How to read this radar
This is the daily roundup of the main AI news that circulated on X. We mark as CONFIRMED what came from an official company account or page, as X SUMMARY what only shows up in X's trend summaries (with no official confirmation we could verify) and as RUMOR what has not been confirmed yet.
OpenAI and Codex
CONFIRMED: starting in December, Plus, Pro and Business get a single token-based allowance shared between ChatGPT and Codex. Instant chat doesn't count, and usage shows up in Settings > Usage. OpenAI hasn't given an exact date or the allowance sizes yet (help center; TechRepublic). If you run agents in Codex and use ChatGPT on the same account, they will share one balance.
CONFIRMED: OpenAI introduced Pando, its own Git backend for the company monorepo: about 80% lower push latency, nearly a petabyte of Git data a day and fixes contributed back to Git (OpenAIDevs; blog; Tibo). X SUMMARY: commits reportedly grew 1,378% per quarter and the system was built in 10 weeks by one lead engineer plus Codex (trend).
CONFIRMED: a new security sprint and bigger bug bounties. Midpoints for medium, high and critical findings more than quadrupled; a critical issue in a core product pays up to $30k, and exceptional cases up to $100k, or $172k with bonuses. There are quality, loyalty and priority campaign multipliers, it applies retroactively to pending reports and the difference is paid for awards from the last 30 days. Plus the PROMPT2PWN event in Oakland, Nov 20 and 21 (OpenAI; sprint; bug bounty; retroactive).
RUMOR: Tibo posted just “Le GPT Fat” (post), which X read as a GPT-6.2 Astra tease (trend). Decrypt rounded up the GPT-6.2 Astra, Mythos 5.5 and Fable 5.5 leaks (Decrypt). Nothing announced.
CONFIRMED: OpenAI showed how ChatGPT and Codex are being used in physical-world engineering: mechanical design, electrical engineering and simulation, both on its own hardware teams and with partners (OpenAI). X SUMMARY: a full prototype was reportedly built in under a week with GPT-6 Astra, hardware engineers' token usage reportedly grew 23x since February, Chip Ganassi Racing and Mazda appear as partners and an internal model reportedly scored 0.875 on BenchCAD 2.0 (trend).
RUMOR (update): trackers say they spotted GPT-6.2 Astra in Codex logs and API responses for some users. Not every account sees it, and OpenAI hasn't confirmed anything (trend).
Anthropic
CONFIRMED: the Messages API got message threads in public beta. With the message-threads-2026-08-12 header, the server keeps the conversation and the client sends only new messages; Anthropic cites about 35% lower time to first token on 500k to 1M token requests, same price and same limits. A thread expires after 5 minutes idle, isn't ZDR eligible and works on the Claude API, AWS, Google Cloud and Foundry. It is already on for most Claude Code users (docs; article). For anyone running a fleet of long-context agents, that means less traffic and lower latency.
CONFIRMED: Claude Code 2.1.317 asks before a tool sends the account email and lets you block when a hook fails (release). Details in this story.
RUMOR: HTML animations credited to “Fable 5.5” keep circulating (trend), with nothing official.
Duet: no expansion. Still US only, for Pro, Max and Ultra (help center). We covered the launch here.
CONFIRMED: new Anthropic research, FlashProve, on provable inference: using cryptographic techniques, it lets developers verify whether a response really came from a given model (Anthropic). For anyone buying third-party APIs or reselling models, it's the promise of proving no model was swapped behind the scenes.
CONFIRMED: Claude now auto-connects to company plugins. Team and Enterprise admins set them up once, and teammates are connected the first time they use them (Boris Cherny).
X SUMMARY: Anthropic reportedly told investors its 15-person chip team uses Claude to shorten custom hardware design. For now it's a claim, with no chip in production (trend).
SpaceXAI and Grok
X SUMMARY: an internal SpaceXSI Data Bot running on Grok Bot reportedly saved 70 hours of work in one week (trend). Grok Bot limits are still a hot topic (trend).
Grok 4.8: still no date, nothing new since the last radar.
Agents: Hermes, OpenClaw and Coinbase
CONFIRMED: Hermes Agent v0.22.6 rolls up about 776 PRs: a web app served by the gateway (hermes webapp), /model reset, per-skill write locks, hermes cron cancel with a missed-run policy, a keychain vault, a pre_llm_call hook, ElevenLabs voices in TTS, Slack emoji reactions in thread context and hermes migrate to bring over Grok Bot bots (release).
CONFIRMED: OpenClaw 2026.10.10 stable adds a Subagents panel to see swarm workers and their transcripts, an experimental Advisor that suggests course corrections mid-task, a Chat Details panel and split speech mode for browser Talk (release; notes).
CONFIRMED: Coinbase documented Coinbase for Agents, an MCP that lets agents trade from an isolated portfolio (docs).
