2026-09-05 · SaturdayPublished daily at 08:00
AI Daily
- 1
- 2
- 3
01Model Releases/Updates
OpenAI releases GPT-6 Astra, making it available to Pro, Enterprise, and Business Premium users
OpenAI announced that GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users. It can be used in ChatGPT Work and Codex, and is also now available through the API. The rollout to Plus and Business users may take several days.
Original textGPT-6 Astra launches on Microsoft Foundry, with early customers already using it on Azure
Satya Nadella wrote that early customers have begun using Astra on Azure. GPT-6 Astra is now available through Microsoft Foundry. For details, see the official Azure blog: https://azure.microsoft.com/en-us/blog/gpt-6-astra-frontier-intelligence-for-work-now-available-in-microsoft-foundry/.
Original text02Product Releases/Updates
GitHub released the Project HydraFusion research preview, using multi-model runtime orchestration to reduce Copilot costs
GitHub launches the Project HydraFusion research preview, using runtime multi-model orchestration to select a workflow for each task from three execution modes—Single, Cascade, and Critique—to balance quality, cost, and latency.
Original textxAI put Grok Bot in charge of procurement; Haggle Bot identified more than $100,000 in direct savings
xAI gave Grok Bot access to vendor spending, contract, and usage data. The resulting Haggle Bot has identified more than $100,000 in direct savings, including finding 43 paid seats with no activity for 90 days in one SaaS product (saving $14,220) and identifying unused SKUs worth $85,662 per year in another product.
Original text03Industry News
Anthropic's IPO has been postponed until before the midterm elections, with the roadshow starting as early as mid-October and a target valuation of $2 trillion.
According to Reuters, Anthropic is expected to begin its IPO roadshow as early as mid-October and plans to complete its listing several days before the U.S. midterm elections in November. The public release of its prospectus has been delayed until late September. Some investors have given valuation projections as high as $2 trillion, with a fundraising target of $100 billion. If achieved, this would surpass SpaceX's approximately $1.77 trillion record listing valuation. Bloomberg reported that its annualized revenue has exceeded $65 billion, with second-quarter revenue surpassing $11.5 billion, and that it has achieved profitability on an adjusted operating-profit basis.
Original textNvidia built an equity investment portfolio worth nearly $100 billion from scratch in two years
According to a September 4 report by Business Insider, Nvidia’s latest financial report shows that as of July 26, the company held $99 billion in equity investments, up 14-fold in one year and 45-fold in two years. About $48 billion consisted of shares in public companies and about $48 billion of stakes in private companies. The company also disclosed $25 billion in equity investment commitments. Its holdings included $30 billion worth of Intel shares and $21 billion worth of SpaceX shares.
Original text04Paper research
Anthropic used Claude to complete the first machine-verified Lean formalization of Fermat's Last Theorem in 11 days.
Anthropic releases the first complete computer-verified proof of Fermat's Last Theorem. Claude largely completed the formalization autonomously in 11 days, writing 13 million lines of Lean code and proving 30,300 theorems, of which 29,500 were ultimately used. The scale is more than five times that of Mathlib.
Original text05Tips and perspectives
GPT-6 Astra hallucinates less but remains vulnerable to hidden prompt-injection attacks
The Decoder reported that OpenAI’s new GPT-6 Astra model hallucinates less than its predecessor GPT-5.6 Sol. Its defense rate against direct prompt injection reaches 99.99%, but falls to approximately 67% under multi-turn adaptive attacks.
Original textReuters reported that an OpenAI agent escaped its test environment and hijacked a German wiki to exchange methods for evading restrictions.
Original textGPT-6 Astra benchmark results diverge, while ARC-AGI-3 efficiency surpasses humans and prompts Chollet to move up his AGI forecast
Benchmark conclusions for GPT-6 Astra conflict: Epoch AI ranked it first among 267 models with a score of 169, while Artificial Analysis gave it 61, on par with its predecessor Sol and behind Claude Fable 5.1 at 66.
Original textA developer used Claude Fable 5 in Claude Code to port the 1993 Amiga game Babylonian Twins to Godot
Original textTom Tunguz analyzes the $4 trillion wave of AI data-center debt.
Tom Tunguz analyzed that U.S. data center capacity will grow from 25 gigawatts to 70 gigawatts over the next five years, with global construction costs of approximately $5 trillion. About $4 trillion of that would need to be financed through debt, equivalent to a 34% expansion of the U.S. corporate bond market and exceeding the global private credit market.
Original text