daily notes

Themes

Grok 4.7 advances long-horizon work at unchanged token pricing

At xhigh effort, Grok 4.7 advances long-horizon coding and terminal benchmarks while retaining $2/$6 input/output pricing. Artificial Analysis records stronger agentic knowledge-work results, but its headline comparison uses xhigh for 4.7 and high for 4.6 and measures 81k output tokens per task versus 36k.

Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index to bring SpaceXAI into the top 4 AI labs. Coding Agent Index performance has also improved, overtaking GPT-5.6 Sol Grok 4.7 scores +2 points over Grok 4.6 on the Intelligence Index, with strong performance on agentic knowledge work tasks. We evaluated the new model at xhigh reasoning effort. Congratulations to @SpaceXAI and @ElonMusk on the release! Key takeaways: ➤ Grok 4.7 joins the frontier of agentic knowledge work: Grok 4.7 gains +111 Elo over Grok 4.6 (high) on AA-Briefcase, our private benchmark for long-horizon agentic knowledge work, scoring 1657 Elo and placing it alongside Claude Opus 5 and Claude Fable 5.1 at the frontier. On GDPval-AA, it scores 1695 Elo, +90 ahead of Grok 4.6 (high). ➤ A leap in coding agent performance: Grok 4.7 (xhigh) with Grok Build scores 56 on the Artificial Analysis Coding Agent Index, up +9 points from Grok 4.6 (xhigh). Among models in their native harnesses, Grok 4.7 + Grok Build now ranks 4th, behind only Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5. ➤ Incremental performance changes elsewhere: Outside of agentic knowledge work, Grok 4.7 broadly matches Grok 4.6 (high) on the other Intelligence Index tasks. It improves on Terminal-Bench 4.0 (+4.5 percentage points) and GDP.pdf (+3.0 p.p.), with regressions on AA-LCR (-3.7 p.p.) and AutomationBench-AA (-1.1 p.p.). ➤ High token use across tasks: Grok 4.7's gains come with higher token usage. Grok 4.7 (xhigh) uses approximately 81k output tokens per Intelligence Index task, compared with 36k for Grok 4.6 (high) and 27k for GPT-6 Astra (max) - 125% and 196% more, respectively. Other model details: ➤ Context window of 500k tokens, unchanged from Grok 4.6 ➤ Pricing of $2/$6 per 1M input/output tokens with cache hits discounted to $0.50 per 1M tokens, matching Grok 4.6 ➤ Configurable reasoning effort spans low to xhigh. Our evaluation uses xhigh.

MiMo V2.6 pairs a large index gain with unchanged listed pricing

MiMo-V2.6-Pro reaches 46 on Artificial Analysis’s index, up from 26 for V2.5-Pro, while Xiaomi lists the same $0.435 per million cache-miss input and $0.87 output token prices. The open-weight 1.02T-parameter MoE is a sharp capability-price move, although Xiaomi’s broader benchmark table and application claims remain first-party.

MiMo-V2.6-Pro debuts as the top open weights model on the Artificial Analysis Intelligence Index (46). At $0.13 per Intelligence Index task, it lands on the Intelligence vs. Cost per Task Pareto frontier @Xiaomi has just released MiMo-V2.6-Pro, an open weights model with major advances in intelligence over its predecessor, MiMo-V2.5-Pro (Intelligence Index: 26). Despite the improvement, it retains the same attractive pricing at $0.435 per 1M input tokens (with a 99% cache-hit discount) and $0.87 per 1M output tokens. This makes MiMo-V2.6-Pro one of the most cost-efficient models to deploy. MiMo-V2.6-Pro is an MoE model with 1.02T total parameters and 42B active parameters. Stay tuned for additional analysis of the model. Check out MiMo-V2.6-Pro full benchmarking breakdown here: https://t.co/czBJhQKuWJ

discussion1 selected reply
@tensorquayreply ↗

42B active, but the rest of the experts still need somewhere to live 😅 At an ideal 4 bits per weight, 1.02T parameters is ~510GB just for weights, before scales or KV cache. That's the local-hosting number I'd keep beside the API price. https://t.co/4D83rq8BvU

Must read

Anthropic publishes a full methodology for tracking AI-led R&D, agent oversight, and compute allocation, including a reported 26% share of model-R&D tasks led by Claude in August. The figures are self-reported and not yet cross-lab comparable, but the scope limits and appendix make the proposal legible enough to test rather than merely endorse.

