AI NewsOpen AI

An engineer paired a fast decision-making model with GPT-6 Astra's planning to defeat Minecraft's Ender Dragon in under 9 minutes for less than a dollar, open-sourcing the code behind the run.
Engineer Ronak Malde paired Jev, a fast decision-making model, with GPT-6 Astra for planning, beating Minecraft's Ender Dragon in 8 minutes 43 seconds from an empty inventory. The run cost under a dollar total: about a penny for Jev, 96 cents for Astra.
The split mirrors "thinking fast and slow": Astra sets high-level goals, while Jev makes near-instant tactical calls during combat and movement. A rival team using a different fast model, Jesse, shaved the run to 7 minutes 6 seconds. The code is open-sourced.
Key Capabilities:

• Split roles: Astra handles high-level planning; Jev makes near-instant tactical decisions.
• Dirt cheap: Total model cost under a dollar for the full run.
• Open-sourced: Full harness code is public for others to replicate or improve on.

OpenAI launched Astra for Law, a version of GPT-6 Astra paired with a legal search index of over 230 million URLs, scoring 54% on legal research benchmarks versus 39% for the base model.
OpenAI introduced Astra for Law on September 17, pairing GPT-6 Astra with a legal search index covering over 230 million URLs. On Vals AI's Legal Research Bench, it scores 54% correctness versus 38.7% for GPT-6 Astra with web search alone.
Harvey and Sullivan & Cromwell helped build custom workflows on it, and access rolls out to select firms through Trusted Access in ChatGPT and Codex first. API access for Harvey and Legora comes soon, as OpenAI frames this as the start of a longer legal push.
Key Capabilities:

• Legal-grade search: Indexes over 230 million URLs of U.S. case law and statutes, drawing heavily from CourtListener.
• Measurable gains: 54% correctness on legal research questions, up from 38.7% for the base model.
• Firm partnerships: Sullivan & Cromwell, Ropes & Gray, and Cooley built custom tools with OpenAI's help.

OpenAI will retire GPT-5.5 from ChatGPT and Codex on October 14, 2026, across all plans, though the model stays available through the API for developers who still need it.
OpenAI will retire GPT-5.5 from ChatGPT, ChatGPT Work, and Codex on October 14, across all plans. Codex users pointing scripts or agents at 'gpt-5.5' need to switch to 'gpt-5.6-sol' before then, or their workflows will break.
GPT-5.5 will remain reachable through the OpenAI API and in Codex sessions using an API key, so it isn't disappearing entirely. The turnover is fast: GPT-5.5 is already being pushed out by GPT-5.6 Sol and GPT-6 Astra.
Key Capabilities:

• Hard cutoff: GPT-5.5 disappears from ChatGPT, ChatGPT Work, and Codex on October 14.
• API stays open: Developers can still reach GPT-5.5 through OpenAI's API after the ChatGPT retirement.
• Migration required: Codex scripts and agents using 'gpt-5.5' need to switch to 'gpt-5.6-sol'.

OpenAI launched a Data agent for ChatGPT Work that connects to company databases, plus ChatGPT for Financial Services, built with Morgan Stanley and Evercore for investment banking research and pitchbooks.
OpenAI introduced two ChatGPT Work tools on September 10. A Data agent connects to sources like BigQuery, Snowflake, and Datadog, turning questions into dashboards. ChatGPT for Financial Services, built with Morgan Stanley and Evercore, pairs GPT-6 Astra with data from PitchBook and Crunchbase.
Both target work usually done by junior staff: pulling numbers and building the pitchbooks that take analysts hours. Every output cites its source, so decisions stay traceable as more research gets automated.
Key Capabilities:

• Plain-language analytics: The Data agent turns questions into shareable dashboards, no queries or new tools needed.
• Wall Street-ready: Built with Morgan Stanley and Evercore for LBO modeling, buyer screening, and pitchbooks.
• Source-linked: Every figure traces back to its underlying data, from company databases to PitchBook filings.

