Anthropic launches Claude Opus 4.8 with faster, safer coding

Anthropic released Claude Opus 4.8 on Thursday, six weeks after Opus 4.7. The model is available on claude.ai and in Claude Code. Existing Opus customers receive the update at the same standard rates.

Opus 4.8 shows numeric gains on several benchmark suites. On SWE‑bench Pro, which measures the ability to solve hard, multilingual production coding problems, Opus 4.8 scored 69.2%, up from 64.3% for Opus 4.7. OpenAI’s GPT‑5.5 scored 58.6% and Google’s Gemini 3.1 Pro scored 54.2% on the same test. On Humanity’s Last Exam, an expert‑level academic question set, Opus 4.8 reached 49.8% without tools and 57.9% with tools. OSWorld‑Verified, a test of real‑world software UI tasks, rose to 83.4% from 82.8%. On Terminal‑Bench 2.1 for command‑line tasks, Opus 4.8 scored 74.6% (up from 66.1%); GPT‑5.5 led that test at 78.2%.

Anthropic reported lower rates of deceptive and misuse‑cooperation behavior compared with Opus 4.7 and fewer missed code bugs. The company wrote that Opus 4.8 “reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest.” Anthropic added that Opus 4.8 is four times less likely than Opus 4.7 to leave bugs in its own code unflagged.

The company contrasted Opus 4.8’s alignment with its restricted Claude Mythos preview, a larger model currently available to vetted cybersecurity organizations through Project Glasswing. Anthropic wrote that Opus 4.8’s deceptive and misuse‑friendly behavior rates are comparable to the Mythos preview on its internal measures and stated it is “making swift progress on developing these safeguards and expect[ing] to be able to bring Mythos‑class models to all our customers in the coming weeks.”

Standard pricing for Opus 4.8 remains $5 per million input tokens and $25 per million output tokens. Anthropic introduced a faster mode that runs the same model at roughly 2.5× speed for $10 per million input and $50 per million output; the company reported that the fast rate is about three times cheaper than previous fast modes on earlier models. The fast option uses more compute and targets latency‑sensitive use cases.

Opus 4.8 includes new user controls and developer features. Effort settings let users choose compute spend: Low and Medium use fewer tokens, High is the default, Extra (labeled “xhigh” inside Claude Code) uses more compute for harder problems, and Max is the deepest setting. Anthropic implemented a new tokenizer that produces more tokens per task, raised rate limits in Claude Code to handle higher token consumption for Extra and Max settings, and released dynamic workflows in research preview for Enterprise, Team and Max plan customers. Dynamic workflows allow the model to spawn parallel subagents and verify outputs in a single session; Anthropic noted these workflows use significantly more tokens than standard sessions.

The release comes amid a price gap between U.S. models and lower‑cost offerings from China. DeepSeek V4 Pro and Xiaomi MiMo V2.5 Pro are available at roughly $0.435 per million input tokens and $0.87 per million output tokens through some providers. Anthropic framed the difference as a tradeoff between cost, performance and safety for regulated industries and sensitive production workloads.

A quick coding test comparing the three models on a single 3D zombie game task showed different tradeoffs: GPT‑5.5 produced a fast but incomplete result, DeepSeek V4 Pro delivered a solid, fast build, and Opus 4.8 produced richer design and mechanics while running slower. Anthropic confirmed developers already paying the $5/$25 rates for Opus models receive the Opus 4.8 improvements without a price change.