본문 바로가기
News · 2026.10.09

Push repetitive work such as summarizing and classification down to Haiku 5.5 — at about 21 cents per task, it cuts cost to roughly a quarter of Sonnet's

AI News: Haiku-5.5, Visual ChatGPT, Grok's New Trick, and more! — Matt Wolfe · Read the original ↗

This page is a translation of the Korean original. Read the Korean original →

What you get

  • Sharply lowers the cost of high-volume repetitive work and coding subagents
  • ChatGPT answers turn visual, with charts, buttons and forms, so you can check them faster
  • Dashboards connect directly to data platforms and CRMs and stay current without separate plugins

When to use it

The recommendation is to choose Haiku 5.5 for high-volume repetitive work such as summarizing, compression, database queries and classification, and as a subagent in coding tasks. The speed and cost advantages are larger when running basic tasks through the API or Claude Code than in a flat-rate chat subscription.

How to do it

  1. In your Claude account, select Haiku 5.5 as the model and set the effort level to medium, the recommended value.
  2. Connect a data platform (BigQuery, Databricks, Snowflake) or a CRM such as Salesforce to build a live dashboard that stays up to date (available on paid plans; Motion is in beta for Team/Enterprise).
  3. In ChatGPT, turn on GPT-6 and set thinking effort to instant to get answers that mix images and step-by-step visuals with text.
  4. In the permissions dropdown of Codex/ChatGPT work, choose approve for me mode to cut down approval requests (free for signed-in users).
  5. Developers can use the Decisions API, which returns only three outputs — predicates, choices and scores — to build fast, cheap judgment calls.

What it means for my work

If you run agents, now is the time to redraw your model-routing rules. Move low-difficulty steps such as summarizing, classification and database lookups to a cheaper model. For decision nodes that only need to pick one option or return a true/false probability, replace long text generation with structured decision output. Cost and latency both fall.

Background — what the original covers

In a single week, Anthropic shipped a low-cost, high-speed model plus dashboard and animation features, while OpenAI announced 28 consecutive days of one launch a day and released GPT-6, a visual UI, a judgment-only API and speed improvements. Grok began using outside models as its backend and reading X directly, and Mistral and Reflection released large open-weight models. The video argues that the agent products of many companies are converging on doing the same job.

#Tools

← All practitioner tips