Claude Opus 4.8 with Dynamic Workflows released
Key Questions
When was Opus 4.8 released and what is its key feature?
Released May 28, 2026, it includes Dynamic Workflows in research preview, enabling hundreds of parallel subagents within a single session.
What benchmark scores does Opus 4.8 achieve?
It scores 69.2% on SWE-bench Pro, 74.6% on agentic terminal coding, and leads a 12-agent comparison at 88.6% on SWE-bench Verified.
How does Dynamic Workflows address common agent failures?
It mitigates agentic laziness, goal drift, and self-preferential bias through isolation patterns and structured orchestration of sub-agents.
What improvements in honesty and calibration are reported?
Opus 4.8 is 4x less likely to pass flaws and shows better self-calibration compared to prior versions, addressing earlier bug reports on false completion claims.
What cost optimization patterns use Opus 4.8?
Multi-agent swarms assign Opus 4.8 to planning while routing execution to cheaper models; Ultra Code mode bundles high thinking with up to 100 parallel agents.
What community feedback exists on Notion's handling of Opus 4.8?
Notion temporarily disabled Opus 4.7/4.8 integrations due to degraded performance, confirming a recurring pattern of quality regressions.
What migration considerations apply to Opus 4.8?
Users should recalibrate effort settings, enable self-checking skills, and note fast mode caveats along with changes in Copilot billing.
How does Opus 4.8 compare to prior models in output speed?
Opper AI benchmarks show faster output at 67 tokens per second versus 57, with lower time-to-first-token of 7.9s compared to 22.6s in 4.7.
Anthropic released Opus 4.8 on May 28, 2026, featuring Dynamic Workflows (research preview) enabling hundreds of parallel subagents in a single session. Benchmarks: 69.2% SWE-bench Pro, 74.6% agentic terminal coding, beats GPT-5.5 and Gemini 3.1 Pro. Effort controls, cheaper fast mode, improved honesty (4x less likely to pass flaws). Community demos show full project generation. Opper AI comparison confirms improvements (Intelligence Index 61 vs 57, Coding Index 57 vs 53, faster output 67 vs 57 t/s, lower TTFT 7.9s vs 22.6s). Six Five Media benchmarks: agentic coding 64→69, knowledge work 1753→1890; effort control as cost guardrail. New settings guide covers agent optimization, auto mode, verification skills, scheduled tasks, bypass permissions. ECC project with 63 sub-agents claims 10× faster shipping. New tutorial: free Opus 4.8 access via AWS $200 credits and Kiro. Anthropic's 2026 roundup: 29 launches in 5 months, four-modes framing (Chat/Cowork/Code/Projects) includes Opus 4.8. Microsoft Build session on Claude Code + Cowork in Foundry (Opus 4.8, Dynamic Workflows, Auto Mode, evaluation-first). Dynamic Workflows demo with 8 parallel researchers, 16 fact-checkers, editor, designer – live agent office visualization. Opus 4.8 migration guide: effort recalibration, self-checking behavior, Copilot billing deadline, fast mode caveats. Claude Code + Hermes Agent integration into unified Agent OS dashboard, Bun rewrite example, faster search, security improvements. New routing rule video compares Opus 4.8 vs 4.7 with cost-aware perspective, highlighting 'annoying honesty' and token burn signals. New Dynamic Workflows tutorial by Ryan MacLean demonstrates repeatable agent orchestration, UltraCode scaling, and hidden docs trick. Russian-language Dynamic Workflows demo on real project finding a vulnerability. Anthropic engineer Thariq's firsthand guide on Dynamic Workflows – concrete example prompts, failure modes (agentic laziness, goal drift), and six common patterns (classify-and-act, fan-out, adversarial verification, etc.). Bun rewrite use case. Opus 4.8 + n8n MCP tutorial: autonomous pipeline building, self-correction behavior, shift from executor to specifier. Ethan Mollick shared Anthropic's agent teams vs workflows chart; community discussion highlights debuggability concerns when AI chooses orchestration, token cost visibility, and intermediate result access in regulated settings – important nuance for Dynamic Workflows tutorials. Ultra Code mode – a bundled setting combining extra high thinking and Dynamic Workflows, turning Claude into a team of up to 100 parallel agents. Comprehensive 12 AI coding agents comparison (Claude Opus 4.8 leads SWE-bench at 88.6%, parallel sub-agents key differentiator). 'Everyone treats Claude Code like one agent, but it can be an army' – beginner overview of sub-agents/Dynamic Workflows. Notion temporarily disabled Opus 4.7/4.8 due to degraded performance, confirming a repeating degradation pattern. Multi-agent swarm using Opus 4.8 for planning and cheaper models for execution – cost optimization pattern. Short tip list for running Opus autonomously with Claude Code. Dynamic Workflows deep dive by Thariq and Sid explaining three failure modes (agentic laziness, self-preferential bias, goal drift) and how isolation fixes them. Opus bug report detailing self-favoring skepticism, unstable calibration, false completion claims – high-signal community troubleshooting.