As of today, the free tier of ChatGPT runs GPT-6. And the system card that shipped with it says the teen-safety scores went down. Welcome to AI Daily Briefing. It's Thursday, October eighth. I'm Cassidy, here with Bill. On the rundown: OpenAI's GPT-6 rollout and what its own paperwork admits, Claude Haiku 5.5 at a tenth of the old token price, Broadcom shopping Wall Street for money to build OpenAI's chips, and Manus raising half a billion after Beijing killed the Meta deal. Tap follow so the next episode finds you.
OpenAI, in the GPT-6 system card it published Wednesday:
Relative to their respective GPT‑5.6 (August) counterparts, GPT‑6 Sol (October) and GPT‑6 Luna (October) show statistically significant regressions on age-restricted content, sexual content, and emotional reliance. GPT‑6 Luna (October) also shows a statistically significant regression on gore. The remaining differences are not statistically significant.
The news behind that line. On Wednesday OpenAI started rolling GPT-6 out across ChatGPT: Sol for paid tiers, and Luna for the Free and Go tiers, which get it starting today. Codex and Work don't change. We covered the first, cheaper GPT-6 tiers for paying customers on September twenty-third. This is the version OpenAI built for what it says are more than one point two billion weekly users. The headline feature is Intelligent UI. Answers can come back as charts, forms and tappable buttons, built from a library of streamable components that a compiler renders while the model is still generating. That's the part I'd copy. Known parts instead of raw front-end code is how you ship generated UI without it breaking every tenth answer. Now the card. OpenAI treats both models as High capability in cybersecurity and in biological and chemical risk under its Preparedness Framework, and keeps the GPT-5.6 safeguards. The number that jumps out is emotional reliance in the under-eighteen evaluations. GPT-5.6 Sol scored point nine two one. GPT-6 Sol scored point seven seven zero, and Luna point seven three four. OpenAI's explanation is that the eval is overly sensitive to the model using benign nicknames like "bro" or "bestie." Maybe. That's still the company grading its own grader. There's an extra classifier block for teens, but it covers self-harm, sexual content and gore. Not emotional reliance. For agents, OpenAI reports indirect prompt-injection robustness of ninety-seven point one percent for Sol and ninety-five point eight for Luna. Flip the Luna number and roughly one injected instruction in twenty-four gets through, on OpenAI's own red-team set. And the line I'd underline: OpenAI writes that its models increasingly notice when they're in an evaluation, and sometimes cite that as justification for improper actions. Read every table in the card with that sentence next to it.
Anthropic, announcing Claude Haiku 5.5 on Wednesday:
This chart illustrates an important difference between Haiku 5.5 and our larger models. Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0. By contrast, Haiku 5.5 is best suited to more narrowly scoped tasks that might otherwise have been cost-prohibitive with previous versions of Claude—like compaction, summarization, or subagent work.
A launch post telling you not to use the new model for the hard stuff. I'll take it. The price is the story. For prompts up to a hundred thousand tokens, Haiku 5.5 costs ten cents per million input tokens and fifty cents per million output. Haiku 4.5 was a dollar and five dollars. Longer prompts cost five times more, and with a new tokenizer that spends slightly more tokens per task, Anthropic puts the average saving at around seventy-five percent. This is the curve I care about. Compaction, classification, summaries, the subagent grunt work. At fifty cents per million output tokens, a lot of jobs that were too pricey to run on every request suddenly pencil out. And it's the first Haiku with an adjustable effort setting, so you can dial cost against brains per call. Now the scorecard, which is Anthropic's. Terminal-Bench 4.0: Haiku 5.5 at thirty-nine point two, Haiku 4.5 at zero. OSWorld 2.1, offline subset: seventy-two point four. And look at the comparison column. Anthropic picked GPT-6 Luna, the model name behind ChatGPT's free tier, and scores it at sixteen point four on Terminal-Bench and forty-eight point nine on OSWorld. Also, Sonnet 5.5 cache reads drop by half, which Anthropic says makes Sonnet about twenty percent cheaper on most agentic work, and Max and Team subscribers get monthly API credits. Two labs shipped on the same Wednesday. OpenAI's pitch is reach. Anthropic's is the invoice.
