Claude Opus 4.8 is more honest, less deceptive, and considerably cheaper

At a glance:

• Anthropic launches Claude Opus 4.8, beating Opus 4.7 on most benchmark scores
• Fast‑mode pricing is three times cheaper, making simple queries far more affordable
• New effort controls and a research‑preview dynamic‑workflow system give users finer control over speed, cost and task complexity

What’s new in Claude Opus 4.8

Anthropic announced the latest iteration of its flagship large‑language model, Claude Opus 4.8, on the same day it rolled out a suite of user‑facing features. According to the company, Opus 4.8 “reaches new highs on our measures of prosocial traits,” meaning it is more inclined to support user autonomy and act in the user’s best interest. In benchmark testing the model outperformed Opus 4.7 across almost every metric, from reasoning accuracy to code generation quality.

Beyond raw performance, Anthropic highlighted a shift in the model’s behavioural profile. Opus 4.8 is described as “more honest” and “four times less likely to not report flaws in code it writes.” It also shows a markedly lower propensity for deceptive outputs or cooperation with misuse, and it is more likely to flag uncertainty instead of presenting unverified claims as facts.

Pricing and fast‑mode improvements

One of the most tangible upgrades for developers and enterprises is the cost reduction on the model’s fast mode. Anthropic says fast‑mode responses are now three times cheaper than they were for Opus 4.7, opening the door for high‑volume, low‑complexity queries without the token‑burn that previously limited usage. The cheaper fast mode does not sacrifice the model’s core safety guarantees, making it an attractive option for cost‑sensitive workloads.

Effort controls let users dial speed versus depth

Anthropic introduced an “Effort” selector in the Claude UI, offering four levels: Low, Medium, High, and Max. Selecting a lower effort setting returns faster, more token‑efficient answers, while higher settings cause the model to spend more computation time, yielding richer, more detailed responses. The company warns that higher effort modes consume tokens at a faster rate and recommends reserving the “Max” setting for the most demanding tasks.

Dynamic workflows in research preview

In a separate research‑preview feature called Dynamic workflows, Claude can now orchestrate hundreds of sub‑agents to tackle larger problems, particularly within Claude Code. Users simply provide a high‑level task, and the system automatically plans, executes, and verifies each sub‑step before presenting the final output. This capability aims to reduce the need for manual prompting and to improve reliability on complex, multi‑stage projects.

Availability and the road ahead

Claude Opus 4.8 is available globally starting today across all Anthropic platforms. The company also hinted at the upcoming release of “Mythos‑class” models in the coming weeks, suggesting a broader family of specialized LLMs is on the horizon. Existing Claude users can immediately opt into the new model and experiment with effort controls and dynamic workflows via the research preview.

What this means for the AI landscape

Anthropic’s dual focus on safety‑centric behaviour and cost efficiency positions Opus 4.8 as a compelling alternative to other high‑end models that often trade one for the other. By making fast‑mode cheaper and giving users granular control over computational effort, Anthropic is addressing two of the biggest friction points for enterprise adoption: budget predictability and output quality. The dynamic‑workflow preview also signals a move toward more autonomous AI assistants that can manage complex pipelines with minimal human supervision, a trend likely to accelerate across the industry.