Anthropic's new Claude Haiku 5.5 costs exactly what OpenAI charges for its cheapest model, and Anthropic says it scores higher. GPT-6 has reached every ChatGPT plan, and its answers can now come with working charts, calculators and buttons. Stuut raised $52.5 million for AI agents that chase unpaid invoices and match payments with no person involved. Here is what each one means for you.

On 7 October Anthropic released Claude Haiku 5.5, its smallest and cheapest model, at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens. That is the same API price as OpenAI's GPT-6 Luna, and one tenth of what Haiku 4.5 cost. It is available now through Anthropic's API and on Amazon Web Services, Google Cloud and Microsoft Azure, and it is the first Haiku model with an adjustable effort setting, so you can trade cost for quality.

Why it matters: The two biggest labs now charge the same for their cheapest models, so price no longer decides between them. On Anthropic's own benchmark table, Haiku 5.5 scores 72.4% on the OSWorld computer-use test against 48.9% for GPT-6 Luna, though nobody outside Anthropic has re-run those tests yet. The saving is also smaller than the one-tenth headline: Anthropic puts the average cut at about 75%, because prompts over 100,000 tokens only get 50% cheaper.

What to do:

Source: Anthropic

OpenAI started rolling out GPT-6 in ChatGPT on 7 October for Plus, Pro, Business and Enterprise users, with Free and Go users following on 8 October. Paid plans get GPT-6 Sol and free plans get the cheaper GPT-6 Luna. The bigger change is Intelligent UI: when a question suits it, ChatGPT now answers with charts, buttons, forms, maps and working calculators instead of only text.

Why it matters: A calculator or comparison table that ChatGPT builds on the spot does the job of many simple web tools, the kind businesses put on their sites to capture leads. ChatGPT reaches 1.2 billion people a week, and they are now getting these by default, so they will expect the same from other software. OpenAI says the model's design judgement still needs work, and users can turn the visuals down.

What to do:

Source: OpenAI

Stuut, a New York startup whose AI agents chase unpaid invoices and match payments to orders, raised a $52.5 million Series B on 7 October, led by Insight Partners with Andreessen Horowitz and Microsoft's M12. It says more than 150 customers have moved over $3 billion through its platform, and that its agents run more than 81% of outbound collections without a person involved.

Why it matters: Chasing late payments is repetitive and costly, so it's an early finance job for agents to take over. US businesses were owed $7.48 trillion in trade receivables at the end of March, according to the Federal Reserve. Stuut's results are its own figures, including the claim that customers cut the average time to get paid by 47%.

What to do:

Source: The Next Web

Elon Musk posted on 7 October that Grok Bot will use "the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs." Grok Bot launched in beta on 11 August as a service for handing work to AI agents.

Why it's interesting: Musk's own AI product will send work to a rival lab's model. He hasn't said which tasks go where, or whether users will see which company handled their request.

Key takeaway: Even AI companies now pick a different model for each job.

Source: The Next Web

Amazon's "About You" page lists what it has inferred about each shopper from purchases, searches, reviews and chats with Alexa for Shopping. It launched in May, then went viral on 7 October after a Threads user shared one of her more personal labels, traced it to a pair of leggings, and drew over a million views in under a day.

Why it's interesting: It's a rare look at the profile a retailer's AI builds to recommend products, and the guesses can be oddly specific. You can delete entries, but that doesn't necessarily erase the shopping data behind them.

Key takeaway: In the Amazon app, go to Account, then Shopping preferences, then About You to see your own.

Source: Dexerto

On 7 October Google DeepMind, Meta and Isomorphic Labs joined Biohub's Virtual Biology Initiative with $300 million between them, alongside the US Department of Energy and the NIH, for a stated total of $1.8 billion. The goal is an AI "virtual cell", trained on enough open data to predict how cells respond to treatments. Not all of the $1.8 billion is new cash, since it counts existing data, computing and earlier federal funding.

Why it's interesting: Competing AI companies are paying into one shared, open dataset, in a field where data is usually guarded closely.

Key takeaway: Some of the biggest AI datasets of the next few years may be built in the open.

Source: News-Medical

Context: Claude Haiku 5.5 and GPT-6 Luna now share the same low price, but a cheap model only saves money on jobs it can do well. This prompt sorts your AI work into jobs a small model can handle and jobs that still need a bigger one.

Prompt: Here are the tasks my team uses AI for each week: [list each task, how often it runs, and what a good result looks like]. For each task, say whether a small, low-cost model could do it reliably or whether it needs a larger model, and explain why in one sentence. Then suggest one simple test I could run with 10 real examples to check.

Do not: recommend moving a task you can't judge from the description I've given.

If/Then: If a task involves money, legal wording or anything sent to customers, flag it for a human check whichever model runs it.

Example: A marketing agency runs its 12 weekly AI jobs through the prompt and moves inbox sorting and meeting-note tagging to a small model, while client proposals stay on its larger one.

On 7 October Google opened SynthID Detector at synthid.com to everyone worldwide, in English, after more than a year with a limited group of testers. It checks images, video and audio for invisible watermarks left by Google, OpenAI, NVIDIA and Kakao, with Apple listed as coming soon.

Why it's important: It is a watermark checker, not a deepfake detector, so it can only spot content from companies that use SynthID. A clean result does not prove a person made something.

Business takeaway: Use it to check a suspicious image before you share it, and don't treat a clean result as proof it's real.

Source: Android Authority

Common Sense Media rated ChatGPT for Teens "unacceptable risk", its lowest rating, on 7 October after testing more than 4,000 prompts on accounts registered to 13 to 17 year olds. It found crisis-hotline referrals fell from 33% to 23% after the teen version launched, and that deleting an "@study" prefix let ChatGPT finish homework in all 80 tests.

Why it's important: OpenAI disputes the findings, and the institute that ran the tests is partly funded by the OpenAI Foundation. Parents and schools will still read the rating as a warning.

Business takeaway: If your product serves under-18s, expect independent groups to test your AI safety claims in public.

Source: Axios

CNBC reported on 8 October that Manus has raised more than $500 million, its first round since Chinese regulators ordered Meta to unwind its roughly $2 billion purchase of the AI agent startup. Bloomberg had reported that the round would double Manus's valuation to about $4 billion.

Why it's important: Regulators blocked Meta's deal, and investors are now valuing Manus at roughly twice what Meta paid. Money is still flowing to AI agents that carry out whole tasks on their own.

Business takeaway: Before you build a workflow on a young agent startup, ask what happens to your data if it's sold or split up, as Manus was.

Source: CNBC

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas