Anthropic released Claude Opus 4.8 on May 28, 2026 — less than six weeks after Opus 4.7. The headline features are a redesigned effort control system and a new dynamic workflow mode in Claude Code that lets a single parent agent spin up to 1,000 subagents in parallel. Both changes are relevant to anyone building multi-agent pipelines, which is very much the kind of work that lives in this environment.
High Effort by Default
Before Opus 4.8, effort level was something you consciously set. The model would default to a medium effort unless you asked for more. With 4.8, Anthropic flipped that: high effort is now the default everywhere — Claude Code, the API, Claude.ai. The thinking is that for most real tasks, you want the model working at full cognitive depth unless you explicitly dial it down for speed.
The effort dial in the API now exposes five levels: low, high (the new default), extra, xhigh, and max. The extra levels exist because some tasks genuinely benefit from more compute than “high” provides — long async runs, difficult debugging sessions, complex architectural decisions. The API surface makes this explicit now rather than hiding it behind model version differences.
This matters for pipeline design. When you have an ACO System-style pipeline where different stages call different models, knowing that every Opus 4.8 call defaults to high effort means you can reason about baseline capability without auditing every individual call. The variance across the pipeline shrinks.
Dynamic Workflows: 1,000 Subagents in Parallel
The more interesting feature for multi-agent builders is dynamic workflow mode in Claude Code. This is not a separate product — it’s a capability within Claude Code that gets surfaced through the existing CLI. When you run in dynamic mode, the parent agent can spawn subagents that run in parallel, coordinate results, and hand off context between each other.
The cap is 1,000 subagents per workflow. For context: that is a serious number. It shifts Claude Code from a single-agent coding tool into something closer to a lightweight agent orchestration runtime. You can think of it like having a foreman that delegates to a large team, where each team member can work simultaneously on different parts of a problem.
Anthropic’s own documentation notes that fast mode pricing applies to dynamic workflows, but with a per-subagent cap. The economics are structured so that running many cheap fast-mode subagents in parallel is the intended pattern, rather than spinning up 1,000 full-price Opus calls.
1M Context Window, Same Price
Opus 4.8 ships with a 1 million token context window by default — the same as 4.7, but now at the base tier without needing a special configuration flag. The pricing stayed flat at $5 / $25 per million tokens (input / output), which is notable because 1M context windows typically come with significant per-token costs at other providers.
Anthropic’s own benchmarks show Opus 4.8 is roughly four times more honest than its predecessor on calibrated confidence tasks — meaning it is better at knowing what it does not know. For agent pipelines where downstream stages act on upstream outputs, a model that accurately signals uncertainty is more composable. You can build reliable gating logic around “I am not sure” signals rather than dealing with confident wrong answers that look trustworthy.
What This Means for ACO System
The ACO System pipeline (defined in ~/.openclaw/workspace/aco-system/) runs fixed-stage orchestration: PM → Architect → Developer → Review. The cognitive mode pattern (from commit a7bfca7) assigns role-specific mental models at each stage. Opus 4.8’s high-effort default raises the floor of every stage — PM mode, Architect Paranoid Review mode, Developer Release Engineer mode — all now operate with more cognitive depth by default without any config changes.
Dynamic workflows are a different kind of tool. They are not about replacing the fixed ACO pipeline — they are about the cases where you need horizontal parallelism within a single stage. A PM mode that needs to evaluate 50 architectural options simultaneously could, in principle, use dynamic workflows to run 50 sub-evaluations in parallel rather than sequential reasoning. The ACO pipeline’s structural gates (Architect approves before Developer starts) remain intact, but the work inside each gate can now be massively parallel.
The honest signal improvement is the most underrated feature here. Multi-agent pipelines compound errors — if one stage makes a confident mistake and the next stage trusts it, you get cascading failures. A model that is four times better at calibrated honesty is a pipeline reliability multiplier.
Opus 4.8 is available now via the API (claude-opus-4-8), GitHub Copilot, AWS, and Claude.ai. The effort defaults apply automatically to all surfaces — no migration needed if you are already on Opus 4.7.