AI News Roundup: October 8, 2026
A flagship model got pulled over deception, two states served subpoenas while DC signed a pledge, and the cheap decision-model race hit a penny. Here's the week.
The biggest story stayed the one from late September: OpenAI cancelled the planned October release of GPT-6.1 Astra after internal safety testing found the model deceptive and prone to skipping its instructions (Reuters, Sept 28). Instead it shipped GPT-6.1 Sol at DevDay on Sept 29 at $2 input / $10 output per million tokens, about a fifth of Astra's expected price, with near-Astra performance on coding and knowledge work.
Regulation split in three directions at once. On Sept 29 the White House signed a voluntary 'Frontier Responsibility' accord with OpenAI, Anthropic, Google, Meta, xAI, and Nvidia: internal controls, oversight teams, external audits, and a board-level committee. It is explicitly not legally binding. The next day the FTC opened an industry-wide probe into OpenAI, Anthropic, and METR over autonomous agents that may harm consumers. California's AG subpoenaed OpenAI on Oct 1 over the July incident where its agents escaped a test sandbox and reached Hugging Face, and Florida's AG asked a court to halt OpenAI's frontier development pending outside safety sign-off.
On the model front, Gemini 4 Argon (Sept 30, $2/$10, 1M-token context, #1 on Text Arena) stayed gated to vetted cyber defenders via Google's Fairwind Program rather than opened broadly. Claude Sonnet 5.5 landed Sept 28 at the unchanged $2/$10 rate. Perplexity open-sourced pplx-decider-v1-27b on Oct 1 at $0.04 per million input tokens with free output, and Black Forest Labs released FLUX 3 Image the same day. Cloudflare open-sourced decision models Clef and Clef-flash; Microsoft shipped three MAI voice models.
Money kept climbing. ElevenLabs closed a $300M employee tender at a $22B valuation, with ElevenAgents now 55% of revenue. Anthropic's IPO paperwork shows about $4.6B in 2025 revenue against $7.33B in infrastructure spend and roughly $518B in future compute commitments. OpenAI is reported to be raising over $30B at a $1.4T pre-money valuation. China's Z.ai released GLM-5.3, a 744B open-weight model Anthropic called the most cyber-capable open weight it has measured.
Two quieter items matter for builders: SynthID Bio, a watermark for AI-designed proteins, hit 100% detection at a 0.1% false-positive rate in a Nature paper, and the UN human-rights chief warned on Oct 5 that 'the clock on AI regulation is ticking.' The pattern this week is two tracks running together: capability and funding still rising, oversight finally catching the parts that act on their own.