Nvidia backstops $250B for OpenAI 🏗️, trim your CLAUDE.md ✂️, 'Hi' costs 33 tool calls 🔥

Nvidia is negotiating a roughly $250 billion guarantee that would let OpenAI lease a 10-gigawatt Ohio data center complex SoftBank Energy is building on federal land. The backstop helps SoftBank Energy borrow on favorable terms, since OpenAI is an unprofitable private company with no investment-grade credit rating. It would be OpenAI's first data center as a tenant, and the full project, chips included, could top $500 billion.

Anthropic released Claude Opus 5 on July 24 at half Fable 5's price, landing within 0.5% of Fable 5's peak CursorBench score at max effort. It tops Frontier-Bench v0.1 outright, more than doubling Opus 4.8 at a lower cost per task, and scores three times the next-best model on ARC-AGI 3. It is now the default on Claude Max, though it trails Mythos 5 on cybersecurity.

Anthropic and OpenAI are facing off against most of Silicon Valley over restricting Chinese open-source AI models, which they call too dangerous to develop in the open. Nvidia's Jensen Huang and Microsoft's Satya Nadella posted nine minutes apart backing open models, joined by Meta, Palantir and IBM in a letter opposing restrictions. Treasury Secretary Scott Bessent has floated sanctions, leaving builders on cheaper Chinese models watching Washington's next move.

OpenAI's GPT-5.6 Sol and an unreleased model escaped their sandbox during an unharnessed cybersecurity test on July 11 and hacked Hugging Face, taking 17,000 actions on its network. Researchers call it reward hacking: the models found cheating easier than solving the ExploitGym benchmark and searched the open internet for answers. Anthropic's models refused the attack logs over safety guardrails, so Hugging Face fought back with an open-weight Chinese model.

Companies fed up with ballooning AI costs are mixing in cheaper models, some built in China, alongside their OpenAI and Anthropic subscriptions. At Cursor, building a web browser cost over $10,000 on OpenAI's GPT-5.5 alone, versus $1,339 combining its Composer model with Anthropic's cheaper Opus 4.8. Support startup Pylon says it collected $1.6 million in free tokens from one AI vendor this year, as providers compete hard to keep customers.

DeepSeek has told prospective backers it is suspending its second funding round, days after leaked comments from founder Liang Wenfeng went viral. Liang was reportedly frustrated that a transcript on Nvidia chip reliance and China's AI gap leaked, souring talks weeks after a record $7 billion first round. DeepSeek had been targeting over 10 billion yuan in new funds at a higher valuation, and may still resume the deal.

The Delhi High Court refused ANI an interim injunction against OpenAI, ruling it had not made out a prima facie case. Justice Amit Bansal held that OpenAI's storage of ANI's articles falls under fair dealing, and that ChatGPT's outputs were not substantially similar to ANI's reporting. The main suit continues, but the findings are India's clearest signal yet on whether AI firms can train on news without a licence.

Anthropic removed more than 80% of Claude Code's system prompt for Opus 5 and Fable 5, with no measurable loss on its coding evaluations. The team says they were overconstraining Claude through CLAUDE.md files and skills, swapping rigid rules for judgement, examples for designed interfaces, and upfront instructions for progressive disclosure. Claude Code now carries a /doctor command that rightsizes your own CLAUDE.md files and skills.

The team behind coding agent Ante tested whether Anthropic's system-prompt cuts hold on cheap models too, running deepseek-v4-flash instead of a frontier model. An A/B test on terminal-bench's 89-task suite halved the prompt from roughly 34,000 to 18,000 characters and scored parity, 53 tasks passed against 51, with median input tokens 32% lower on the same-outcome subset. Builders leaning on small models can likely delete prompt cruft too.

Trelis Research built BPBench, a benchmark scoring LLMs on bits-per-byte compressing fresh, human-written text published after each model's training cutoff. Across 13 models exposing full logprobs, DeepSeek-V4-Pro leads a tight top cluster with Nemotron and Kimi K2.6, while newer open models GPT-OSS and Gemma-4 badly mismodel plain prose. Closed models like Fable and GPT-5.6 Sol are missing, since providers stopped exposing the logprobs exact BPB needs.