 The new tool combines GPT-6 Astra's reasoning with premium financial data for research and modeling.
OpenAI unveiled ChatGPT for Financial Services, a specialized enterprise version powered by GPT-6 Astra combined with premium datasets from LSEG News, PitchBook, Daloopa, Crunchbase, and Quartr. It's shaped by design partners including Morgan Stanley and Evercore.
Teams can develop research, build financial models, and create customized client materials with firm-specific templates. OpenAI's Nick Turley said it teaches the AI to research and back conclusions like a real analyst, automating entry-level financial tasks.
Key Capabilities:

• Built-In Financial Data: Combines premium datasets from LSEG, PitchBook, Crunchbase, and Quartr directly into the tool.
• Analyst-Style Research: Trained to back conclusions with evidence the way a real financial analyst would.
• Enterprise Partners: Shaped by design partners including Morgan Stanley and Evercore for real workflows.

 Fugu Max and Ultra v2 route tasks across a pool of open models, cutting cost while boosting capability.
Sakana AI introduced Fugu Max and Fugu Ultra v2, the next evolution of their multi-agent orchestration system. Fugu Max dynamically routes tasks to the least capable model that can still handle them, delivering performance within striking distance of elite models at two to six times lower cost.
Fugu Ultra v2 pushes peak capability higher, outperforming Opus 5 and Fable 5 on Chartography, and beating models costing three to five times more per token on DeepSWE, all without using Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool.
Key Capabilities:

• Smart Routing: Dynamically sends tasks to the cheapest capable model instead of always using the biggest one.
• Major Cost Savings: Matches elite model performance at two to six times lower cost.
• Vendor Independence: Avoids reliance on any single frontier model, protecting against API cutoffs or lock-in.

 GPT-Live-1 brings full-duplex voice with natural interruptions and backchannels to third-party apps.
OpenAI launched GPT-Live-1 in its API, letting developers add ChatGPT's natural back-and-forth voice to their own apps and agents, complete with the ability to listen while speaking. It excels in noisy environments like cafés and supports new accents and languages.
It leads benchmarks with 80.1% on interactivity and 0.8-second latency, priced at $0.05 per minute for the voice layer. Builders are already creating poker coaches with live avatars, language tutors, and responsive hardware assistants.
Key Capabilities:

• Full-Duplex Voice: Listens while speaking, enabling natural interruptions and backchannels in conversation.
• Noise Resistant: Performs well in noisy environments like cafés without losing accuracy.
• Fast Response: Delivers 0.8-second latency, priced at $0.05 per minute for the voice layer.

 A group of AI agents produced a proof for the Navier-Stokes Millennium Prize Problem, OpenAI claims.
OpenAI announced it's sharing a solution to the Navier-Stokes Millennium Prize Problem, one of math's deepest unsolved challenges, produced by a group of agents using a next-generation model more capable than GPT-6 Astra. This would be a landmark achievement if independently verified.
However, mathematician Tristan Buckmaster has alleged the proof builds on his and Levent Alpöge's recent related work without proper attribution, a claim OpenAI denies. Given the dispute, this claim should be treated as unconfirmed pending independent mathematical review.
Key Capabilities:

• Major Claim: OpenAI says its agents solved one of seven Millennium Prize Problems in mathematics.
• Disputed Origin: A mathematician alleges the proof builds on his uncredited prior research, which OpenAI denies.
• Unverified Status: Not yet confirmed by independent mathematical peer review, so treat with caution.

 GPT-6 Astra (Medium) is temporarily open in Arena's Direct Mode with no subscription required.
Arena.ai opened GPT-6 Astra for free testing directly in Direct Mode. GPT-6 Astra (Medium) is available in Direct Mode for the next 24 hours, through September 10 at 8am PT.
It will remain accessible in Battle and Agent Mode after the Direct Mode window closes, as it already is today. This gives users a quick, no-cost way to try OpenAI's newest flagship model firsthand.
Key Capabilities:

• Limited Window: Free Direct Mode access lasts only 24 hours, ending September 10 at 8am PT.
• No Signup Needed: Lets users test the model directly without any subscription commitment.
• Still Accessible After: Remains available in Battle and Agent Mode even once the free window ends.

