Another Daily AI Newsletter - August 13
Top Story: Grok 4.6 puts SpaceXAI back at the frontier
SpaceXAI released Grok 4.6, less than three years after the first Grok began as a two-month training sprint. The new model is no longer merely chasing OpenAI and Anthropic. It belongs in the same frontier tier.
Artificial Analysis gives Grok 4.6 a score of 61 on its Intelligence Index, level with GPT-5.6 Sol and one point behind Fable 5. Axios called the release SpaceXAI’s return to the AI elite.
The economics strengthen the result. Grok 4.6 currently leads CursorBench at 70.8%, narrowly ahead of Fable 5 at 70.5%, while costing $2.81 per task versus $17.32. Cursor cautions that small score differences may not be statistically meaningful. Cursor also collaborated on Grok 4.6 and supplied anonymized workflow data for supplemental training, making the benchmark highly relevant to coding work but less neutral as universal proof.
The results vary by task. SpaceXAI’s 36-page model card shows Grok leading several knowledge-work and inference tests while trailing rivals on DeepSWE, TerminalBench, and other agent evaluations. One factuality test also produced a higher error rate than Grok 4.5. Grok has reached the frontier without sweeping it.
The development speed may be the bigger story. Grok 4.6 arrived one month after 4.5, and SpaceXAI trained it on work from its own model-development process. An earlier checkpoint reportedly explored 297 inference optimizations in five hours, opened seven pull requests, and produced three changes now serving production traffic. SpaceXAI reports those changes improved prefill throughput by 3.1% and decoding by 1.5%.
Interesting Perspectives
A real implementation exposed the cost-and-speed tradeoff. DHH had Grok 4.6 repeat a Rust implementation from Fable 5’s existing plan. Grok finished in 1 hour 24 minutes for roughly $55; the original Fable run took just over 45 minutes and cost about $550. DHH emphasized that Grok reused Fable’s plan, so this was not a clean planning comparison.
Verification matters more than elaborate prompting. After using the model for several weeks, Cursor’s Eric Zakariasson found that short prompts worked well when they clearly defined completion and required the agent to test its work. Grok handled websites and visible interfaces particularly well because it could inspect the result. Video, 3D, and physics still needed more human review.
The model card includes behavioral regressions. Nathan Calvin highlighted that MASK-Rectified dishonesty rose from 0.67% on Grok 4.5 to 3.8% on 4.6, while compliance with unsafe self-harm requests rose from 0.5% to 3.7%. SpaceXAI’s card reports stronger broad refusal and jailbreak results elsewhere, leaving a mixed safety picture alongside the capability gains.
Autonomous agents are coordinating real attacks
Dream Research Labs recovered the working directory from a four-day campaign that ran as many as eight agents in parallel. The system reportedly cracked 85 employee accounts, entered 84 internal systems, took more than 2,500 personnel records, installed persistent access, and widened its scanning to government suppliers, energy companies, and a nuclear-safety agency.
The Financial Times identified Taiwan as the target and reported simultaneous activity against government and critical-infrastructure systems. The Register separately highlighted the nuclear-safety target. The operator still had to launch the campaign, but the agents handled reconnaissance, credential attacks, verification, and adaptation at a scale that previously required a larger human team.
Open models are becoming a policy fight
The White House may reverse a week-old exclusion. WIRED reports that frontier-capable open models may be added to the administration’s voluntary prerelease safety-testing framework. The framework is not public, and the change has not been independently confirmed. Axios reported last week that open models were excluded.
Three AI pioneers argued for keeping models open. Geoffrey Hinton, Fei-Fei Li, and Andrew Ng disagreed on specific safeguards but said open research and access remain important for competition, scrutiny, and scientific progress.
Alibaba released its largest open-weight model. Qwen3.8-Max has 2.4 trillion total parameters, activates 95 billion per token, and supports context windows up to one million tokens. NVIDIA published a serving recipe for GB300 NVL72 systems.
Open weights are also a data-residency tool. OneAdvanced deployed more than 50 agents inside a UK-sovereign AWS architecture. It self-hosted Llama 4 Maverick and Llama Guard 4 because its preferred managed versions were not available in the required UK region.
AI is moving into work people already do
DeepMind brought sign-language dictation to the Pixel 11. Its sign-language-to-text model lets people sign into Gboard and Live Transcribe anywhere they would normally type. The first release translates American Sign Language into English, with additional devices and languages planned.
