Anthropic Didn’t Release a Better Model Because It Got Too Good
The Opus 4.7 degradation
Opus 4.7 dropped recently. Most people saw “better benchmarks” and moved on.
But tucked inside Anthropic’s own documentation was a sentence that changed how I think about the whole thing.
Mythos Preview is their most robustly aligned model. The model they released to the public is NOT the most aligned version they have. And more importantly: the most aligned version is also not the most capable.
The Numbers
Yes, Opus 4.7 outperforms 4.6 on most headline benchmarks. Finance agent tasks, long-context reasoning, complex coding, vision. They went up. The press release is real.
But benchmarks measure what they measure. And what they often don’t measure is behavior.
Opus 4.7 produces overly detailed harm-reduction advice in areas where 4.6 was more restrained. Flagged in Anthropic’s own safety documentation as a regression. Not a bug they’re fixing. A known shift when they tuned the alignment parameters to a different position.
This is what aligned AI looks like in practice, it seems. Not a clean improvement. A tradeoff. Some things go up. Some things go sideways. Some things get dialed back in ways that don’t show up in leaderboard screenshots.
The Alignment
There’s a concept in AI research called the alignment tax.
Making a model safer costs you something. You can’t maximize both raw capability and corrigibility at the same time. At some threshold, the guardrails constrain what the model will do.
Anthropic says this isn’t entirely true. They’ve argued publicly that alignment and capability aren’t opposing forces. That a well-aligned model is more valuable, not less capable.
Probably…
But “not opposing forces” is different from “no tradeoffs.” And Mythos Preview is the proof. If there were truly no tradeoffs, the most capable model and the most aligned model would be the same model. They’re not.
Two separate products. Two separate positions on the tradeoff curve.
Coding
I’ve been building with Claude Code daily for the last few months. Flutter apps, React frontends, PHP backends. I used Anti-Gravity but went back to Claude. Claude Code keeps winning.
More aligned models add hedging. They ask clarifying questions before writing code that touches user data. They attach security warnings to implementations that are pretty standard. They occasionally refuse edge cases the previous version handled without complaint.
Individually, every one of those behaviors is reasonable. Together, they add some friction.
Not a lot. But some.
The “safer” the model, the more it second-guesses itself. I’ve started dictating longer context prompts using Wispr Flow*, it’s faster by voice than typing the setup every time.
The Capability
Anthropic had a more capable version of Opus 4.7. Probably significantly more capable on certain tasks. And at some point in testing, that version crossed a threshold. Maybe a harmful threshold. Maybe a threshold where the behaviors became harder to constrain without meaningful capability loss.
So they didn’t ship that version.
They shipped what we got. Which is very good. But it’s not the ceiling of what Anthropic built. It’s the ceiling of what they were comfortable releasing.
I’ve written before about how AI coding is changing everything right now, and I’ve also written about the AI privacy problem that vibe-coded tools ignore. A more capable model that’s harder to align might be a worse product in those contexts, not a better one.
Mythos Preview exists as a separate, more constrained product. You can have more alignment. It just costs you somewhere else on the spec sheet.
Options
Sonnet 4.6 is faster, cheaper, and has slightly fewer default constraints on code generation. For most development tasks, it’s close in quality and outperforms Opus on pure throughput.
Opus 4.7 is excellent for complex agentic tasks, long document analysis, and anything where reasoning depth matters more than raw speed. That’s still a lot of real work.
If you want to understand the tradeoffs before picking: that’s the most interesting position to be in. The answer is it depends on what you’re building and how much you want the model to push back on you while you build it.
I mostly run Claude Code in the terminal, switch to Sonnet 4.6 for fast iteration, and pull in Opus for the hard stuff.
The Bottom Line
Anthropic is a safety-first AI lab that happens to ship incredibly capable models. The two things are in tension. Not irreconcilably. But meaningfully.
Every release is a negotiation between what the model can do and what Anthropic is willing to let it do. Opus 4.7 landed at a certain point in that negotiation. Mythos Preview landed at a different one.
The version you wanted might have existed.
It just never shipped.
*this is an affiliate or SparkLoop* partner link. I may earn a commission.