AI agents are AI tools that do a job for you, step by step, instead of only answering a question. Here are four stories from the last few days. I opened every source myself. When a number comes from the company selling the product, I say so.
Anthropic launches Claude Haiku 5.5, a much cheaper small model
Anthropic published Claude Haiku 5.5 on 7 October. It says the model is available now on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.
The price is the headline. Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. A token is a small piece of text, about part of a word. Anthropic lists the older Haiku 4.5 at $1.00 for input, and says the average cost is about 75% lower.
On OSWorld 2.1, a test of how well an AI can use a computer, Anthropic reports 72.4% for Haiku 5.5 against 15.7% for Haiku 4.5. Anthropic also says this is its first Haiku-class model with an adjustable effort setting, so you can pick lower cost or higher intelligence.
There are two more changes. Cache reads for Sonnet 5.5 drop 50%, from $0.20 to $0.10 per million tokens, which Anthropic says cuts costs by about 20% on most agentic tasks. And Max and Team subscribers now get monthly API credits: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team.
My take: every number here comes from Anthropic, which makes the model and sells it. I read one source, Anthropic's own page, and no outside group had tested it in what I read. Anthropic says it is also available on AWS, Google Cloud and Microsoft Azure, and those companies sell it too, so they gain from its success.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
Grok Bot will now send tasks to rival models
Gizmodo reported on 7 October that SpaceX's Grok Bot agent service will now route requests to other companies' models. The news comes from a post by Elon Musk, so this is his claim.
Musk named Anthropic's Claude Opus 5.5, Midjourney for images and Suno for audio. He said Grok Bot will pick "whatever is most likely to give you the best outcome." He also said simple questions go to small, fast models, and complex ones go to large models.
Some things are not clear. Musk did not give the full list of models, and did not confirm whether Grok 4.7 is one of the options. Gizmodo says the shift may mean SpaceX is moving toward AI infrastructure, but that is the writer's view, not a company statement.
My take: I read one source, Gizmodo, and it leans on Musk's own posts. Claude is one of the named models, so this story touches Anthropic. I would wait for a clear list of models and prices before judging it.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
HackerRank's AI interviewer Chakra is now generally available
RuntimeWire reported on 5 October that HackerRank made Chakra generally available, after about six months in beta. Chakra is an AI agent that runs a coding interview.
Candidates work on a task from a real code repository, with an optional AI assistant. Chakra asks follow-up questions about their choices and how they judged the AI's output. HackerRank says it looks at problem-solving, judgment, communication and "AI fluency". Beta testers named include Snowflake, Snorkel and Capgemini.
CEO Vivek Ravisankar told TechCrunch: "Now, because of AI, anybody can produce an artifact." His point is that output alone says less about a person than it used to.
Hiring teams get scores, feedback and evidence. HackerRank's candidate guide says Chakra does not make the hiring decision; the employer does. Its notice also says the AI may look at coding, behavioural signals, webcam images and screen activity.
My take: the 500,000 figure, the testers and the fairness claims all come from HackerRank. RuntimeWire says the launch material does not show independent checks against real hiring results. It also does not say whether 500,000 counts people or sessions. If you are a candidate, read the data notice first.
Realtor.com adds RealAssist AI for real estate agents
RISMedia reported on 6 October that Realtor.com launched RealAssist AI inside Realtor.com+, its workspace for real estate agents.
It does three things. It guides agents through adding clients and their search preferences. It sends a daily digest of a client's chats, searches and price changes, with suggested next steps. And its "Agentic Actions" help with tasks such as refining searches or drafting client messages.
The article describes agents as "always in the driver's seat, reviewing and approving every action." It says Realtor.com+ is live in 21 markets, with four more coming, and serves more than 192,000 agents.
My take: I like that a person approves each step, because it is a safe way to start with agents. But I read one article, and it repeats Realtor.com's own description. It does not say who built the AI inside, so I will not guess.
What I take from this week
- Small, cheap models are becoming a bigger part of how agents are built, but the test results so far are the makers' own.
- Agents are starting to hand work to each other's models, so who does the work behind the scenes is harder to see.
- The safest agents so far keep a person in the loop, and the companies still need to show outside proof.
I will test these tools on real tasks with my five checks when I can use them.