Grok 4.5: SpaceXAI's coding-and-agents flagship, at a price that undercuts everyone
SpaceXAI released Grok 4.5 on July 16 — its strongest model, aimed squarely at coding, agentic tasks, and knowledge work, and trained alongside Cursor on tens of thousands of NVIDIA GB300 GPUs. Its published numbers put it in the frontier pack rather than ahead of it (Terminal Bench 2.1: 83.3 vs Claude Fable 5's 84.3; SWE Bench Pro: 64.7 vs Fable 5's 80.4; it does lead the SWE Marathon long-horizon benchmark at 29.0), but the pitch is economics: $2 per million input tokens and $6 output — a third of Claude Sonnet-class pricing — served at 80 tokens/s, with roughly 4× fewer output tokens per SWE Bench Pro task than Opus 4.8. It's the default model in Grok Build, free there and in Cursor for a limited time, and available via API. Independent measurement broadly agrees: Artificial Analysis places it at 54 on its Intelligence Index, seventh overall, at among the lowest cost-per-task of any frontier-tier model.
Sources & depth
Wire-level analysis showed Grok Build CLI uploading entire repos; xAI killed the pipeline after publication
A security researcher publishing as cereblab wire-captured what xAI's Grok Build coding CLI actually transmits and found it uploaded the entire tracked repository — every file plus full git history, packaged as a git bundle — to a Google Cloud Storage bucket (grok-code-session-traces), independent of what the agent read: on a 12 GB test repo, 5.1 GB left the machine for a task that needed 192 KB. Files the agent was explicitly told not to open were recoverable from the captured upload, and turning off the "Improve the model" setting did not stop transmission. After the analysis and a public reproduction harness circulated, xAI disabled the upload server-side, added a /privacy retention control, and Elon Musk publicly committed to deleting previously uploaded data — a commitment not yet independently audited. The practical lesson for anyone running coding agents: read the wire, not the marketing, and gitignore anything you can't afford to ship.
Sources & depth
Apple targets dozens of OpenAI employees with legal letters
Apple has sent legal preservation letters to about 40 former employees now working at OpenAI — roughly a tenth of the ~400 ex-Apple staff there — directing them to preserve documents and communications and to meet with Apple's lawyers. The letters escalate the trade-secrets lawsuit Apple filed last week against OpenAI, its device subsidiary io Products, and two named ex-Apple engineers: hardware chief Tang Tan (a 24-year Apple veteran) and Chang Liu. Apple alleges OpenAI's hardware program is built on misappropriated designs and calls the evidence filed so far the "tip of the iceberg"; OpenAI says it has seen no evidence the complaint has merit and that its device timeline is unchanged. Preservation letters make deleting the requested material a legal offense, so the dispute now hangs directly over OpenAI's first AI device — and any discovery that follows.
Sources & depth
NVIDIA pitches Vera Rubin as the post-training platform — and joins the intelligence-per-dollar metric war
NVIDIA's newsroom argues that continuous post-training — reinforcement-learning loops that keep adapting agentic models to new tools and environments — is becoming the dominant AI workload, and positions the Vera Rubin platform as doing that work with roughly a quarter of the GPUs the Blackwell generation needs. Evidence offered: Nemotron 3 Ultra (550B parameters) reaching 71.7% on SWE-bench Verified after RL post-training, and Prime Intellect measuring ~30% more throughput per CPU from Vera CPUs versus x86 in RL sandbox workloads; Perplexity and Together AI are also cited as running RL post-training at scale. No new product here — this is a positioning post, and the third metric NVIDIA has led with this month after performance-per-watt. Its headline phrase, intelligence per dollar, is the same one OpenAI's CFO published the same day.
Sources & depth
Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers
NVIDIA and Hugging Face shipped an integration that lets NeMo Automodel fine-tune Diffusers models straight from the Hub — point it at a model ID and train, no checkpoint conversion or model rewrite. It launches with recipes for FLUX.1/FLUX.2 text-to-image (12–32B), Wan 2.1/2.2 text-to-video (1.3–27B), HunyuanVideo (13B), and Qwen-Image (20B), with both full fine-tuning and LoRA. Parallelism — FSDP2, tensor, expert, context, pipeline — is YAML configuration rather than code, scaling the same recipe from a single GPU to multi-node clusters; on 8× H100, FLUX.1-dev sustains ~35.5 images/s for full fine-tuning and ~53.7 with LoRA. Apache 2.0 throughout. This moves large video-model fine-tuning from a bespoke engineering project to a config exercise.
Sources & depth
A scorecard for the AI age
OpenAI CFO Sarah Friar argues businesses should measure AI by work accomplished, not tokens consumed — a scorecard she names useful intelligence per dollar. It asks four questions: is AI completing work that matters; what does each successful task cost (counting retries, human review, and rework, not just tokens); can people depend on the result; and does each AI dollar produce more value as usage grows. The pointed claim: the cheapest tokens rarely produce the cheapest outcomes, because a frontier model that succeeds in one pass can undercut a budget model on cost per successful task. Landing the same day as NVIDIA's identically-phrased Vera Rubin pitch, it makes the industry's ROI-metric land-grab explicit — vendors are now competing to define the denominator.
Sources & depth