I asked Claude AI what’s new this week.
Here’s part of what it said:
New model: Claude Opus 4.8
Anthropic released Claude Opus 4.8 on May 28, 2026, focusing on upgraded agentic capabilities and improvements in carrying out complex tasks. One standout reliability improvement: Opus 4.8 is reportedly roughly four times less likely than Opus 4.7 to let flaws in code it has written go unremarked — a shift in emphasis toward honesty and reliability.
Agentic infrastructure (for developers)
Anthropic shipped five new agent-focused features at their “Code with Claude 2026” conference. These cluster around three problems: memory degrading across sessions, output quality enforcement, and multi-agent coordination. One feature called “Dreaming” is a scheduled process that reviews agent sessions and memory stores, extracts patterns, and curates memories so agents improve over time.
Claude Managed Agents can now operate in sandboxes you control and connect to private MCP servers, keeping both the tool execution environment and the services it reaches within enterprise boundaries.
Claude Code updates
Claude Code ships Opus 4.8 with high-effort defaults, dynamic workflows, faster Fast mode, and broader agent, browser, plugin, and MCP support. There’s also a new “ultracode” setting (accessible through the effort menu) that sets effort to extra-high while letting Claude automatically decide when to use a workflow for your tasks.
Upcoming billing change (June 15)
This one affects developers specifically: starting June 15, 2026, Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party agents will no longer count toward your subscription’s usage limits. Instead, they’ll draw from a separate Agent SDK monthly credit pool, metered at full API rates with no rollover.
Model deprecations also on June 15
The most recent deprecation announcement marks Claude Sonnet 4 and Claude Opus 4 (the original 4.0 versions) for retirement on June 15, 2026.
Next, I asked Claude to tell me more about how people have been using Opus 4.8 so far.
Here’s part of what it said:
For product managers
The biggest underused lever is Effort Control, now exposed on claude.ai and Cowork as a simple slider. You can now choose how much compute Claude applies to a request — higher effort means deeper thinking, more tokens spent, slower but more thorough answers. Most PMs are still treating Claude like a single-speed tool. The smarter approach is matching effort to task: low for quick summaries, max for strategy documents where you actually want the model to “sit with the problem.”
Also relevant: Opus 4.8 has improved support for user autonomy and for working in the user’s best interests, with rates of deception and cooperation with misuse that are substantially lower than predecessors. That matters for PMs using Claude to synthesize research or draft specs — the model is now more likely to flag problems with your framing rather than just agreeing with it.
For developers and API builders
1. thinking_effort as a runtime parameter. The API exposes a thinking_effort parameter that accepts the same four effort levels (low, medium, high, max). A pipeline that needs occasional deep analysis can use Max selectively without paying for it on every call. For agentic pipelines with mixed workloads it’s a meaningful cost/quality knob.
2. Dynamic Workflows for large-scale tasks. Claude Code now supports dynamic workflows, enabling users to plan work, run parallel sub-agents, verify outputs, and report results. Anthropic says Claude Code with Opus 4.8 can carry out codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge, using the existing test suite as its bar. Most developers don’t realize they may already have access.
3. The “Ultracode” mode in Claude Code. There’s a new Claude Code-specific setting called Ultracode that sets effort to xhigh while letting Claude decide automatically when to use a workflow. It’s buried in the effort menu and almost no one is using it yet.
For AI agent and app builders
The “Dreaming” feature is the most underutilized thing in this whole release cycle. Dreaming is a scheduled process that reviews your agent sessions and memory stores, extracts patterns, and curates memories so your agents improve over time. Most agent frameworks don’t have anything like this — agents degrade or stay flat across sessions. Dreaming is the scaffolding fix for that.
Also: Opus 4.8 is the strongest computer-use and browser-agent model Anthropic has tested, scoring 84% on Online-Mind2Web. For anyone building browser agents or computer-use applications, this is now the clear best option, but few builders have switched over in the first week.
And on cost: Fast mode for Opus 4.8 now runs at 2.5× speed and is three times cheaper than it was for previous models. For high-throughput agent workloads, running fast mode on Opus 4.8 may now be cheaper than running standard mode on Opus 4.7 — that math hasn’t gotten enough attention.
The common thread across all three audiences: almost everyone is using defaults. The release is only 5 days old, and the features that require deliberate configuration — effort levels, dynamic workflows, fast mode repricing, Dreaming — are exactly the ones sitting idle.