OpenAI says GPT-6 scored worse on self-harm safety tests

GPT-6 safety card: OpenAI’s safety card for the GPT-6 models now in ChatGPT, published Wednesday, shows lower safe-response rates on a deliberately hard set of self-harm prompts. GPT-6 Sol had a safe-response rate of 0.896, down from 0.934 for GPT-5.6 Sol; GPT-6 Luna scored 0.901, down from 0.932 for GPT-5.6 Luna, the card says. OpenAI called the drops statistically significant and said the failures it reviewed were “borderline but still generally safe.” OpenAI ran the tests without ChatGPT’s system-level safeguards and said error rates are not representative of average production traffic.

If you or someone you know is struggling, call or text 988, the Suicide & Crisis Lifeline (U.S.).

Claude Haiku 5.5: Anthropic released Claude Haiku 5.5 on Wednesday and listed it at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. The company said the model costs around 75% less to run on average than Haiku 4.5. That is Anthropic’s average running cost, not the cut on every call; above 100,000 tokens the price is 50% lower, Anthropic said.

Samsung guidance: Samsung Electronics estimated third-quarter operating profit at about 107.4 trillion won in preliminary guidance released Thursday. That compares with 12.17 trillion won a year earlier and 89.49 trillion won in the second quarter, Samsung said. Sales were estimated at about 195 trillion won, up from 86.06 trillion won a year earlier. The release gives no breakdown by business.

TSMC revenue: TSMC said Thursday its September revenue was about NT$511.86 billion, up 54.6% from a year earlier and down 0.6% from August. Its July, August and September reports add up to about NT$1.494 trillion for the third quarter, the highest quarterly total in TSMC’s monthly reports. The monthly figures are unaudited. TSMC reports quarterly results Oct. 15.

Anthropic usage policy: Anthropic will bar “sustained and needless abusive or cruel behavior toward our models” starting Nov. 12, the company said Thursday. It said the rule applies only in extreme cases where users repeatedly act cruelly “with no discernible purpose,” not to common user frustration, pushback, dark creative themes or model testing and research. Claude’s ability to end such chats will remain the primary enforcement mechanism, Anthropic said.

Pentagon Tradewinds: The Pentagon’s Tradewinds program lets companies pitch AI products in videos no longer than five minutes, Wired reported Wednesday. Accepted pitches get “post-competitive” status, which Wired said could help shave months off award timelines; a Defense Department official told Wired some awards came in less than a week. This year the military is seeking pitches for AI that can “enhance force lethality” and agents for joint targeting, including so-called kill chain execution, Wired reported.

Finland data centers: The Finnish Supervisory Agency (LVV) on Tuesday told Google subsidiary Tuike Finland Oy to suspend preparatory work that would significantly alter the environment at two planned data centers in Muhos and Kajaani until environmental impact assessments are finished, CNBC reported. The agency wants tree clearing, excavation and road building stopped by Oct. 23 and is seeking voluntary compliance first, TechTarget reported. “We understand the concerns and have fallen short of our own high standards in this instance,” a Google spokesperson told CNBC.