harness engineering" to describe the governance infrastructure required to safely deploy permission-hungry AI agents, which need broad access to private data and systems while remaining vulnerable to prompt injection. A secondary observation: agentic tools are pulling developers back to the terminal as a primary interface after years of abstraction.