앤트로픽, 오푸스 4.8 출시…“수주일 내 미토스 공개”
Source: Byline Network
Announcement
Anthropic has released the latest version of its Frontier model, Opus 4.8, just one month after launching Opus 4.7. The company also announced that the most powerful model, Mitos, will be publicly available within a few weeks.
Performance Improvements
Opus 4.8 shows a substantial jump in performance over previous versions across coding, active reasoning, inference, and practical‑knowledge tasks. Early testers reported that the model is more stable and demonstrates superior judgment when performing agent‑based work.
Integrity Enhancements
Anthropic says Opus 4.8 has been hardened to reduce unfounded claims. The model now better signals uncertainty and makes fewer baseless assertions. Internal evaluations indicate that Opus 4.8 is about four times less likely to miss defects in generated code compared with earlier releases.
Claude Code Dynamic Workflow
Anthropic also unveiled a dynamic workflow feature for Claude Code:
- Allows Claude Code to handle much larger tasks.
- After planning, a single session can launch hundreds of sub‑agents.
- With Opus 4.8, agent execution times are longer, enabling Claude Code to process tens of thousands of lines of code from planning through merge in one go.
- This capability is offered as a research preview and is available on the Claude Code Enterprise, Team, and Max plans.
Effort‑Control Feature
Both Claude.ai and Co‑Work now include effort‑control options, letting users choose how much computational effort the model should expend on a response.
Message API Updates
- The Message API now permits a system item within a message array.
- Developers can update Claude instructions mid‑task without corrupting prompt caches or requiring an extra user turn.
- During agent execution, permissions, token budgets, and environment context can be modified on the fly.
Pricing
Opus 4.8 is offered at the same rates as the previous version:
- Standard mode: $5 / 1 M input tokens, $25 / 1 M output tokens.
- High‑speed mode: 3× cheaper than earlier models – $10 / 1 M input tokens, $50 / 1 M output tokens.
Future Plans
While Opus 4.8’s cybersecurity performance trails the Claude Mitos preview, Anthropic notes that a Mitos‑level model will require stronger security safeguards before a general release. The company expects to make a Mitos‑class model available to the public within the next few weeks.