Claude Opus 4.8 Arrives with Higher Honesty and Record‑Breaking Valuation

Anthropic released Claude Opus 4.8 on Friday morning, building on Opus 4.7 with sharper judgment, higher self‑reported honesty, and the ability to work autonomously for longer periods. API pricing remains the same (input $5 / M tokens, output $25 / M tokens), while Fast Mode costs only one‑third of the regular rate (input $10 / M, output $50 / M).

Benchmark tables show Opus 4.8 outperforming its predecessor and other competitors, including GPT‑5.5, across programming, agent capabilities, logical reasoning, and knowledge‑intensive tasks. Early testers report more reliable outputs, sharper judgment, and apparently lower token consumption.

The model incorporates “honesty training” that discourages unsupported claims. Anthropic’s internal evaluation indicates the probability of the model generating code with undisclosed defects drops to about one‑quarter of Opus 4.7. Alignment tests reveal new highs in prosocial traits and a markedly lower rate of unaligned behavior, matching the best‑performing Claude Mythos Preview.

A major new feature is Dynamic Workflows, currently in research preview. In Claude Code, the model can plan its own workflow and launch hundreds of parallel sub‑agents in a single session. This enables large‑scale tasks such as rewriting the Bun JavaScript runtime from Zig to Rust—achieving 99.8% test‑suite pass, 750 k lines of Rust code, and completion in 11 days.

Claude.ai and Cowork now expose an “Effort Control” setting. High‑Effort mode makes the model think more deeply and produce higher‑quality responses at the cost of speed and token usage, while Low‑Effort mode is faster and consumes fewer rate‑limit tokens. The default for Opus 4.8 is High‑Effort, which consumes a token amount comparable to Opus 4.7 but delivers better performance.

The Messages API now accepts System Entries, allowing developers to update Claude’s instructions on the fly without resetting the prompt or simulating user turns, facilitating dynamic adjustments to permissions, token budgets, or context during agent runs.

Anthropic announced a $65 billion Series H financing round, raising its post‑money valuation to $965 billion. Lead investors include Altimeter Capital, Dragoneer, Greenoaks, and Sequoia, with additional commitments from Amazon ($5 billion) and memory manufacturers Micron, Samsung, and SK Hynix. The capital will fund further AI research and capacity expansion.

Looking ahead, Anthropic is developing a new class of models under the “Glasswing” project that will surpass Opus capabilities. Early adopters are testing the Claude Mythos preview for cybersecurity tasks, and Anthropic is building stronger safety measures before a broader release.