Anthropic's Claude Opus 4.8 Takes the Top Spot on AI Benchmarks — Coding and Honesty Leap Forward
Anthropic released Claude Opus 4.8 on May 28, reclaiming the #1 position on the Artificial Analysis Intelligence Index with major gains in coding, scientific reasoning, and code quality honesty.
Anthropic’s Claude Opus 4.8 is the new leader on the Artificial Analysis Intelligence Index, scoring 61.4 — up 4.1 points from Opus 4.7 and 1.2 points ahead of GPT-5.5. Released on May 28, 2026, it reclaims the benchmark crown Anthropic briefly ceded to OpenAI with a meaningful jump in coding, scientific reasoning, and a long-overdue fix for how the model handles its own mistakes in code.
Coding Performance
Opus 4.8 scores 69.2% on SWE-Bench Pro, up from 64.3% for Opus 4.7, and beats both GPT-5.5 and Gemini 3.1 Pro on that benchmark. On SWE-Bench Verified it reaches 88.6%. The Elo on GDPval-AA hits 1,890, implying roughly a 67% win rate against GPT-5.5 in direct head-to-head agentic evaluations.
The standout behavioral change: Opus 4.8 is approximately four times less likely than its predecessor to let flaws in code it just wrote pass unremarked. That’s not a benchmark number — it’s a qualitative shift in how the model behaves during long coding sessions. Fewer silent errors means fewer downstream debugging sessions caused by Claude’s own oversights.
Scientific Reasoning
The model leads Humanity’s Last Exam by one point over GPT-5.5 and overtakes Gemini 3.1 Pro on CritPt, a frontier physics evaluation. Accuracy sits at 46.6%; hallucination rate holds roughly flat at 35.9%, consistent with Opus 4.7.
New “Dynamic Workflow” Tool
Opus 4.8 ships with a new dynamic workflow tool that lets the model plan and adapt multi-step tasks without requiring the user to pre-define every step. It’s particularly useful for long-horizon coding and research tasks where the path to completion isn’t known upfront — the model can adjust its plan when it hits a dead end instead of failing or looping.
Pricing
Pricing is unchanged: $5 per million input tokens, $25 per million output tokens. Prompt caching brings costs down by up to 90%; batch processing cuts them by 50%. No pricing change for the API tier compared to Opus 4.7.
Where It Falls Short
GPT-5.5 still leads on terminal-coding benchmarks. Hallucination rate didn’t drop meaningfully. And the dynamic workflow tool is currently API-only — Claude.ai users won’t see it surface as a user-facing feature at launch.
The overall picture: Opus 4.8 is a genuine step up from 4.7, particularly if your use case involves agentic coding loops or long research tasks. The four-times-less-likely-to-miss-code-flaws improvement alone is worth the upgrade for serious engineering workflows.