Friday, September 4, 2026

OpenAI releases GPT-6 Astra on Sept. 3 with staged access

OpenAI released GPT-6 Astra on Sept. 3, 2026, initially to a limited set of organizations, with rollout to ChatGPT Plus, Pro, Business and Enterprise to follow over the coming days, according to OpenAI's announcement as relayed by several channels.

Evidence: 1 first-party, 0 hands-on, 7 relaying

Watch: OpenAI: Introducing GPT-6 Astra for developers

ARC Prize reports Astra 62.7% on ARC-AGI-3 in standard harness, 99.9% with adapter

ARC Prize reported that GPT-6 Astra scored 62.7% on the ARC-AGI-3 semi-private set under its standard harness at max reasoning, and 99.9% under a provider adapter that preserves private reasoning state and uses OpenAI compaction, according to AI Code King and Prompt Engineering, who relayed ARC Prize data.

Evidence: 0 first-party, 0 hands-on, 5 relaying

Watch: Prompt Engineering: GPT-6 Astra: The harness matters more than you think

OpenAI prices GPT-6 Astra at $10 input and $50 output per million tokens

Astra's API price is $10 per million input tokens and $50 per million output tokens, matching Anthropic's Claude Fable 5.1 and 2.5 times GPT-5.6 Sol's $4 and $20, according to AI Code King and Theo, who relayed OpenAI's launch figures.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: Theo - t3.gg: It's Here.

OpenAI reports Astra at 71.6% to 73% on OSWorld 2.0, versus 65.7% for GPT-5.6 Sol

OpenAI's launch materials, as relayed by four channels, put GPT-6 Astra's OSWorld 2.0 computer-use score at 72.6% versus 65.7% for GPT-5.6 Sol, with tasks finishing in about 40 minutes versus about 75.

Evidence: 0 first-party, 0 hands-on, 4 relaying

Watch: AI Code King: GPT-6 Astra (Benchmarks Deep-dive): This is not a good coding model an

OpenAI classifies Astra as first model at critical cyber level in its Preparedness Framework

OpenAI said GPT-6 Astra is the first of its models to reach the critical cyber threshold in its Preparedness Framework, according to AI Code King, Fireship and Julian Goldie, who relayed the system card.

Evidence: 0 first-party, 0 hands-on, 4 relaying

Watch: AI Code King: GPT-6 Astra (Benchmarks Deep-dive): This is not a good coding model an

System card: Astra shows lower chain-of-thought monitorability in OpenAI tests

OpenAI's system card for GPT-6 Astra reports lower chain-of-thought monitorability than earlier models, according to AI Explained, and OpenAI researchers cited by the speaker worry it may sandbag on safety tasks.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: AI Explained: GPT 6 Astra, so good even OpenAI are worried

Astra scores 97.6% to 98% on Frontier Math Tier 4; Epoch says it solved 2 of 68 unsolved problems

OpenAI's launch materials put Astra at 97.6% on Frontier Math Tier 4, against 83% for GPT-5.6 Sol and 87.8% for Fable 5.1, according to AI Code King; AI Explained gave about 98% at peak and 83% without reasoning.

Evidence: 0 first-party, 0 hands-on, 3 relaying

Watch: AI Code King: GPT-6 Astra (Benchmarks Deep-dive): This is not a good coding model an

Astra scores 57.9% on Terminal Bench 4.0 versus 55.8% for Fable 5.1, with near-ties elsewhere

On Terminal Bench 4.0, GPT-6 Astra scored 57.9% versus 55.8% for Claude Fable 5.1, 52.3 for Claude Opus 5 and 37.3 for GPT-5.6 Sol, AI Code King reported from OpenAI's table.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: Theo - t3.gg: It's Here.

Artificial Analysis rates Astra 61 on its Intelligence Index, level with GPT-5.6 Sol and below Fable 5.1

Artificial Analysis gave GPT-6 Astra 61 on its Intelligence Index, equal to GPT-5.6 Sol and five points behind Claude Fable 5.1 at 66, according to AI Code King and Fireship.

Evidence: 0 first-party, 0 hands-on, 3 relaying

Watch: AI Code King: GPT-6 Astra (Benchmarks Deep-dive): This is not a good coding model an

Testers report Astra completing long computer-use tasks, with refusals and skipped steps noted

Testers who used GPT-6 Astra's computer use reported long unattended runs.

Evidence: 0 first-party, 4 hands-on, 0 relaying

Watch: Nate Herk: GPT-6 Astra Made This Entire Video (high hype)

Users report Astra Codex runs lasting days, and a repo latency cut from 800 ms to under 30 ms

Matthew Berman said he ran /goal with Astra for five straight days to build a SimCity clone that was still unfinished, with lag fixed after he asked for frame-rate optimization.