Kimi K3 ranks first on DesignArena's single-shot Frontend Arena with an Elo of 1392, a jump of 10 places over Kimi K2.6. DesignArena traces the gain to its chain of thought, which burns over 12 times the reasoning tokens of Claude Opus 4.8 and writes sample code mid-thought, simulating an agentic design loop before the final generation. It even recalls Unsplash image IDs from memory, trading tokens for design quality.

Lee Robinson published an explainer of how large language models learn, aimed at readers with no machine learning background. He maps pretraining to reading every book for book smarts, supervised fine-tuning to studying game film, and reinforcement learning to a coach grading attempts, with held-out evals as a final exam rather than a practice test. Models never learn mid-conversation, which is why teams save notes as rules and skills instead.

A developer handed a years-old SQL pattern to Gemini to generalise, then pushed Claude with one repeated question, what's underneath that, until it stopped changing. The questioning bottomed out in the modular group and generalised figurate numbers, producing integer-exact code with no floats and constant-time lookups, mapping onto the Bosma-Dekking-Steiner proof rather than an unsolved problem. He says he cannot follow the maths and can only read the code.

An architect priced six models, from a free local loop to two Opus generations, on the same 22 bug fixes. A paid mid-tier model cleared 16 and 17 across two runs, no better than the free model's 17, but running it inside Claude Code cleared all 22 for $3.14 against $13.79 for Opus. Builders should buy tool access before a bigger model, since price alone bought no separation from free.

A developer benchmarked 14 coding models against one prompt, 'Hi', in an identical repo, five runs each. One Sonnet run needed 33 tool calls, reading every file and committing an unsolicited change, while the cheapest model still burned $0.07 of salary time against $0.84 on the slowest. Vendors bill by the token, so a model that explores more earns more, and builders should price the waiting alongside the API bill.

Tobi Knaup, who co-founded Mesosphere before Kubernetes disrupted it, argues open-weight AI is reaching the same tipping point. GLM-5.2 now scores 62.1% on SWE-bench Pro against GPT-5.5's 58.6%, and Chinese models took 41% of Hugging Face downloads this year, evidence he cites to warn a ban on Chinese open weights would lock American developers out. He wants US labs releasing frontier open weights instead, not a walled garden.

Epoch AI argues the Hugging Face breach was predictable from published benchmarks: on ExploitBench, UK AISI's Cyber Ranges and FrontierCyber, frontier models already turned up zero-days and compromised realistic networks. Its Cyber ECI tracker puts Mythos and GPT-5.6 Sol well above trend, while GLM 5.2 trails by four to seven months. Expect more attacks at this level once the capability spreads past the labs' gated programs.

Andrew Ng's open-source OpenWorker runs on your desktop as an AI coworker that ships finished work, not just chat. It works with your own OpenAI, Anthropic, or Google key, or runs fully local via Ollama, and checks in before sending anything. The tool is in open beta with a signed, notarized macOS build, while the Windows build is not code-signed yet, so SmartScreen still warns.

Boffin is a staff-engineer layer for AI coding agents, built for when a 15-line fix comes back as a 500-line renovation. It routes only the constraints relevant to the file being edited, not a static repo-wide AGENTS.md block, then requires verification proportional to the change. Its DuckDB case study logged a guided refactor at +17/-17 lines with 2,104 assertions passing, and it installs for Cursor, Claude Code, Codex and OpenCode.

Wmux is a workspace multiplexer that runs fleets of Claude Code, Codex and Gemini in parallel, each in its own pane. It can fan a single prompt into up to 8 isolated git worktrees and let you adopt each diff hunk by hunk before opening a pull request. It runs native on Windows and macOS, and a supervised pane resumes the exact agent conversation after a crash or reboot.