[10:08 AM]

ChatGPT Images 2.5 keeps details consistent across multiple edits instead of drifting each time.
OpenAI released ChatGPT Images 2.5, calling it faster, sharper, and smarter with better tools for creating whatever you can dream of. It offers improved fidelity for more natural, recognizable images and keeps details consistent across multiple rounds of edits.
A new comment-based editing feature lets users change only the specific part they want, without touching the rest of the image. Sam Altman noted it's not built for solving difficult math problems but should be genuinely enjoyable to use. It's also now available through third-party platforms like fal.
Key Capabilities:

• Faster Generation: Speeds up image creation to keep your creative ideas flowing without long waits.
• Comment-Based Edits: Change only what you specify, leaving the rest of the image untouched.
• Edit Consistency: Maintains consistent details and subjects across long, multi-step editing sessions.

It scans your Gmail, Slack, and documents to learn your specific phrases, sign-offs, and quirks.
OpenAI's ChatGPT Work can now pick up on what makes your writing sound like you, including favorite phrases, specific sign-offs, and capitalization habits. It connects to everyday tools like Gmail, Google Drive, Slack, and SharePoint to learn your personal writing style.
Setup takes about 30 seconds through a connection page, with the feature available on web for paid plans. Early users praise the quick setup and its help drafting Slack messages and documents, though some raise privacy questions around linking accounts.
Key Capabilities:

• Personal Style Match: Learns your specific phrases, sign-offs, and formatting habits from real writing.
• Multi-Tool Connection: Pulls context from Gmail, Google Drive, Slack, and SharePoint simultaneously.
• Fast Setup: Takes about 30 seconds to connect your tools and start learning your style.

 GPT-6 Astra can use a computer like a human, handling browsers, spreadsheets, and full workflows on its own.
OpenAI launched GPT-6 Astra, its most powerful model yet, calling it a generational leap in computer use, coding, cybersecurity, and scientific research. It's rolling out today to a limited set of organizations, expanding to all ChatGPT Plus, Pro, Business, and Enterprise users in the coming days.
Astra scores 98.6% on ARC-AGI-3, up from GPT-5.6 Sol's 7.8%, and hits 100% on ExploitGym, a tough cybersecurity benchmark. OpenAI president Greg Brockman closed the launch briefing by saying "Welcome to the AGI era."
Key Capabilities:

• Computer Use: Navigates browsers, spreadsheets, and desktop apps like a human, completing full workflows.
• Massive Benchmark Jump: Scores 98.6% on ARC-AGI-3, compared to just 7.8% for its predecessor.
• Critical Cybersecurity: Meets the Critical threshold in OpenAI's safety framework, rolling out cautiously to vetted partners first.

Astra is the first OpenAI model to officially cross the Critical cybersecurity threshold, with limited rollout planned.
OpenAI confirmed Astra is its first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework, meaning it can find and exploit previously unknown security flaws without human guidance. The model discovered two real zero-day vulnerabilities during testing.
OpenAI paused parts of Astra's development to strengthen safeguards before release. Advanced cyber capabilities will roll out first to a small group of testers through a program called Daybreak, expanding access to defenders over time.
Key Capabilities:

• Critical Rating: First OpenAI model to meet the highest cybersecurity risk tier in its safety framework.
• Real Discoveries: Found two genuine zero-day vulnerabilities during internal testing, unprompted.
• Limited Rollout: Advanced capabilities launch first to vetted testers through the Daybreak program.

Cursor's direct access to OpenAI models ends November 12, following its acquisition by SpaceX.
OpenAI is ending its partnership with Cursor, the coding tool now owned by SpaceX. Under the proposal, Cursor's direct model access shuts down on November 12, giving developers time to switch.
The decision reportedly stems from trust concerns tied to SpaceX's history with compliance. Cursor says OpenAI models make up only 5% of its traffic, softening the blow.
Key Capabilities:

• Hard Cutoff: Direct OpenAI model access for Cursor ends completely on November 12, 2026.
• Small Impact: OpenAI models reportedly account for just 5% of Cursor's total traffic today.
• Workarounds Exist: Developers can switch to personal API keys or other AI providers to keep working.