Claude browser sessions now follow the user. Claude Cowork for Chrome saves browser tasks to conversation history and lets them continue in Claude’s desktop, web, and mobile apps. Skills and connectors are available inside the browser session.
Sierra wants agents to follow every sales lead. Its long-running Horizon agents can stay with a prospective customer for days, weeks, or months instead of waiting for another inbound message.
Investors are paying for AI output and its verification
Lovable raised $400 million at a $13.3 billion valuation. The company says it reached $500 million in annualized revenue in June, and its valuation has roughly doubled since December.
Blacksmith raised $45 million at a $550 million valuation. Its valuation was $60 million less than a year ago. Faster code generation is increasing demand for the testing and validation layer around it.
Cognition is reportedly discussing a $40 billion valuation. The talks come three months after the Devin maker raised at $26 billion and are reportedly tied to reaching a $1 billion annualized revenue run rate.
Thrive Holdings raised $2 billion at a $12 billion valuation. The OpenAI-backed company buys traditional service businesses and rebuilds their workflows around AI, beginning with accounting and IT.
Quick Hits
BMG and Suno formed a global music partnership. Participating artists and songwriters can opt into new AI music products while BMG licenses its catalog under the agreement.
Amazon will use Twitch streams for AI training by default. Creators must opt out if they do not want their audio and video included.
MAI-Thinking-1 is available in Microsoft Foundry. It is Microsoft’s first internally developed reasoning model for coding, complex reasoning, and enterprise deployment.
DeepSeek released V4-Pro 0813 through its API. OpenRouter carries the model, but DeepSeek had not published an announcement or confirmed open weights when Simon Willison checked.
SpaceX’s Bastrop Gigasat factory is taking shape. The terrestrial factory is intended to manufacture AI1 satellites for orbital computing beginning as soon as late 2027; the footage does not show a data center being built in orbit.
The mass job losses predicted from AI have not appeared clearly in labor data. The Guardian examines the gap between prominent forecasts and the employment numbers available so far.
One Thing Explained: Why AI benchmark rankings can flip
A reasoning budget is the maximum amount of output a model may generate while solving a problem. It acts like a time limit on a test: a model that excels at quick answers may lose its advantage when every competitor receives more room to reason.
A new 56,476-run study tested four models across three reasoning benchmarks with budgets ranging from 64 to 4,096 tokens. Model rankings reversed on every benchmark as the budget changed. More tokens did not always help: accuracy declined on 3% to 19% of questions, depending on the model and test.
This matters because a benchmark score is partly a measurement of the evaluation setup. Reasoning effort, token limits, tool access, prompts, and agent harnesses can all change the order of the leaderboard.
When comparing models, look beyond the headline score. Check the reasoning budget, total tokens, cost per completed task, tool configuration, and whether the difference is statistically meaningful.
Tools to Try
If you want to run a model inside a webpage, try Needle 2 from Cactus Compute. The small model downloads into the browser and runs locally; its weights are available on Hugging Face.
If you build voice agents, try LiveKit’s Expressive mode. It adjusts the agent’s emotional tone to better fit what a caller is saying.
If your Notion inbox keeps growing, try context-aware triage. Notion AI can flag items that need attention and help clear lower-priority messages.
For Builders
Agent Plugins 1.0 now works across VS Code, Copilot CLI, and the Copilot app. The open standard packages skills and MCP servers into one installable plugin instead of requiring a separate integration for every agent client.
Vercel Sandbox added managed images and preinstalled coding agents. Ubuntu is now the default base, and builders can customize open-source images for repeatable agent environments.
Databricks compared three ways to classify against 100,000 labels. The evaluation covers vector search, reranking, and agentic approaches for very large taxonomies.
AWS published a tiered KV-cache architecture for large-model inference. It moves reusable attention state beyond GPU memory to reduce repeated prompt computation without requiring oversized instances for every workload.
Research to Read
Models may know more facts than they can reliably retrieve. Google’s knowledge-profiling study argues that recall, rather than missing stored knowledge, is the main factuality bottleneck in the frontier models it tested.
A locally grounded medical system beat broader models on one benchmark. VITA retrieves Indian treatment guidelines, antimicrobial-resistance data, formularies, and resource constraints instead of relying only on general model knowledge.
A stronger model can improve a weaker one without changing its weights. AI4AI at Test-Time has a builder model create an inference harness that helps a smaller target model solve tasks more reliably.
Scientific diagrams are getting their own multimodal benchmark. Diagram-MMU contains 3,700 diagrams and 18,300 human-validated questions spanning diagram parsing, understanding, and generation tasks.