2026-09-04 · FridayPublished daily at 08:00
AI Daily
- 1
- 2
- 3
01Model Releases/Updates
OpenAI releases GPT-6 Astra: a computer-use model with a 1.05M context window, with access restricted after reaching the Critical cybersecurity threshold
OpenAI releases GPT-6 Astra, positioned as a computer-use model, with a 1,050,000-token context window, 128,000 maximum output tokens, and a knowledge cutoff of April 30, 2026. It scores 72.6% on OSWorld V2-Offline (GPT-5.6 Sol scores 65.7%), while average task time falls from approximately 75 minutes to 40 minutes.
Original textOpenAI releases its new Astra model, focused on computer and browser operation but sparking controversy over opaque recurrence
OpenAI Releases Its Latest Model, Astra, Calling It the Most Powerful Model to Date. It Focuses on Computer and Browser Operation, Initially Available to customers of the Daybreak cybersecurity program, followed by Pro, Plus, Enterprise, Business paid accounts, and the API within a week.
Original textOpenAI begins rolling out GPT-6 Astra to all Plus users
OpenAI announces that it has begun rolling out GPT-6 Astra, saying it is proceeding as cautiously and quickly as possible, with a focus on making it available to all Plus users rather than limiting it to Pro, Business, and Enterprise plans. The rollout will take several days to complete; multiple entirely new systems will be running at large scale for the first time, and the team is bringing substantial compute capacity. See openai.com/index/gpt-6-astra/ for details.
Original textOpenAI releases GPT-6 Astra, achieving state-of-the-art results on multiple benchmarks
OpenAI releases GPT-6 Astra, achieving state-of-the-art results on FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0, and leading scores on Terminal-Bench Science 0.1 and HealthBench Pro.
Original textIFM Releases Six Open-Source K2 Horizon Models, Covering 0.9B to 375B-A23B, with the Full Training Lifecycle Opened
Original textOpenAI releases GPT-6 Astra, the first model to reach the critical-level cybersecurity capability threshold
OpenAI released its next-generation large language model GPT-6 Astra on September 3. It is the company's first model to reach the Critical cybersecurity capability threshold in its Preparedness Framework and can discover unknown vulnerabilities in heavily protected systems without step-by-step guidance.
Original textOpenAI releases GPT-6 Astra, with an official benchmark score of 99.9% on ARC-AGI-3
Original textOpenAI Releases GPT-6 Astra, Comprehensively Surpassing Claude Fable 5.1 on Benchmarks
Original text02Product Releases/Updates
Hugging Face Releases Open-Source Tool funes, Providing Coding Agents with a Locally Held Memory Layer
Hugging Face releases funes, an open-source tool that provides a local memory layer for coding agents such as Claude Code, Codex, pi, and Hermes. It indexes existing conversation logs as Lance datasets, and a single funes add command lets an Agent retrieve original sources on its own (Agent, timestamp, session, and turn).
Original textxAI designs Grok Bot: Rebuilding the interface for persistent agents
xAI published a design article explaining how Grok Bot’s interface is designed for persistent agents that go beyond a single session. The product centers on Bots rather than conversations; Bots have identities, memories, their own computers, and tools. Avatar animations convey states such as idle, working, waiting, blocked, thinking, and completed. The workspace provides three levels of access: status, preview, and takeover.
Original textOpenAI launches Daybreak for Frontline Defenders, investing $1 billion to support frontline cyber defense
OpenAI launches the global Daybreak for Frontline Defenders initiative, committing $1 billion in subsidized Daybreak access, training, technical support, and collaboration, to be used over the next six months. Priority will go to frontline defenders with limited resources, including water treatment facilities, power grids, state and local governments, community banks, nonprofits, and open-source maintainers.
Original textxAI releases the enterprise edition of Grok Bot, with Grok and Cursor Enterprise customers receiving two weeks free
xAI Announces Grok Bot for Businesses, Allowing Grok and Cursor Enterprise Customers to Use It Free of Charge for the Next Two Weeks and Invite Their Entire Organization, Including Employees Without Existing Seats.
Original text03Industry News
NVIDIA announces the acquisition of Hugging Face for $12.9303 billion
NVIDIA announced that it has agreed to acquire Hugging Face for $12,930,300,000, with Huang Renxun announcing the news on the official blog. Hugging Face currently has more than 18 million developers, hosts over 3 million models, 500,000 datasets, and 1 million applications, and serves more than 200,000 companies.
Original text04Tips and perspectives
Artificial Analysis evaluates GPT-6 Astra: matches Fable 5 on coding agents, but its price rises to 2.5 times as much
Artificial Analysis Releases Its Evaluation of GPT-6 Astra. It Scores 67 on the Coding Agent Index, Roughly Equivalent to Claude Opus 5 and Fable 5, at Less Than Half the Cost of Fable 5; Its Token Efficiency Is About 70% Higher Than GPT-5.6 Sol (max).
Original textFrançois Chollet evaluates GPT-6 Astra's performance on ARC-AGI-3
François Chollet writes that GPT-6 Astra delivers a step change in capability on interactive reasoning tasks, scoring 66% on ARC-AGI-3 with a standard harness and approaching 100% with a continuous-conversation harness and custom compaction, at a cost of about $360 per game.
Original textRohan Paul interprets the safety findings in the 117-page system card for OpenAI GPT-6 Astra
Rohan Paul Summarizes the Key Points from OpenAI GPT-6 Astra's 117-Page System Card: Astra's Ability to Control Its Own Chain of Thought Leapt from 16.1% for GPT-5.6 Sol to 60.9%, with a Corresponding Decline in Monitorability.
Original textGary Marcus evaluates GPT-6 Astra: clear progress, but robustness and monitorability remain questionable
Gary Marcus comments on GPT-6 Astra, saying that multiple reports indicate genuine progress and that OpenAI's product explicitly creates and manipulates symbolic world models, validating arguments he has made for nearly a decade.
Original textTom Tunguz analyzes the data-for-compute logic behind Meta Muse Spark's dual-track pricing
Tom Tunguz Analyzes Meta's Release of the Muse Spark Model and Its Two-Tier API Pricing: Standard Tier (muse-spark-1.3) Input at $1.25/m.
Original textGoogle Cloud teaches you how to build an always-on Agent with Cloud Run instances for $5.70 per month
Shir Meir Lador explains on the Google AI Developers Blog how to use Cloud Run instances to run a persistent agent in the cloud 24/7 for $5.70 per month (1 vCPU, 1Gi memory, shared CPU).
Original textMeta Muse Spark 1.3's comparative evaluation results against Claude on the Artificial Analysis Coding Agent Index are released
The Artificial Analysis Coding Agent Index Shows Meta Muse Spark 1.3 (max) Scoring 68 with Muse Code, Second Only to Claude Code + Opus 5 (xhigh), Which Also Scores 68.
Original textARC-AGI-3 was saturated by Astra just six months after launch, progressing twice as fast as François Chollet expected
Sherwin Wu says he once thought ARC-AGI-3 was very difficult, but Astra has now saturated the benchmark. Quoting François Chollet, he notes that when ARC 3 launched, Chollet expected frontier models to take about a year to saturate it; in fact, it took only 6 months, roughly 2 times faster than expected. The capabilities of the new generation of models will challenge views of AI formed around older models.
Original text