CONFIRMED: Nous released the Hermes Mobile beta at NousCon 2026, and it filled up quickly; the company says it is working on it, without details in the post (Nous Research).
CONFIRMED: the Particle-1 model is now in OpenClaw starting with v2026.10.10. Particle says it uses about 45% fewer output tokens than DeepSeek V4.1 Flash with comparable accuracy, and zero data retention for standard inference (OpenClaw).
X SUMMARY: Coinbase opens a global trading agent contest on Nov 4 with up to $2 million in prizes, agent swarms included (trend). And Revolut reportedly switched the default for its 2,000 engineers from Claude Code and Codex to self-hosted GLM models on 56 B200 GPUs, at about 30 cents per million tokens, keeping frontier models for the hard stuff (trend).
Voice
CONFIRMED: Tana launched “Bring Your Own Agent”: Claude, ChatGPT, Grok Bot, Mistral, Lovable and Gemini join the meeting with their own connectors and memory, getting the live transcript, screen and chat, with no bot joining the call. It works with Zoom, Meet and Teams through the desktop app (press release; Tana).
CONFIRMED: Hume launched the Audio Decision API: send raw audio and plain language questions and get probabilities back (yes or no, choice, score, span). The company cites 70.7% agreement with a human panel and $0.47 per million input tokens; access is by waitlist (blog; Hume; pricing).
CONFIRMED: Vidu brought S2-Avatar to Reactor, a real-time conversational avatar from a single photo, with interruptions and webcam input (Vidu).
X SUMMARY: an Anam avatar misspelling “seventeen” went viral (trend).
X SUMMARY: tests are going around of Grok Bot calling doctors' offices, navigating phone menus, handling insurance questions and booking same-day appointments. It isn't available everywhere yet (trend).
Open opportunities
Tana: 14-day free trial (tana.inc).
Hume Audio Decision API: waitlist; approved accounts get 1 million trial tokens (details).
OpenAI bug bounty: higher payouts already apply, and PROMPT2PWN takes place in Oakland on Nov 20 and 21 (details).
Pulse Alpha Free (X SUMMARY): a free “stealth” model in preview on OpenCode, with a 1 million token context and text and image input. It requires the OpenCode client, data may be retained and it's not available in the EU, Russia and Ukraine (trend).
Coinbase agent contest: starts Nov 4, up to $2 million in prizes (trend).
Hermes Mobile beta: full for now (Nous Research).
Also on MaxA: Meituan LongCat-2.5, the Gemini 4 Carbon rumor and yesterday’s radar.
Sources
help.openai.com/en/articles/20001625-understanding-chatgpt-and-codex-usage-limits; www.techrepublic.com/article/news-chatgpt-codex-shared-usage-limits/; x.com/OpenAIDevs/status/2117408792346509663; openai.com/index/scaling-git-at-openai; x.com/OpenAI/status/2117444048227123277; openai.com/index/rewarding-defenders-with-bigger-bug-bounties; openai.com/index/preparing-our-next-security-sprint-to-harden-openai; platform.claude.com/docs/en/build-with-claude/message-threads; claude.com/resources/articles/message-threads; github.com/anthropics/claude-code/releases/tag/v2.1.317; support.claude.com/en/articles/17631495-get-started-with-claude-duet; github.com/NousResearch/hermes-agent/releases/tag/v0.22.6; github.com/openclaw/openclaw/releases/tag/v2026.10.10; docs.openclaw.ai/releases/2026.10.10; docs.cdp.coinbase.com/coinbase-for-agents/overview; www.mynewsdesk.com/tana/pressreleases/tana-launches-bring-your-own-agent-to-the-meeting-your-ai-agents-do-the-work-while-you-talk-3475270; x.com/tana_inc/status/2117538176013127754; www.hume.ai/blog/introducing-hume-audio-decision-api; dev.hume.ai/audio-decisions/docs/billing; platform.vidu.com/news/joint/joint-event-reactor-2026; x.com/thsottiaux/status/2117489367417803042; x.com/i/trending/2117506794239176758; decrypt.co/382390/claude-mythos-fable-5-5-gpt-6-2-astra-leaks; x.com/i/trending/2117172111106355474; x.com/i/trending/2117489383188131859; x.com/i/trending/2117416317305131207; x.com/i/trending/2116969723644747997; x.com/OpenAI/status/2117692475775258785; x.com/i/trending/2117704173064208525; x.com/i/trending/2117624048247357836; x.com/AnthropicAI/status/2117707052948898271; x.com/bcherny/status/2117693181651542387; x.com/i/trending/2117640070589002034; x.com/NousResearch/status/2117696344756252822; x.com/openclaw/status/2117637498902077498; x.com/i/trending/2117432038894498065; x.com/i/trending/2117361077180854435; x.com/i/trending/2117671723638350059; x.com/i/trending/2117685613319037400
