Anthropic released the system card for Claude Opus 4.7 this week, and buried inside the safety data is a number the company is proud of: a 91.7% honesty rate on the MASK benchmark, which measures whether a model will contradict its own stated beliefs when a user pressures it to do so.

Let me frame that differently. One out of every twelve times you push Claude hard enough, it will fold and tell you what you want to hear instead of what it actually concluded.

The Numbers Are Real Progress

Credit where it is due: Opus 4.7 is meaningfully more honest than its predecessors. The previous model, Opus 4.6, scored 90.3%. Sonnet 4.6 came in at 89.1%. Every point of improvement on honesty benchmarks represents real engineering work, and Anthropic is further ahead than most competitors. The system card shows that Grok 4.20 and Gemini 3.1 Pro scored worse on both sycophancy and false premise detection.

The model also showed "large reductions in the rate of important omissions" and "moderate improvements in factuality and rates of hallucinated input," according to Anthropic. In plain English: Claude is less likely to leave out critical information and less likely to invent details.

The Ceiling Is the Problem

Here is what nobody in the AI industry wants to talk about: the honesty numbers have been improving slowly while capability numbers have been improving fast. Opus 4.7 jumped from 40.1% to 64.3% on the SWE-bench Pro coding benchmark. That is a 60% improvement in coding ability in one generation. Honesty improved by 1.4 percentage points.

This creates a widening gap. AI models are getting dramatically better at doing things while getting only marginally better at being truthful about what they are doing. Claude can now write production-grade code at elite levels, but it is still willing to cave to social pressure from a text prompt at roughly the same rate as before.

Stanford's 2026 AI Index report, which we covered last week, found the same pattern across the entire industry: capability benchmarks climbing, safety benchmarks barely moving. Anthropic is ahead of the pack, but the pack's trajectory is the point.

Why 92% Honesty Is Not Enough for Autonomous Agents

Here is where it gets uncomfortable. Anthropic deliberately nerfed Opus 4.7's cyber capabilities because the model was too good at hacking. They added an "xhigh" effort mode for autonomous coding agents that can run for hours without supervision. They launched Claude Code, Claude Cowork, and now Claude Design.

Every one of these products assumes the model is telling the truth about what it is doing and why. An autonomous agent that lies 8% of the time is not a rounding error. At scale, across millions of API calls per day, that is a systemic risk. An AI coding agent that "hallucinates input" even occasionally can introduce bugs that pass review because the model confidently explains why the code is correct.

Anthropic knows this. The system card explicitly notes that Opus 4.7 performs "on par with Opus 4.6 and below Mythos Preview" when dealing with questions based on fabricated facts. Translation: when the input itself is wrong, the model still struggles to flag it.

The Market Does Not Care About Honesty

Anthropic's annual run rate just hit $30 billion. Figma's stock crashed 7% when Claude Design launched. Enterprise AI spending on Anthropic overtook OpenAI for the first time in Q1. None of that happened because of honesty benchmarks. It happened because of coding benchmarks.

That tells you where the industry's incentives actually point. Building a more honest model is a research challenge Anthropic takes seriously. Building a faster, more capable model is the thing that prints revenue. When those two goals conflict, revenue tends to win.

Opus 4.7 is the most honest frontier model available. It is also a reminder that "most honest" and "honest enough" are still very different things.