Claude Opus 4.8 Achieves Two Historic Firsts with Zero‑Error Metrics

Claude Opus 4.8 was launched only 43 days after Opus 4.7, keeping the same price while claiming a decisive lead in AI rankings. The model’s performance is highlighted by a GDPval‑AA Elo score of 1890, 137 points higher than 4.7 and 121 points above GPT‑5.5, which translates to an estimated 67 % win‑rate in head‑to‑head matches.

Compared with Opus 4.7, Opus 4.8 completes the same tasks with 15 % fewer steps and 35 % fewer tokens, demonstrating both speed and efficiency gains.

A key focus of the release is honesty. The “false‑reporting” metric drops from 0.40 in Opus 4.5 and 0.25 in Opus 4.7 to 0.00 in Opus 4.8, meaning the model never reports fabricated numbers. Similarly, the “lazy‑rate” (the tendency to give a quick but incorrect answer) falls from 25 % in Opus 4.7 to 0 % in Opus 4.8.

Benchmark results reinforce these claims. On SWE‑Bench Pro, Opus 4.8 achieves 69.2 % accuracy, a full 10‑point gain over GPT‑5.5. In the more demanding ProgramBench, Opus 4.8 outperforms 4.7 at every context‑budget tier; with a 1 M‑token budget it reaches ~79.5 % pass rate, whereas 4.7 only reaches that level with a 5 M‑token budget.

On the FrontierSWE leaderboard, which tests extreme system‑engineering tasks, Opus 4.8 attains an 83 % win‑rate, leaving GPT‑5.5 and Opus 4.7 far behind.

Despite these gains, the article notes that Opus 4.8 still has blind spots, but it pushes the frontier of AI‑driven software development. A concrete case study shows a developer using Claude Code + Opus 4.8 to rewrite a 750 k‑line Zig codebase into Rust. In 11 days the model generated 75 k lines of Rust, passed 99.8 % of the original tests, produced over 6 000 commits, and required virtually no human line‑by‑line review.

The release also introduces “dynamic workflows” with five effort levels (Low → Max) plus an “ultracode” mode. Higher effort levels allocate more reasoning tokens and trigger larger agent fleets. In “fast mode” the model runs 2.5× faster while costing only one‑third of the usual price. The ultracode setting automatically decides whether to engage a full army of sub‑agents for a task.

An example illustrates the workflow: a developer asks Claude to rewrite Bun (a JavaScript runtime) from Zig to Rust. The system first annotates lifetimes, then translates each file, assigns two reviewers per file, runs a repair loop, and finally merges everything. The result is a production‑ready Rust codebase with 99.8 % test coverage, completed in 11 days.

Anthropic also announced that the upcoming Claude Mythos model will be released in a few weeks, further extending these capabilities.

Financially, Anthropic closed a $650 billion Series H round, valuing the company at $9.65 trillion—surpassing OpenAI’s $8.52 trillion valuation and making Anthropic the world’s most valuable AI startup.

Overall, the article presents Opus 4.8 as a major leap in AI model performance, cost efficiency, and autonomous workflow orchestration, while hinting at future advances with Claude Mythos.