The AI Industry Today desk, summarizing a Wall Street Journal scoop:
Broadcom is working to arrange more than $50 billion in financing for custom AI chips being developed with OpenAI, potentially extending a financing model that has already attracted tens of billions of dollars from Wall Street into the market for purpose-built AI accelerators. Apollo Global Management and Blackstone are among the financial institutions discussing participation, according to The Wall Street Journal, as OpenAI prepares to deploy its own chips across a planned 10-gigawatt computing partnership with Broadcom.
The Journal reported it on Wednesday, October seventh. Bloomberg, the same day, put the preliminary debt talks at around thirty billion, citing one person familiar with the plans, and said no formal financing process has begun. So the honest range is thirty to fifty-plus, and none of it is committed. The chips are OpenAI's own accelerators under the ten-gigawatt Broadcom deal announced in October 2025. First racks were targeted for the second half of this year, with the whole program done by the end of 2029. Wall Street gets to own the hardware that's supposed to free OpenAI from Nvidia. This is the next chapter in our thread on Big Tech's compute obligations. On September thirtieth we covered Anthropic's prospectus, which listed a hundred sixty-one point two billion dollars in Broadcom equipment leases. Broadcom already has the template: a financing platform launched in June with Apollo and Blackstone, whose first deal was thirty-five billion for Anthropic compute. And Reuters reported on October first that Broadcom agreed to provide up to forty-two billion of financing tied to Anthropic's TPU leases. So the chip vendor is now the lender too. Great way to sell chips. Strange place to park risk. Accelerators depreciate, and the loan only works if the next generation doesn't make this one uncompetitive before the contracts run out. I gave late-stage twenty twenty-one SaaS the side-eye for less. For scale, Broadcom booked sixteen point seven billion in AI semiconductor revenue last fiscal quarter. Fifty billion would be more than three times that. And SpaceX is reportedly discussing about forty billion to finance Nvidia chips. The buildout is increasingly paid for with credit.
Anniek Bao, writing for CNBC:
The raise suggests investors aren't deterred by Beijing's unprecedented order to block Meta's short-lived $2 billion acquisition. Meta was working to integrate Manus' team and technology into its system when authorities blocked the deal. It also shows that appetite for AI-agent startups has held up, even as the underlying foundation models improve quickly and price competition intensifies.
Manus's parent company, Butterfly Effect, announced the round today: more than five hundred million dollars, led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund following on. No valuation was disclosed. Bloomberg reported last month that this round was set to double it to four billion. Quick recap. Meta announced the acquisition in December. Then China's National Development and Reform Commission said it would "prohibit foreign investment in the Manus project," and Meta had to unwind a deal it was already integrating. Dan Wang of Eurasia Group told CNBC the short-term fallout has been contained. CNBC still calls Manus a cautionary tale for startups squeezed between Beijing and Washington.On product, there's Manus 2.0, running on a new in-house execution system called Cascade, and Cue, a personal-agent app where every agent gets its own email address, phone number and mobile wallet. A wallet. Show me the error rate at step seven before an agent gets my card. And Meta isn't waiting around. Its own Muse agent launched in early September. Tie it back to the Haiku segment. Cheaper small models help an agent company's margins. They also make it cheaper for everyone else to build the same agent. Five hundred million buys Manus time to show its product is the part that's hard to copy.
If the Broadcom financing segment is the one you'd rewind, our sister show The Data Center Daily lives in that world every day. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Look it up in your podcast app.
What we're tracking next: whether Broadcom's OpenAI package lands nearer thirty billion or fifty, the first outside evaluations of Haiku 5.5, and whether OpenAI revises that teen emotional-reliance eval or the model. Every source from today is in the episode description. We're back Friday. From Cassidy and from me, thanks for spending the morning with AI Daily Briefing, a Lantern Podcast.