Anthropic's Claude Opus 4.8 Bets on Honesty Over Raw Scale

Anthropic released Claude Opus 4.8 on 28 May, and the headline the company chose to lead with was not a benchmark. It was honesty. Opus 4.8 is, by Anthropic's own measure, four times less likely than its predecessor to let a flaw in code pass without flagging it, according to the company's announcement. For a model increasingly trusted to write and review software on its own, that is the number that matters most.

The coding gains are real too. On SWE-bench Verified, a test of whether a model can resolve genuine software issues, Opus 4.8 scored 88.6 percent against 87.6 percent for Opus 4.7. On the harder SWE-bench Pro, it reached 69.2 percent, up from 64.3 percent, and ahead of OpenAI's GPT-5.5 at 58.6 percent and Google's Gemini 3.1 Pro at 54.2 percent, as reported by Vellum. Those are incremental steps, not a leap, and that fits the mood of the moment. May has been a month of architecture and efficiency rather than giant jumps in capability.

Better and cheaper to run

What stands out is the cost side. Anthropic says Opus 4.8 reaches those scores using roughly 35 percent fewer output tokens than Opus 4.7, and it ships at the same price, five dollars per million input tokens and twenty-five per million output. The model is available immediately through the Claude API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, with the API identifier claude-opus-4-8.

There is also a new feature aimed squarely at how people now use these systems. TechCrunch reported on a dynamic workflows tool in Claude Code, where the model writes its own orchestration scripts, spins up dozens or hundreds of parallel sub-agents to attack a problem from different angles, then deploys adversarial agents to try to refute the findings before reporting back. It is a long way from autocomplete.

A loud week for the company

The launch landed on the same day Anthropic closed a $65 billion funding round that pushed its valuation past OpenAI's. It also comes weeks after the company's Mythos preview model demonstrated autonomous cyber capabilities in lab tests that regulators have been watching closely. A flagship model that is measurably more willing to admit when something is wrong reads, in that context, less like a marketing line and more like a deliberate answer to the trust problem these tools have created for themselves.

I would still want to see independent testing before treating the honesty figure as settled. Vendor benchmarks are vendor benchmarks. But the direction is worth noting: a frontier lab choosing self-correction as its lead pitch, rather than another point on a saturated leaderboard.