Anthropic released Claude Opus 4.8 on May 28, 2026, pairing modest model improvements with dynamic workflows that can spawn hundreds of parallel sub‑agents and a user-facing effort control to tune response depth and token use.
Anthropic released Claude Opus 4.8 on May 28, 2026, calling the update a "modest but tangible" improvement and packaging it with platform features aimed at larger, more agentic workloads. The launch centers on orchestration and usability — dynamic workflows and a user-facing effort control — rather than positioning Opus 4.8 as a pure raw-model leap. That combination matters because it is intended to change how teams distribute large engineering tasks and manage cost versus result depth.
Anthropic published benchmark results for Opus 4.8 that show measurable gains on coding and multidisciplinary tests. Opus 4.8 scores 69.2% on agentic coding (SWE‑Bench Pro), up from 64.3% for Opus 4.7 and ahead of GPT‑5.5 at 58.6%. On the multidisciplinary Humanity's Last Exam, Opus 4.8 posts 49.8% without tools and 57.9% with tools. The company also reports the model is more likely to flag uncertainties and is about four times less likely to leave coding bugs unremarked compared with Opus 4.7.
Anthropic frames Opus 4.8 as outperforming its prior release and several competitors across most tested categories, and says it scores higher on prosocial traits such as support for user autonomy. The company refers readers to the Claude Opus 4.8 System Card for fuller safety and alignment metrics and tests.
The most operationally significant change is dynamic workflows: the platform can plan tasks and spawn hundreds of parallel sub‑agents within a single session. Anthropic says Claude Code running Opus 4.8 can execute codebase‑wide migrations across hundreds of thousands of lines, carrying work from planning through merge. That capability is positioned to let teams parallelize large refactors and distributed coding jobs inside one orchestrated session.
Anthropic also introduced an effort control exposed next to the model picker on claude.ai and in Cowork. Users can set a default of 'high' or ramp to 'extra', 'xhigh' and 'max' to request deeper, more token‑intensive responses; the control is presented as a direct tradeoff between answer depth, latency and token cost.
Pricing for Opus 4.8’s API remains the same as Opus 4.7: $5 per million input tokens and $25 per million output tokens. Fast Mode runs Opus 4.8 at 2.5× speed and is priced at $10 per million input and $50 per million output — about one‑third the prior fast‑mode price. Anthropic says higher rate limits for Claude Code customers will help offset heavier token use, and reporting cites outside analysis suggesting Opus 4.8 could mitigate the practical cost increase seen when 4.7 replaced 4.6.
Sources
Replies (0)
No replies in this topic yet.
