Claude Opus 4.8: one million tokens of context and a new Dreaming feature

Is it worth jumping from Opus 4.7 to 4.8 if you already use Claude for work? It is, if you rely on long sessions or agents that span multiple phases.

Anthropic shipped Opus 4.8 on May 28 as the newest model in its top-tier class. The release went straight to general availability through the Claude API, Amazon Bedrock and Google Vertex. Pro plan users get the model with no changes to their account.

This is the second Opus release in under two months. The previous 4.7 dropped on April 16. Three things separate the new build from the old one: context, output and agent behavior.

One million tokens of context becomes the default

The biggest technical change. The default context window (how much text the model can hold in “short-term memory” during a single session) is now one million tokens. A token is the basic processing unit, roughly four characters or half a word.

In practice, the model can now analyze a mid-size software project or a full book in one pass. Old Opus 4.7 worked with 200,000 tokens by default, with an optional bump to one million only on request.

Max output also climbs to 128,000 tokens. That is enough for detailed technical documentation or a long legal brief in a single response. Pricing stays in line with 4.7: five dollars per million input tokens and twenty-five per million output.

Dreaming, an agent that learns from its own history

Anthropic also unveiled a new feature for Claude Managed Agents called Dreaming. Per the company’s description, an agent can review its prior sessions, spot patterns in its own work and reshape memory for future tasks.

Dreaming lets agents review past sessions, discover patterns in their work and curate memory so behavior improves over time between individual runs.

Concrete example: an agent that triages business email each day can learn how the company tends to reply to certain inquiries, and adjust the next batch of replies to match without further instruction.

The other important agent-side addition is orchestrating multiple agents in the same session. A lead agent can delegate work to smaller specialized agents that run in parallel on a shared filesystem, each with its own model and tools.

System prompts mid-conversation

Another important change for developers. Until now the system prompt (the instruction defining model behavior in a session) could only be set at the start. In Opus 4.8 it can be swapped mid-stream without resetting state.

A scenario where this matters in practice:

• the agent first analyzes a user request as a planner
• then shifts into a code executor role
• then closes as a reviewer that checks the result

The old approach required three separate sessions or a pile of conditional logic in the prompt. Now a single session with three system contexts is enough.

What users in Serbia should know

Serbian payment cards work on claude.ai and through the API with no restrictions. The Pro plan runs twenty dollars per month (around 2,350 dinars) with expanded limits for Opus 4.8.

The Mac app ships for free. The web version runs in any modern browser on Windows and Linux, where there is still no official desktop app.

Bedrock and Vertex are available to any company in Serbia with an Amazon or Google Cloud account. Payment goes through a card or a business contract.

Is it enough to justify leaving Sonnet?

For most jobs Sonnet 4.6 stays the right pick on speed and price. Opus is around three times slower on a typical prompt and five times pricier per output token.

The gap shows up on complex tasks that involve large context, multiple phases, or tasks the older models could not solve on the first attempt. There Opus 4.8 genuinely pushes the line, especially for companies building internal AI agents.

Do you already use Opus for daily work, or only for the heavier projects where speed is less of a factor?