Anthropic (ANTH.PVT) on Thursday released its latest flagship AI model, Claude Opus 4.8. The new model, an upgrade over the company’s prior leading offering Opus 4.7, comes as Anthropic and archrival OpenAI (OPAI.PVTl) race to make their public market debuts later this year.
According to Anthropic, Opus 4.8 tops OpenAI’s GPT-5.5 as well as Google’s (GOOG, GOOGL) Gemini 3.1 Pro in a number of synthetic AI-focused benchmarks. That includes things like agentic coding, agentic financial analysis, and agentic computer use.
Agentic capabilities are particularly important as more companies begin experimenting with and deploying AI agents across their businesses.
Think of AI agents as semi- and fully autonomous digital helpers that can perform tasks on your behalf.
Anthropic says Opus 4.8 is also more “honest” than its prior models, meaning it’s less likely to present false information as fact-based on little supporting information. Instead, the company claims, Opus 4.8 will let you know if it sees “uncertainties about its work” and will be less likely to provide users with “unsupported claims.”
In addition to Opus 4.8, Anthropic says it’s also debuting capabilities for its broader product portfolio. That includes what it calls dynamic workflows, which the company says can set hundreds of small subagents to work and deliver results.
You’ll also now be able to control how much effort Claude puts into performing tasks, helping to cut down on the amount of time or money you spend on certain capabilities.
For example, you’ll be able to tune Claude in both your browser and in Claude Cowork to put more effort into operations you want the AI to noodle on longer and less effort into things you want done quickly.
Using less effort will also cut down on the amount of tokens a user burns through as they have back-and-forth chats with AI chatbots and services.
Tokens are the basic units AI services use to measure inputs and outputs. So each time you ask Claude, or ChatGPT, or Gemini a question, whether that’s how long turkey chili lasts in the fridge or to code a basic web page, you’re using tokens.
Each token represents portions of media, whether that’s text, audio, images, or video. For instance, the prior sentence is approximately 33 tokens based on a sample tokenizer created using Claude.
According to Anthropic, Claude Opus 4.8 is available today and is the same price for regular users as Opus 4.7 — $5 per million input tokens and $25 per million output tokens. Fast mode will cost $10 per million input tokens and $50 per million output tokens.
