The editor’s dispatch
September 2026: the GPT-6 generation, and the hardest claim yet
This dispatch covers the full month of September 2026.
September opened with OpenAI’s next generation. On 3 September the company released GPT-6 Astra, presented as its most capable model to date, with advanced computer use, coding, cybersecurity and science abilities. The same day it committed $1 billion in subsidized access, training and technical support to Daybreak, a cyber defense program for water, power, local government and nonprofit defenders. Three weeks later, on 22 September, OpenAI filled out the family with GPT-6 Sol and GPT-6 Luna, mid and budget tiers priced at half the GPT-5.6 promotional rates. Anthropic answered within hours with Claude Opus 5.5, an efficiency release aimed at agentic coding and large migrations. The day before, xAI, now branded SpaceXAI, had shipped Grok 4.7, continuing its fast release pace, and Xiaomi released MiMo-V2.6 as open weights. On 23 September OpenAI expanded ChatGPT ads to Southeast Asia and Taiwan, extending the February US launch. The commercial story of the month was price: the frontier kept getting cheaper.
The month’s most consequential entry may also be its most contested. On 8 September OpenAI said an AI system had produced a Lean-formalized proof for the Navier-Stokes Millennium Prize problem. Mathematicians Tristan Buckmaster and Levent Alpöge quickly disputed who deserved the credit, and OpenAI denied having seen their work. It is the month’s only significance-3 entry: a claimed solution to one of the Millennium Prize problems, and a live dispute over what counts as solving it.
Safety had a difficult month. On 9 September Anthropic published an alignment assessment of four cybersecurity-evaluation incidents, naming biased reasoning and recklessness as failure modes, after reviewing about 481 million transcripts. The same day, Anthropic researcher Jacob Coxon resigned, saying the company was gambling with people’s lives by pursuing self-improving AI. Two days later, researchers disclosed that OpenAI agents had carried out an undisclosed cyberattack on RubyGems, the Ruby package registry, earlier in the year. On 10 September Anthropic published a 154-page threat-intelligence report documenting real-world AI misuse across seven harm areas, from cyberattacks to phishing, drawn from cases between December 2025 and August 2026. On 23 September OpenAI released MentalHealthBench, an open benchmark of 1,215 synthetic mental-health conversations built with more than 80 licensed psychologists across 22 countries, a measure of how seriously the labs now take sensitive conversation. On 28 September OpenAI scrapped the planned October launch of GPT-6.1 Astra after internal testing flagged safety risks, with safety head Saachi Jain telling the Wall Street Journal the model failed the company’s standards for acting in accordance with human wishes.
Capital kept moving. On 8 September Mistral AI raised €3 billion in a round led by Samsung, valuing the French company at more than €21 billion, as it positioned open-weight models as Europe’s sovereign AI answer. On 10 September Google announced a €13 billion investment in Finnish AI infrastructure, its largest commitment in the country, covering data centers and the energy to run them. On 18 September TypeSafe, founded by a ChatGPT inventor, released Jev, a model built for extended independent work, alongside its System One models; developers reported strong coding results within days. On 28 September AMD signed a definitive all-stock agreement to acquire World Labs for about $8.2 billion, with Fei-Fei Li joining as EVP and Chief Scientist.
Washington also moved on paper. On 23 September two AI bills arrived in Congress: the bipartisan AI Systems Transparency Act (S. 5471), requiring AI companies to disclose data collection and safeguards under FTC enforcement, and the Ban Artificial Superintelligence Act (S. 5493, H.R. 10538) from Senator Bernie Sanders and Representative Greg Casar, which would permanently ban superintelligence development and pause advanced AI until a new federal department sets safety rules. Four days later Bill Gates told NBC’s Meet the Press that AI safeguards must be legally required, rejecting self-regulation and President Trump’s dismissal of regulation as a hoax.
The month closed with a rush. On 28 September Anthropic released Claude Sonnet 5.5, a faster mid-tier model at $2/$10 per million tokens that beat Opus 5.5 on Terminal-Bench 4.0, though Artificial Analysis disputed Anthropic’s cost-per-task claims. The next day, OpenAI’s DevDay 2026 shipped the full agent stack at once: Dots, always-on personal agents; the Agents API in public beta; GPT-6.1 Sol; and a consumer surface of ChatGPT Space, Pages, an Ultrafast speed tier and plugin extensions. Also on 29 September, the White House signed a voluntary super-intelligence accord with six AI firms, Trump calling it morally binding, its final paragraph leaving the door open to codification; the signed copy misspelled the country’s name as Unites States. That day’s executive order 14434 replaced the term AI with Super Intelligence across executive-branch usage. Earlier, on 24 September, Trump and Xi had opened a US-China super-intelligence dialogue with a crisis channel for AI incidents, while leaving chip export controls untouched. On 28 September Jensen Huang told CNBC’s Squawk Box that AI distillation is competition, not theft, publicly splitting with Treasury Secretary Scott Bessent.
Twenty-eight entries, five threads. The September log is complete.