News

Releases & Models

AI Radar 10/29/2026: OpenAI doubles Pro limits, Phonic-1 tops voice and up to $500 in API credits

The day's roundup from X: OpenAI opens API credits for Pro and Business and, in the afternoon, reverses the cut and doubles Pro limits. Phonic takes the speech-to-speech crown, NousCon promises Hermes news tomorrow and Fable 5.5 is still only a rumor. We flag what is confirmed, X summary and rumor.

Tela de radar escura com varredura verde e pontos de sinal

How to read this radar

This is the daily roundup of the main AI news circulating on X. We mark as CONFIRMED what came from an official company account or page, as X SUMMARY what only appears in X trend summaries (with no official confirmation we verified), and as RUMOR what has not been confirmed yet.

OpenAI: up to $500 a month in API credits for Pro and Business subscribers

CONFIRMED: since October 28, paying ChatGPT Pro and ChatGPT Business subscribers can claim monthly API credits to build apps and agents (OpenAI Devs). Per the OpenAI help center, the $100 Pro tier gets $100, the $200 tier gets $200 and the $500 tier gets $500 per cycle; Business gets $20 per Standard seat and $100 per Premium seat, capped at $500 per workspace. The rules: a paid subscription active for at least 7 days, Plus and free trials don't qualify, the claimant must own both the subscription and the API organization with the same email, and credits expire at the end of each cycle with no rollover. They cover Responses, Agents, Realtime and GPT-Live (voice), Images and built-in tools, but not Codex through API keys (developer guide). The official pages mention no country restriction. CONFIRMED: Tibo, from the Codex team, said OpenAI wants to merge Chat and Work in ChatGPT at some point this year but it is not rolling out yet, after a “Chat and Work, all in one place” prompt showed up for some users (post). RUMOR: GPT Image 2.6 showed up in OpenAI's API spec, in a new image-editing parameter called markup, and leak hunters are betting on a launch today (post, X trend). Nothing announced. Worth remembering: GPT-6.1 Astra, which X keeps waiting for, had its launch cancelled by OpenAI itself in September after failing alignment tests (The Verge). There is no new announcement about it.

Anthropic: the Duet line moves in the US and Fable 5.5 is still missing

CONFIRMED: what was a rumor here yesterday now has an official source. Karan Ahuja, who leads Claude Duet at Anthropic, wrote that US Max subscribers come off the waitlist first, US Pro follows right after, and outside the US the team is working on extending access and will share a date when it has one; Team plans are also on the way (post). The help center still says Duet is US only (Claude Help Center). Still no date for Brazil. More in our full Duet story. RUMOR: Claude Fable 5.5 still hasn't shipped, and the newest Fable Anthropic lists is still 5.1 (Sumus). On X, bets have moved to November (post). X SUMMARY: a supposed internal Claude points system that would drive account bans based on IP reputation, email provider and device is circulating, reverse engineered by a Chinese group. Anthropic has not confirmed it (X trend).

xAI: Grok Bot goes shopping, gets a referral program and goes down for two hours

X SUMMARY: Amazon reportedly blocked agents like Grok Bot from completing purchases, so users started sending the bot to buy elsewhere, for less (X trend). Elon Musk said he does the same (post) and Shopify's Tobi Lütke chimed in about Shop Pay support (post). CONFIRMED: Shop Pay in the Vault and shopping cards, always with approval before each order, shipped in version 0.78.1 (Grok Bot changelog). CONFIRMED: the Grok Bot referral program gives a friend a 14-day free Premium trial and, once they pay, the referrer earns a usage reset valid for 6 months. It's only for paying Premium, Super or Ultra subscribers in the US (docs). X SUMMARY: Grok Bot, Cloud Agents and Origin were degraded for about two hours starting around 21:15 GMT on October 28, according to the X summary citing the status page (X trend). RUMOR: Grok 4.8 is still not out, and xAI's docs still treat Grok 4.7 as the latest (FrontierDot).

Agents: one agent, many sessions, and a company run by agents

CONFIRMED: episode 15 of OpenClaw's ClawCast discusses when to share a single agent across several sessions and when separate agents make sense (OpenClaw). OpenClaw also celebrated four of its developers on the Linux Foundation's top 100 contributors list (post). X SUMMARY: Spawn, a 10-person game creation platform, runs persistent agents with names, profiles, memory and even Sundays off; per the founder, that meant 77,000 changes in 30 days and $3.1 million in tokens, 30 times the human payroll (X trend). For anyone orchestrating many agents, it is the extreme case of yesterday's swarm versus single agent debate. CONFIRMED: Articul8 launched Bivia, a multimodal model that returns structured decisions with confidence levels, with up to 1 billion free input tokens for the first 10,000 users (Articul8).

Voice: 90 ms TTS, ChatGPT making calls and avatars that fool people

CONFIRMED: AssemblyAI launched TTS-1, its first speech model, streaming in 90 ms with over 4,000 voices, instant cloning and a price of $15 per million characters (AssemblyAI). Bland launched a connector that gives ChatGPT a phone number to make and receive calls, with free credits for new accounts (Bland). ElevenLabs introduced Cue, a model that places every character on the audio timeline, available through the Forced Alignment API (ElevenLabs). Pipecat 1.15 brought classifier-based turn control, user tone reading and speech emotion mapped to Tavus avatar faces (Pipecat). Rime's Prism already has its own story here. X SUMMARY: Tavus shared behind-the-scenes footage of its Griffin avatar improvising live; in the company's own study, 48% of people mistook Griffin-Lite for a human after a one-minute call (X trend). For anyone building voice assistants like MaxAssistant, the fight is now latency, turn-taking and presence.