Xiaomi’s long-form release gives the architecture, availability, API pricing, benchmark table, and concrete tool-use case studies behind MiMo-V2.6-Pro. Artificial Analysis supplies an independent capability-and-cost observation, while Xiaomi’s wider application and benchmark claims remain attributable to the builder.

Move video decoding from the CPU to the GPU in vLLMvLLM and NVIDIA Computer Vision Team

This engineering note gives the configuration, scaling experiment, and caveats for moving multi-GPU video-captioning decode work to NVDEC. Its 8×H100 test reports more than double the CPU-decoder throughput, while explicitly flagging the VRAM reservation trade-off that determines whether the approach fits a real deployment.

Signals

US and China agree to dialogue on AI ahead of Trump-Xi meeting https://t.co/kyoNOBFXB9

The Financial Times reports that the US and China agreed to AI dialogue ahead of a Trump-Xi meeting. The available article metadata confirms the report’s framing but no primary bilateral readout was available in this run, so the scope and commitments remain unverified.

Sentiment

Measured capability gains widen deployment choices +0.22

38 source items · 67% editorial confidence

Must read

Credit to Meta's @DinaPowellMcC for bringing Louisiana officials into the @theallinpod discussion to explain how communities are benefiting from data-center investment. Satya Nadella and Brad Smith’s recent posts/discussions of Quincy, Washington is another useful example of the work every hyperscaler needs to do: engage residents early, explain the economics and demonstrate how local people can share in the wealth creation. These are first steps in earning sustained community support. Trying to force projects through or dismissing opposition as propaganda won’t substitute for tangible benefits and credible commitments on costs. A few specifics from Richland Parish: >Teacher compensation: The June sales-tax distribution for full-share certified employees rose from $10,200 to $50,935; support staff payments rose from $3,323 to $17,472. Construction-related receipts flowed through existing taxes dedicated to employee compensation, so these are variable distributions rather than permanent salary increases. >Talent recruitment: Superintendent Sheldon Jones said every teacher interviewed that year was fully certified, a first in his 30-year career. Higher compensation is helping the district compete for educators. >Local businesses: Meta reported more than $1.6B in contracts awarded to Louisiana businesses since breaking ground in December 2024. Local restaurant and service-business owners described opening or expanding to meet the additional demand. >Changing attitudes: Officials described initial uncertainty giving way to visible benefits, with other rural parishes now calling to ask about attracting similar projects. Residents described opportunities to return home or build businesses locally, while the superintendent emphasized giving students reasons to stay. >Access to jobs: Meta announced $5M for community-college scholarships, including data-center-related trade training for eligible local graduates. McCormick also described paid training with guaranteed contractor jobs at Meta sites, addressing the income barrier that prevents many workers from retraining. >A longer-running example: Microsoft describes Quincy’s expanded tax base helping fund schools, a hospital and other public facilities, alongside local technical training. The next test for newer projects is how much benefit persists after construction, and whether developers deliver on their commitments to residents.

Mathew uses a Louisiana case to treat local benefits as a project-risk input: teacher payments were variable distributions, and the enduring test is whether workforce, infrastructure, and community commitments outlast construction. Meta’s first-party release supports the cited local contracts and training commitments, but it cannot establish the long-run outcome by itself.

Sentiment

Buildout economics include local delivery +0.03

72 source items · 48% editorial confidence