Latest Artificial Intelligence News & Technical Dispatches

Patch Plugin4Shell zero-click RCE across Claude Code, Codex, Copilot and Gemini CLI with version pins, plugin audits and sandbox escapes blocked in tests.

Evaluate StepFun Step 5 Preview with 600B sparse MoE and 27B active weights at $1 per million input, plus a cache-discount migration check in staging tests.

Benchmark coding agent reasoning tiers from none to xhigh with pass rates, token bills and latency, proving medium effort wins 73% of tasks in tests.

Track Union Alpha from OpenRouter stealth to unbiased Pareto 26.9 with Astra-level scores, and gate unproven models before production in staging tests.

Test Grok Voice Transcribe 2.0 with short-phrase WER down to 6.8% across 19 languages at unchanged batch pricing, plus a swap harness in staging tests.

Deploy PrismML Ternary Bonsai 2 with Qwen3.8 27B at 5.9GB and 1.71 bits per weight, keeping 98.2% benchmarks with a local rollout check in tests.

Cover the Vals AI $40M a16z round for confidential professional benchmarks with contamination math, plus a held-out eval harness you can run in staging.

Compare Voyage, OpenAI and open-weight embeddings on BEIR, code and finance benchmarks with per-million bills, proving domain models win by 8 points in tests.

Run staged context compaction with tool-result offloading and pinned safety constraints, cutting session tokens 74% with zero violations in staging.

Benchmark the Sep 2 launches head to head on coding, long context and cost per task, showing 75.4 DeepSWE and 98.5 MRCR decide the winner in tests.

Trace agent LLM calls in a background thread with span trees and per-model cost tables, capturing 12k spans per min with zero added latency in tests.

Route open-weight inference across DeepInfra and Together AI with live price checks and cache math, cutting monthly model bills 41% in staging tests.

Compare Claude Opus 5 vs GPT-5.1 Codex on SWE-bench, price per task, and throughput. Codex saves $18.75 per million with 50 tok/s speed. Full math.

Test GPT OSS 20b at $0.02 per million against DeepSeek V3 on coding and MMLU Pro. Open-weight routing cuts task cost by 92%. Full benchmark inside.

See how the MCP Registry tracks 26479 servers at 98.8% alive with 15-minute health checks. Learn what separates live servers from dead ones. Full data.