Opus 5.5: Anthropic lowers the cost of working with its larger Claude
The larger Claude has new rates: $4 per million input tokens and $20 for output. Let’s look at what the claimed savings mean for longer jobs.
Loading page…
Model releases and practical articles about research, writing, documents and checking answers. Explore news, guides and tutorials for everyday work.
Start with Opus, Sonnet or the DevDay overview. Then try a writing or research task with material you know well.
The larger Claude has new rates: $4 per million input tokens and $20 for output. Let’s look at what the claimed savings mean for longer jobs.
Anthropic says Sonnet 5.5 generates responses more than 30 percent faster. Token rates stay unchanged; the claimed savings come from a shorter route to the result.
Close your laptop and let Codex keep working. Add a cheaper Sol and Dots for recurring tasks, and there is plenty to unpack from DevDay. Here is where we would start.
The assistant quoted a ticket price. The trouble is, the source had no price list. We have twelve such situations: see how JEV, GPT-6 Luna and you judge them.
“Make it sound human” can lead to another batch of adjectives. We show what to give the model so it improves the text without making promises on your behalf.
The model shortened your description and added a service you do not offer? Let’s build an instruction and check its answer using one short, concrete example.
Claude Code, ChatGPT, Codex, skills, plugins, MCP… If you want to try AI and are already fed up with all these names, let’s start with one ordinary task. Open a chat and turn a short note into something useful.
One model answers; another judges it. Sounds sensible, until we ask who checks the judge. We take a look at JEV and share an exercise of our own.
A large context window sounds impressive. What can you do with it? We suggest three tasks for Astra: untangling documents, an editable budget and a search feature you can run.
Which model should you choose? Give both the same proposal and see what they do with the price, deadline and missing information. Time spent correcting them counts too.
We put documents into NotebookLM and ask questions. But what if one describes a new feature and another an old limit? Let’s sort the sources before producing a tidy summary.
An answer has three links, so it looks trustworthy. Let’s open them. They may all lead back to one announcement, while the most interesting sentence has no support.
You upload the current policy, but the chat still gives the old terms. Before changing the model, let’s inspect the passages it actually received.
A good result, a saved prompt, a new conversation — and something goes wrong. The important agreements may have stayed in the old chat. Here is how to bring them along.
Someone gathers sources, someone writes from them, someone revisits the suspicious sentence. Let’s divide the work into those three roles, even if one person takes all of them.
You message an assistant in a chat app, and it uses connected tools. That was the idea behind Clawdbot, now developed as OpenClaw. We explain it and correct a common shortcut about local data.
Instead of “press one,” a conversation about an available appointment. We look at what voice agents need to book a visit and how to tell whether they actually help.
A folder full of material and a report to prepare. Cowork lets you delegate file work in ordinary language. Here is how to define the task and check what comes back.
Sales fell. But why? AI can help inspect data and find a lead, but first we need to check the numbers. Start with one question, rather than the promise of an automatic report.