For five weeks, OpenAI had the crown.

On May 28, Anthropic took it back.

Claude Opus 4.8 shipped today, and the pitch is blunt: better coding, better agents, better judgment, same price as Opus 4.7. The API name is claude-opus-4-8. Input is still $5 per million tokens, output still $25. No premium for the upgrade. That detail matters more than it looks, and we will get to why.

First, the leaderboard story.

How OpenAI grabbed the lead in April

GPT-5.5 landed April 23. It was a real jump, not a point release.

OpenAI posted state-of-the-art numbers across agentic coding and computer use: 82.7% on Terminal-Bench 2.0, 84.9% on GDPval knowledge work, 78.7% on OSWorld-Verified. It beat Claude Opus 4.7 on most of those boards. It even cleared FrontierMath Tier 4 — the hardest tier — at 35.4%. Testers were not subtle. One NVIDIA engineer said losing access "feels like I've had a limb amputated."

So for five weeks, if you wanted the strongest agent for long-horizon coding, the honest answer was GPT-5.5.

That is the part that stung. Anthropic had been the default for Claude Code users, and OpenAI walked in and took the benchmark high ground.

What Opus 4.8 actually beats

Anthropic did not answer with a bigger number on one chart. It answered on the boards that decide real agent work.

On the Super-Agent benchmark, Opus 4.8 was the only model to complete every case end-to-end — beating prior Opus models and GPT-5.5 at cost parity. On Online-Mind2Web, the browser-agent test, it scored 84%, which Anthropic calls the strongest computer-use model it has tested, ahead of both Opus 4.7 and GPT-5.5. On a partner Legal Agent Benchmark, it posted the highest score on record and became the first model to break 10% on the all-pass standard.

Read that "at cost parity" line again.

GPT-5.5 charges $30 per million output tokens. Opus 4.8 charges $25. So Anthropic is not just claiming the win — it is claiming the win while costing less per output token than the model it passed. That is the whole argument in one sentence.

This is not a clean sweep. GPT-5.5 still leads some boards, and Claude Opus 4.7 already held SWE-Bench Pro at 64.3% against GPT-5.5's 58.6%. If your work lives in raw FrontierMath or pure terminal benchmarks, OpenAI is still competitive. The frontier is a tie in a lot of places. But on end-to-end agent completion — the thing that actually ships work — Opus 4.8 is back in front.

The honesty number is the real headline

Here is the feature nobody put on a billboard.

Anthropic says Opus 4.8 is about 4x less likely than its predecessor to let a flaw in code pass unremarked. It flags uncertainty more. It makes fewer unsupported claims. It stops doing the thing every model does — confidently announcing progress on thin evidence.

If you have ever watched a coding agent declare victory on a broken build, you know exactly why this matters.

A faster model that lies to you about its own output is not faster. It is a liability with low latency. An agent that pushes back when a plan is unsound, catches its own mistakes, and admits what it does not know is worth more than two points on a benchmark. The Devin team at Cognition said it directly: 4.8 fixes the comment-verbosity and tool-calling problems they saw in Opus 4.7.

That is the quiet shift. The race stopped being only about capability. It is now also about whether you can trust what the agent tells you it did.

What shipped alongside the model

Three things landed with it, and they are aimed at people running long jobs:

  • Effort control. A new dial on claude.ai and in Claude Code. Higher effort thinks longer; lower effort burns your rate limits slower. Default is high. For hard async work, Anthropic recommends "extra."
  • Dynamic workflows in Claude Code. A research preview that plans work and runs hundreds of parallel subagents in one session, verifying outputs before it reports back. The example use case is a codebase-scale migration across hundreds of thousands of lines, kickoff to merge, using your existing test suite as the bar.
  • Fast mode. 2.5x the speed, now 3x cheaper than the previous generation.

None of that is a chatbot feature. All of it is built for agents that run unattended for hours.

What you should actually do

Do not switch your whole stack on a launch-day chart.

If you run Claude Code or build agents on the Claude Platform, test Opus 4.8 on your own workload this week — not on Anthropic's benchmarks, on your repo, your migration, your failure cases. The cost-parity upgrade means there is little reason not to.

If you live in Codex and GPT-5.5, you are not stuck on a losing model. GPT-5.5 still holds real ground, especially in terminal-heavy and math-heavy work. Switching agents has a tax — retooling prompts, re-testing pipelines, re-learning quirks. Pay it only when the win clears the cost.

And whichever side you land on, ask the harder question. Not "which model scores higher." Ask which model lies to you less about its own work. After this release, that is the benchmark that should move your decision.

The leaderboard will flip again. It always does. OpenAI will answer, and the spread between these two will stay narrow for a while yet. What changes slower is trust — and right now Anthropic just made that the battleground.

For the wider context on where this race started, see our coverage of OpenAI Codex GPT-5.5 as a work agent, our GPT-5.2 review on how OpenAI's model line evolved, and our breakdown of AI agents vs chatbots for why agentic benchmarks are the ones that matter now.

The AMA Report cuts through the noise — practical intelligence for business operators making real decisions about which AI tools actually earn their cost. Subscribe free →

The AMA Report cuts through the noise — practical intelligence for business operators making real decisions about which AI tools actually earn their cost. Subscribe free →