Anthropic Releases Claude Opus 4.8 With Parallel-Subagent Workflows
Anthropic shipped Claude Opus 4.8 on May 28, pushing SWE-bench Verified to 88.6% and adding Dynamic Workflows — a Claude Code preview that runs hundreds of parallel subagents in one session. Pricing is unchanged from Opus 4.7.
Anthropic released Claude Opus 4.8 today, less than two months after Opus 4.7. The headline is agentic coding: SWE-bench Verified climbs to 88.6% (from 87.6%), Terminal-Bench 2.1 hits 74.6%, and Anthropic’s internal agentic coding score jumps from 64.3% to 69.2%. On GPQA Diamond it scores 93.6%, and it leads the GDPval-AA economic-task benchmark at 1890 Elo.
The more interesting numbers are about reliability. Anthropic says Opus 4.8 is 4x less likely than Opus 4.7 to miss flaws in code it wrote itself — a direct answer to the failure mode where a model confidently ships its own bugs. Multidisciplinary reasoning with tools rose from 54.7% to 57.9%, and the model is the first to break 10% on the all-pass standard of the Legal Agent Benchmark.
Dynamic Workflows
The standout feature is Dynamic Workflows, in research preview inside Claude Code. Instead of one model grinding through a task linearly, Opus 4.8 plans the work, spins up hundreds of parallel subagents in a single session, and verifies their outputs before reporting back. Anthropic’s example: codebase migrations spanning hundreds of thousands of lines — the kind of mechanical-but-massive job that previously meant babysitting an agent for hours.
This is a real shift in how the model is meant to be used. The unit of work is no longer “a prompt” but “a goal decomposed across a fleet of workers,” with verification built into the loop rather than bolted on afterward.
Honesty and judgment
Anthropic leans hard on behavioral gains. The company describes Opus 4.8 as having “sharper judgement, more honesty about its progress, and the ability to work independently for longer than its predecessors.” In practice that means flagging uncertainty instead of claiming unsupported progress — the alignment team reports new highs on “prosocial traits like supporting user autonomy and acting in the user’s best interest.”
Pricing and speed
Pricing is unchanged from Opus 4.7: $5 per million input tokens, $25 per million output. Fast mode — which runs roughly 2.5x quicker — sits at $10/$50 per million and now costs about three times less than previous fast tiers.
On head-to-head agent benchmarks, Anthropic claims Opus 4.8 beats OpenAI’s GPT-5.5 at cost parity and edges Google’s Gemini 3.1 Pro. On the Super-Agent benchmark it was the only model to complete every case end-to-end. The competitive gap at the frontier is now measured in weeks, not quarters.