Grok 4.7 outscores Sol 📊, Anthropic's safety rhetoric pays its bills 🧷, Your prompts belong in git 📍

SpaceXAI released Grok 4.7 on 21 September in Cursor and Grok Build, priced at $2 per million input tokens and $6 per million output, the same as Grok 4.6. At its xHigh setting it scores 46.3% on CursorBench 4.0, against 41.7% for GPT-5.6 Sol Max and 40.4% for Grok 4.6 High. Fable 5.1 Max still leads that benchmark at 51.8%, but lists at $10 and $50 per million.

Xiaomi released and open-sourced the MiMo-V2.6 series on 22 September, led by MiMo-V2.6-Pro, which scores 46.32 on the Artificial Analysis Intelligence Index and passes Kimi K3 to become the strongest open-source model to date. API pricing is unchanged from V2.5, at $0.435 per million input tokens and $0.87 per million output for Pro. The livestreamed RL run behind it cost about $2.62 million and lifted DeepSWE v1.1 from 58.4 to 72.57.

Harvey, the $15.6 billion legal startup, saw gross margins fall from about 50% at the start of the year to -50% by June, according to a person familiar with the matter. Customer usage spiked after a March update to its AI agents. Harvey has now joined a group of software companies adopting open-weight models, including increasingly capable alternatives from China, which are typically cheaper than US proprietary technology.

Amazon blocked Meta's Muse agent from shopping on Amazon.com as of Sunday night, showing users a popup citing its Conditions of Use. Amazon says Muse appears to capture and store customer credentials, scrape account data, and browse without identifying itself. Amazon lost its Perplexity injunction on 4 August, when the Ninth Circuit ruled the user rather than the AI company accesses its computers under federal anti-hacking law, leaving contract claims as its remaining route.

OpenAI began training a new internal model on 28 August, and says it has now resolved the Navier-Stokes Millennium Prize problem and more than 100 long-standing open problems. A recent open letter from mathematicians criticises solving open problems as a benchmark for new AI systems. OpenAI has set up an advisory group hosted at the Institute for Advanced Study, whose unpaid members will not advise it on the pace of its internal progress.

Treasury Secretary Scott Bessent said on Sunday that the United States has proposed a new notification mechanism with China for AI incidents that could affect national security. He described it after talks with Chinese Vice Premier He Lifeng in New York, ahead of a Trump and Xi meeting at the White House on Thursday. Chinese state media reported only that both sides discussed issues relating to AI, without specifics.

Agility Robotics is going public through a merger with Churchill Capital XI at a $2.5 billion pre-money valuation. Its S-4 shows $1.8 million of 2025 revenue against $4.5 million of cost of goods sold and operating losses over $140 million, with about 7 to 8 humanoids sold and 5 going to shareholder Amazon. From here the plan is to rent Digit v5 as labour at an illustrative $8,500 a month plus a $25,000 deployment fee.

Linear's CTO assigned an engineer a ticket titled CI costs are high, and the team rebuilt the pipeline rather than buying more machines. They moved workloads to third-party runners, switched to tsgo, the native TypeScript compiler, which cut the weekly median typecheck by 73%, and split API tests into 8 shards. Despite the suite nearly quadrupling this year, pull request wait time fell from more than 6 minutes to just over 5.

Months of prompting frontier coding agents, from Opus 4.5 through GPT-6 Astra, to rewrite machine learning algorithms in Rust showed that vague instructions stall: told to make code 'as fast as it can be', Opus 4.5 tweaked a few hyperparameters and stopped. Demanding every benchmark beat a measured baseline by at least 1.2x made the agents keep iterating until they hit it. The technique compounded into 2x to 20x speedups, with one UMAP implementation ending 7.5 to 32 times faster.

Daniel Lemire rebuilt every commit of 6 widely used libraries, including simdjson, ada and simdutf, and benchmarked them on one Xeon against an August 2024 baseline. Throughput had been flat for years, then six jumped within weeks: ada's URL parsing went from 0.54 to 1.28 GB/s, and simdutf's ASCII validation from 83 to 160 GB/s. He says he cannot know how much AI was involved, only that trying new ideas became cheap.

Carson Gross argues Markdown is becoming source code rather than documentation, and should be checked into /src beside the code it produces. His proposed convention keeps README, TODO, overview, feature, data-model and API files that developers write directly, with code and tests derived from them instead of from throwaway prompt sessions. The point is locality: the reasoning behind a change sits in the same directory as the change, not in a wiki or a ticket.

Shopify is rewriting its largest app, more than 300 screens, from React Native to native Swift and Kotlin, having already rebuilt and shipped the Shop app in 12 weeks. Its internal tool Helix splits each screen into small ordered checkpoints, and every one must clear CLI behaviour tests, a Gemini visual comparison against the React Native reference, and two independent adversarial code reviews. Engineers can leave it running overnight, and the gates do not relax when nobody is watching.

Stratechery's Ben Thompson argues Anthropic's push to pace the frontier on safety grounds also happens to resolve the capability, product, pricing and capital problems facing frontier labs. He cites Microsoft confirming Copilot now rotates through multiple models beyond Claude, and Anthropic dropping its data-retention provision once Fable usage stayed low, as evidence that model capability alone no longer locks customers in. He warns that slowing progress favours attackers, since an attack only needs to work once.

In testimony prepared for Congress, Nathan Lambert says Chinese open models GLM-5.3 and Kimi K3 score 45 and 44 on the Artificial Analysis Index, against 26 for the leading American open models. He puts Chinese open weights 2 to 5 months behind the closed American frontier, versus 6 to 9 months for US open models, and estimates distillation explains only 1 to 2 months of that gap. He wants Congress funding US open models rather than restricting Chinese ones.

Tim Dettmers' academic lab says its CliffCompaction technique cuts AI agent costs by about 50%, with one partner measuring a 45% reduction in its total AI budget after deploying it. The same open-source stack runs DeepSeek V4.1, a 550-billion-parameter model, on a single MacBook with 128GB of memory. Dettmers argues problems that are cheap to attack and valuable to solve are now uncontested ground for small university labs.

Spyglass argues Google should already have matched Meta's Muse personal agent, given it runs the largest email service, calendar, search engine and smartphone platform. Instead its answer, Gemini Spark, is gated behind paid tiers and available in the US only, while overlapping products called CC, Daily Brief and a planned Gmail AI Inbox target the same job. The author, a former partner at Google's venture fund GV, has Spark access and says he does not know what to do with it.

Camille Fournier argues management is not disappearing, despite Jack Dorsey and Coinbase's Brian Armstrong calling for its end, pointing to Meta rolling back ratios of 40:1 and 50:1 around 13 September after what it recognised as bad outcomes. She warns that letting AI draft a performance review hands over the judgment that earns a team's trust. She notes 1 in 3 managers are now considering a return to individual-contributor roles.

White Circle released Halo 1.0.0 on 20 August 2026, an open-source framework for pre-training, fine-tuning and reinforcement learning of large language and multimodal models. On 8 B300 GPUs it delivers up to 2.8 times the training throughput of stock TRL, with larger margins over the other frameworks it benchmarked. Checkpoints stay standard Hugging Face files that load with from_pretrained, and adding a new model family typically takes under 140 lines of integration code.