Live Feed
IBM Releases Granite 4.2 8B: Dense Reasoning at $0.10 Input per Million Tokens, 12-Language Support
IBM's Granite 4.2 8B brings multi-step reasoning, code generation, and multilingual dialogue to 12 languages at $0.10/$0.15 per million tokens. The 131K-context dense model targets agentic workflows on a single inference host.
OpenAI Tests GPT-6 Astra on 20 Recent Chrome V8 Vulnerabilities
ExploitBench uses 20 high-severity V8 flaws from 13 stable Chrome releases between June and August 2026. Astra beat GPT-5.6 Sol on arbitrary code execution while using fewer output tokens, sharpening the case for phased cybersecurity access.
SpaceXAI Brings Colossus 2 Online at 946 MW With $35.8B in Memphis Compute
Colossus 2 became operational on September 3 with roughly 946 MW of IT power and 1.1 million H100-equivalent compute units. The build uses temporary gas turbines as a bridge to a 1.2 GW plant due in mid-2027, turning SpaceXAI into a merchant compute provider as much as a model lab.
Spotify's Portal Cut Claude Code Token Use by 90% — by Solving the I/O Problem
Spotify Engineering published findings from Portal, an internal proxy that intercepts Claude Code's filesystem reads and caches repeated calls. Token consumption dropped 90% per session. The finding reframes where AI coding agent cost actually comes from.
Gimlet Labs Raises $300M at $3B to Route AI Inference Across Any Chip
Andreessen Horowitz leads a $300M Series B in Gimlet Labs, which claims the first multi-silicon inference cloud. The round values the company at $3B — six months after its $80M Series A. Total funding is now $392M.
OpenAI, Anthropic, and xAI All Went Down Simultaneously on September 3 — No Shared Cause Confirmed
Three frontier AI providers suffered overlapping outages on the same morning. xAI traced it to a Memphis compute center. OpenAI cited a routing error. Anthropic said nothing publicly. AWS, Azure, and Cloudflare all reported no issues.
GitHub HydraFusion Matches Opus 5 Quality on TerminalBench at 67% Lower Estimated Cost
GitHub Copilot's new research preview selects from models across multiple providers at runtime, using three execution patterns to hit frontier quality without always paying frontier prices. On TerminalBench 2.1, it beat Opus 5 by 4.9 points at 67% lower estimated cost.
Claude Formalized Fermat's Last Theorem in 11 Days — 13 Million Lines of Lean, 29,500 Proofs
Working largely autonomously, Claude produced the first end-to-end computer-checked proof of Fermat's Last Theorem. It wrote 13 million lines of Lean code and proved 29,500 intermediate theorems, five times the size of the entire Mathlib library.
16,893 Sessions: Claude Code Skips Web Search, Codex Always Does, Agents Disagree More Than They Agree
A controlled study of 16,893 coding agent sessions reveals sharply different tool-selection behaviors between Claude Code, Codex, and Cursor. Repo context overrides brand preference. Some tools get mentioned in every session and chosen in almost none.
FrontierCode Diamond: Claude Opus 4.8 Merges 13.4% of PRs Under Real Maintainer Standards
Cognition's new benchmark measures whether AI code would survive actual code review, not just pass tests. The best model clears the bar on 1 in 7 tasks. The rest are much worse.
LiveBench September 2026: Fable 5.1 Leads at 83.4, Muse Spark 1.3 Scores 81.6 at $0.22 per Task
September LiveBench rankings confirm Claude Fable 5.1 at the top with 83.4 overall, 0.4 points above Fable 5. Meta's Muse Spark 1.3 enters at 81.6 — matching GPT-5.6 Sol on agentic coding (64.1 vs 56.2) while costing $0.22 per task against Sol's $0.52.
GPT-6 Astra Matches Fable 5 on Coding Agents but Costs 75% More Per Task Than Its Own Predecessor
Artificial Analysis benchmarks GPT-6 Astra across two indices: it ties Fable 5 on the Coding Agent Index at less than half the price, while landing at the same Intelligence Index score as GPT-5.6 Sol (61) at 75% higher per-task cost. The split verdict turns on which workload you are running.
Google Doubles Its AI Chip Rollout Cadence, Targeting Two Generations Per Year
Google's AI infrastructure chief says the company is shifting from a two-year design cycle to releasing two custom chips per year, with ambitions to move even faster. The change puts Google's internal silicon on a cadence that rivals NVIDIA's accelerated Blackwell roadmap.
GPT-6 Astra Scores 99.9% on ARC-AGI-3, Leaving Every Competitor Below 31%
ARC Prize results show GPT-6 Astra effectively solved the newest iteration of the benchmark, used fewer actions than the median human on 96% of levels, and widened the gap between OpenAI and the rest of the frontier to a degree not seen on any prior AGI evaluation.
MBZUAI Releases K2 Horizon: Six Fully Open Models From 0.9B to 375B, First Complete Agentic Fleet
Abu Dhabi's Institute of Foundation Models at MBZUAI launches K2 Horizon, six models spanning wearable edge to enterprise deployment. Every model ships with weights, training code, intermediate checkpoints, and data recipes — the most comprehensive open release to date and the first fleet to expose the full agentic training lifecycle.