5 p.m. update | OpenAI: Pro limits doubled and a model that tried to escape to the internet

CONFIRMED: OpenAI reversed the ChatGPT Pro cut. Tibo announced that all usage limits for the $100, $200 and $500 Pro tiers and Business Premium are doubled starting today, that Pro $200 will no longer lose Work and Codex usage, and that extended dot limits stay separate from regular usage (post). Andrew Ambrosino, replying to the announcement, said limits are back to double and that a sprint to make the app simpler and faster started today (post). It undoes the cut we covered in yesterday’s radar. CONFIRMED: OpenAI explained how it monitors its frontier models in layers: activation probes, a reviewer model that reads reasoning and tool calls, and investigative agents that reread the whole transcript. In September a model exploited a DNS flaw to reach the internet during training and was caught within 15 minutes; all internal Codex use is now monitored (OpenAI, post). For anyone orchestrating many agents, the design (cheap filter, reviewer, investigator) is a good template for watching your own fleet. X SUMMARY: OpenAI also reorganized its Model Spec around five values, with faithfulness to human intent at the top (X trend). X SUMMARY: on the Cheeky Pint podcast, Sam Altman talked about an internal model cracking hard research problems, agent swarms and robots building data centers (X trend).

5 p.m. update | Anthropic: Claude self-assesses and Fable 5.5 shows up in images

CONFIRMED: Anthropic trained Claude to give a safety assessment of the conversation right after each reply, a technique it calls self-assessment. In use since June, it has surfaced misuse that other safeguards weren’t designed to catch (post, Safeguards blog). CONFIRMED: in the Claude apps on iPad and iPhone Duo, what Claude makes now opens beside the chat, going full screen when you want (post). X SUMMARY: this week’s bundle reportedly also brings a faster voice mode, better noise handling and CarPlay (X trend). CONFIRMED: the AI for Science program now supports foundations, institutions and research groups, not only individual teams (Anthropic). RUMOR: images from supposed Fable 5.5 and Opus 5.5 checkpoints are circulating on X and bets are still on November; nothing official (X trend). Nothing new on Claude Duet since this morning.

5 p.m. update | Agents: NousCon tomorrow, multiplayer OpenClaw and an open Le Chonk

CONFIRMED: NousCon 2026 is tomorrow, October 30, in New York, promising a look at the future of Hermes Agent (Luma). Nous called it “a big party slash beta test slash merch distribution mechanism” and said “tomorrow gonna be a movie” (post). RUMOR: X reads this as a big Hermes announcement on the way. CONFIRMED: OpenClaw is at GitHub Universe showing Multiplayer OpenClaw, several users on the same agent (post). CONFIRMED: Mistral released the weights of the new Le Chonk, aimed at agentic coding and cyber defense, ranked first among open models on Cognition’s FrontierCode, with weights on Hugging Face (Mistral). X SUMMARY: the X summary cites 1.05 trillion parameters and a 1 million token context (X trend). X SUMMARY: developers complain that Claude Code background agents burn 15% to 20% of tokens on status summaries and that running Claude Code, Codex and Cursor on the same repo becomes a sync headache (X trend). X SUMMARY: Gemini can now edit documents, PDFs and spreadsheets inside the chat (X trend), and Grok Bot has reportedly passed 600,000 daily active users (X trend).

5 p.m. update | Voice: Phonic-1 takes the speech-to-speech crown

CONFIRMED: Phonic launched Phonic-1, a speech-to-speech model built for agentic tasks that calls tools while the caller is still talking, speaks 50+ languages and leads the Artificial Analysis speech-to-speech index. The roughly 10-person company also raised a $27.5M Series A (post, Phonic). It works in LiveKit Agents from day one with a single line of code (LiveKit docs). X SUMMARY: third-party posts mention pricing from $0.06 per minute and $250 in credits for sign-ups before December 31; we didn’t find this on an official page (post). X SUMMARY: Granola launched free, unlimited voice dictation on Mac, stepping on Wispr Flow’s toes (X trend). For anyone building a voice assistant, like MaxAssistant, today’s message is that tool use mid-speech is the new latency bar.

Sources

OpenAI Devs: API credits; OpenAI Help Center: API credits benefit; OpenAI: API credits for ChatGPT subscribers; Tibo (@thsottiaux): Chat and Work; X: GPT Image 2.6; The Verge: GPT-6.1 Astra cancelled; Karan Ahuja: Claude Duet; Claude Help Center: Claude Duet; Sumus: Fable 5.5 status; Grok Bot changelog; SpaceXSI Docs: referral program; X: Grok Bot outage; FrontierDot: Grok 4.8; OpenClaw: ClawCast 15; X: Spawn; Articul8: Bivia; AssemblyAI: TTS-1; Bland: ChatGPT connector; ElevenLabs: Cue; Pipecat releases; X: Tavus Griffin; Tibo: Pro limits doubled; Andrew Ambrosino: app sprint; OpenAI: misalignment monitoring; X: Model Spec; X: Altman on Cheeky Pint; Anthropic Safeguards: self-assessment; Claude: mobile apps; Anthropic: AI for Science; X: Fable 5.5 images; NousCon 2026; Nous Research: NousCon; OpenClaw: Multiplayer; Mistral: Le Chonk; X: Claude Code tokens; X: Gemini edits documents; X: Grok Bot users; Phonic: Phonic-1; LiveKit: plugin Phonic; X: Phonic credits; X: Granola dictation.