Evidence: 0 first-party, 2 hands-on, 1 relaying

Watch: Theo - t3.gg: It's Here.

OpenAI presents ChatGPT Work mode with plugins for files, email and scheduled automations

OpenAI described ChatGPT Work as a switchable mode in which plugins connect to Drive, Office, email, Figma, Canva and Notion so ChatGPT can create documents, decks and apps and run scheduled computer or browser automations.

Evidence: 1 first-party, 1 hands-on, 0 relaying

Watch: Riley Brown: 8 ChatGPT Agents That Do My Work for Me (Steal These)

Testers report Fable 5.1 finishing long tasks, including a 7-sheet DCF model and a Blender film

Nate B Jones tested Claude Fable 5.1 on several tasks.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: Nate B Jones: Everyone's Testing Claude Fable 5.1 On Code. It Made Me A 37-Second Fi

Institute of Foundation Models releases six K2 Horizon open models under Apache 2

The Institute of Foundation Models (MBZUAI) released six K2 Horizon models from sub-1B to a 375B mixture-of-experts flagship, with Apache 2 weights, training data, code and checkpoints, according to Fahd Mirza and Julian Goldie, who relayed the lab's claims.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: Fahd Mirza: K2 Horizon: 0.9B, 7B, and 32B Tested Locally, Real Results

MiniMax M3 open model reported at roughly 400B parameters with 1M context

A MiniMax guest on the AI Engineer stream said the company released M3, an open model with roughly 400B total and 20B active parameters, vision input and 1M-token context, and is already building M3.1.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: AI Engineer: Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, M

Alibaba's Qwen 3.8 Max described as 2.4T-parameter model with 1M-token context

Julian Goldie and GitHub's presenter said Alibaba's Qwen 3.8 Max has 2.4 trillion parameters and a one-million-token context.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Testers find Astra comparable to Fable 5.1 on writing, design and coding, with mixed results

Every's team gave mixed early-access verdicts on GPT-6 Astra versus Claude Fable 5.1: Dan Shipper uses Astra daily but reaches for Fable on the largest tasks.

Evidence: 0 first-party, 3 hands-on, 2 relaying

Watch: Every: VIBE CHECK: GPT-6 ASTRA

Google releases Gemini 3.8 Flash on Sept. 2 with 1M-token context and three thinking levels

Google released Gemini 3.8 Flash on Sept. 2, 2026, available in the Gemini API, AI Studio, Antigravity, Stitch and the Gemini app for AI Pro and Ultra subscribers, according to Julian Goldie's summaries of Google's announcement.

Evidence: 0 first-party, 0 hands-on, 2 relaying

Watch: Matt Wolfe: AI News: The Most Insane Week So Far This Year!

Sentdex reports GLM 5.3 Flash running about 170-180 tokens per second locally

Sentdex said GLM 5.3 Flash, about 320B parameters with vision, runs at about 170 to 180 tokens per second at native precision on his RTX Pro 6000 setup, versus about 350 for DeepSeek V4 Flash 0731.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: Sentdex: All Roads Lead back To GLM!

Hugging Face releases 200+ WebGPU kernels and a Kernels JavaScript library

Hugging Face released more than 200 open-source WebGPU kernels and a JavaScript Kernels library, shipped as Jinja templates that render WGSL for the device and load from the Hub.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Hugging Face: We shipped 207 WebGPU Kernels for Browser AI

MiniMax describes sparse attention and native multimodal pretraining for M3

A MiniMax guest said M3 uses MiniMax Sparse Attention, an index branch selecting context blocks and a sparse branch computing on them, designed by an intern for 1M-token agent contexts.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: AI Engineer: Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, M

Bijan Bowen tests Fish Audio S2.1 Pro voice cloning from 15 seconds of audio

In a sponsored test, Bijan Bowen judged Fish Audio S2.1 Pro's instant clone from a 15-second clip very convincing, though the judgement was subjective and he tested Chinese output with two takes of differing accent quality.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Bijan Bowen: Fish Audio S2.1 Pro Full Test – Building A Video Translation Pipeline!

GPT-5.6 Sol builds a transcribe-translate-dub pipeline in Bijan Bowen's test

Bijan Bowen used GPT-5.6 Sol in the ChatGPT Mac app at extra high to build a local pipeline using a Whisper-style transcriber, Qwen 3.8 Next (Q4_K_M) for Chinese translation, Fish API dubbing and a web UI, on a 6000 Pro machine.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Bijan Bowen: Fish Audio S2.1 Pro Full Test – Building A Video Translation Pipeline!

Reviewers discuss Meta paper on LLM research preference models for choosing experiments

Hugging Face paper reviewers discussed a Meta paper in which research preference models pick which candidate experiments to evaluate, with an inference-only variant and an agentic variant that can run pilot experiments.

Evidence: 0 first-party, 0 hands-on, 1 relaying