Live Feed
Hurston-1: First Open TTS for Southern American and Afro-English Accents, Built for the Phone Line
Destined AI releases Hurston-1, an Apache 2.0 TTS model trained on 190 hours of consented Southern American, Nigerian, West African, East African, and Southern African English speech. Its NB-TTS mode outputs 8 kHz G.711 audio ready for Twilio — the format phone callers actually hear.
JetBrains ThinkRail 0.2.0: Agentic IDE Adds Per-Turn Receipts and MCP Permission Gates
ThinkRail 0.2.0 ships the two things agentic IDEs have been missing: a verified diff of what the agent actually changed (computed from the worktree, not the agent's own claims) and MCP servers with explicit per-call permission gates. JetBrains' Apache-licensed agent IDE has 1,658 GitHub stars and is trending on Hacker News.
Agent Supervision Becomes a Product Category: Six OSS Tools Ship in Three Days
Between October 22 and 24, six open-source tools launched targeting a single problem: AI agents operating without accountability. task-view audits every agent change against Git, Token Beggar routes subtasks to cheaper models saving 1.42x on quota, Lcc cuts context 21.8% without losing recall, and Peras pauses Claude Code before budget overruns. Product Hunt's Oct 24 leaderboard confirmed the shift from building agents to controlling them.
NVIDIA Nemotron 3.5 Bolt E4B: 4GB Agent Model Trained With Reinforcement Learning Only
NVIDIA released Nemotron 3.5 Bolt E4B on October 23, an 8B-param model for voice assistants, game NPCs, and robot interfaces on consumer hardware. Post-trained from Gemma 4 E4B with RL only, it hits 272.6 tokens/s on RTX 5070 and leads Scale Labs MultiChallenge for small models at 56.03%.
OpenAI Targets $30B at $1.25T Valuation in UAE-Led Pre-IPO Bridge Round
OpenAI is finalizing an up to $30 billion funding round at a $1.25 trillion pre-money valuation, led by UAE sovereign investors and Coatue, down $150 billion from the $1.4T discussed last month. Signatures targeted by end of October.
Kimi K3.1 Pro Launch Marred by Alleged Undisclosed Influencer Campaign
Moonshot's K3.1 Pro debut was shadowed by screenshots allegedly instructing creators to hide paid partnerships, a likely FTC violation. The model itself posts genuine gains: 65.7% on Terminal-Bench 4.0 and AA Intelligence Index 55, fourth overall.
Claude Haiku 5.5 Enters Text Arena: Haiku Jumped from 0% to 39.2% on Terminal-Bench 4.0 in One Generation
Claude Haiku 5.5 (High) joined Arena.ai's Text leaderboard on October 22. Its predecessor scored 0% on Terminal-Bench 4.0; it scores 39.2%. At 75% cheaper than Haiku 4.5 and the first Haiku with adjustable effort, the tier is no longer nominal on agentic tasks.
Kimi K3.1 Pro Debuts on Agent Arena at Rank 11 with the Lowest Tool Hallucination Rate on the Board
Moonshot's K3.1 Pro enters Agent Arena across 4,071 preliminary sessions with a 0.75% tool hallucination rate, lowest of 55 models. It ranks 5th in coding tasks at $2.73 per task, undercutting Claude Opus 5.5 Max by 5x.
Ornith-1.5-397B Reaches 86% SWE-bench Verified with a Self-Generating Training Loop
DeepReinforce's open-weight 397B MoE claims the top spot on the HuggingFace open-model SWE-bench leaderboard, up from 82.4% on Ornith-1.0. The gain comes from a closed training loop where the model proposes its own tasks, scaffolds, and rollouts.
GPT-6 Pro Gets Parallel Agents in ChatGPT: One Task, Multiple Agents, Progress Visible
OpenAI updated GPT-6 Pro in ChatGPT on October 16 to split tasks across multiple agents working simultaneously. Pro, Business, and Enterprise subscribers can track each agent's progress and inspect completed work. No benchmarks were released for the update.
Anthropic Launches Claude Duet: Opus 5.5-Powered Agent Monitors Your Apps, Costs Nothing Extra on Pro
Claude Duet, in research preview from October 21, is Anthropic's answer to Meta Muse and OpenAI Dots. Built on Opus 5.5, it monitors connected apps every few hours, generates daily briefings, and completes background tasks in a cloud VM. Available to Pro ($20/month) and Max subscribers in the US via waitlist.
Harvey Legal Agent Benchmark v1.1: Gemini 4 Argon Leads at 27%, Claude Opus 5.5 Ranks 17th
The Harvey Legal Agent Benchmark v1.1, updated October 22, flips the standard frontier rankings. Gemini 4 Argon leads at 27.1%. Meta Muse models sweep positions two through four. Claude Opus 5.5 finishes 17th at 9.2%. GPT-6 Astra ranks 19th at 7.1% despite the highest cost per test.
Gray Swan Rates Claude Opus 5.5 the Safest Coding Agent; Anthropic Claims 80% Cost Advantage Over GPT-6 Astra on FrontierCode
Anthropic's Opus 5.5 model page documents two claims absent from September benchmark coverage: the lowest prompt-injection success rate of any model in Gray Swan's evaluation, and a cost-per-task advantage of roughly 80% over GPT-6 Astra at default FrontierCode effort.
Same Model, Three Scaffolds: GPT-5.5 Spans 2.5 Points on Terminal-Bench 2.0 Depending on Agent Harness
Three separate GPT-5.5 entries on Terminal-Bench 2.0 run from 82.2% to 84.7% — a gap wider than the distance to the next model tier. The harness is now a primary variable in agentic benchmark results.
Kimi K3.1 Pro Sets Terminal-Bench 4.0 Record at 67%, Lands 4th on Intelligence Index with 11-Point Gain
Moonshot AI's K3.1 Pro posts the highest Terminal-Bench 4.0 score of any model at 67%, climbs to 4th on the AA Intelligence Index at 55 points — 11 above Kimi K3 — and will become the top open-weight model once weights are released. GDPval-AA agentic ELO jumps 312 points to 1,846.