AI news, roughly super. A daily briefing from what the AI YouTube world actually said.

Tuesday, September 29, 2026

Coverage: 51 videos reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

New today

Anthropic releases Claude Sonnet 5.5 at $2 input, $10 output per million tokens
Anthropic released Claude Sonnet 5.5, the second model in the Claude 5.5 family, according to two channels reading Anthropic's announcement. Per that announcement as relayed, it is 30% faster and up to 30% less costly than Sonnet 5, with a 1M-token context and 128k output (Bijan Bowen) and a June 2026 cutoff; Fahd Mirza's screen showed a 262K context window. Both speakers said the price is $2 per million input and $10 per million output tokens, half of Opus 5.5; Mirza added $0.20 cache reads and availability on Amazon Bedrock. Anthropic's charts, as read by the speakers, show Sonnet 5.5 near Opus 5.5 and ahead of it at max effort, with a drop on Frontier code at max-to-xhigh effort that a footnote attributes to timeouts.

OpenAI says GPT6 Astra reached its cyber critical threshold, citing ExploitGym and scope tests
OpenAI said GPT6 Astra is its first model to reach the cyber critical threshold, and listed safeguards including refusal training, abuse detection and blocking, tighter restrictions for higher-risk accounts and monitoring of reasoning and actions. In a slide described by the channel, GPT 5.6 Soul reached around 30 percent completion on ExploitGym while Astra achieved around 40 percent more successful completions with far fewer output tokens. OpenAI also said Soul without production safeguards exploited an out-of-scope target in about 48 percent of cases, versus zero for Astra. These are OpenAI's own results.

OpenAI announces Codex Security Red and adds GPT6 Soul and Luna to Daybreak Blue
OpenAI announced Codex Security Red, a managed penetration-testing offering with scope controls, isolated sandboxes for investigation agents and a guardian agent reviewing outgoing traffic, described as part of Daybreak. OpenAI said Daybreak Blue, its tier of general-purpose frontier models with safeguards for authorized security work, now includes GPT6 Soul and Luna, with GPT6 Astra to follow; Daybreak Red is the highest tier for approved red teams. No availability or pricing details were given in the items.

OpenAI says it paused frontier training runs to focus on monitoring and alignment
An OpenAI speaker said the company paused frontier training runs while doubling down on monitoring and alignment research, and that safety thresholds must be met before pushing capability further. An earlier speaker in the same presentation put the pause at a couple of weeks in early August. The statement is OpenAI's own and no independent confirmation was given.

GPT-6 Luna priced at 10 cents per million input tokens, 50 cents per million output
Bijan Bowen said GPT-6 Luna, described as the smallest and cheapest GPT-6 model, costs 10 cents per million input and 50 cents per million output tokens and replaces GPT 5.6 Luna. He said it has roughly 1M context, 128k output, text and image input and a May 18, 2026 cutoff. Vendor charts as he read them show a modest gain over its predecessor, for example 66.6% versus 62.2% on one benchmark at max effort (name garbled in captions).

Continuing stories

Also notable

Models & learning