If you’ve been using AI lately, you know the discourse. It’s been a frustrating few weeks for the power users and there’s a growing chorus on Reddit and X that the latest updates from Google and Anthropic have been “nerfed”—shaved down at the edges, made more cautious, and frankly, a bit dumber in the name of safety and “alignment.” I fully agree with this assessment as someone who uses them daily and can’t believe how dome Gemini in particular has gotten.
Both Anthropic and Google have spent the last two months getting hammered by power users who say their flagship models have been quietly degraded — what developers have started calling “AI shrinkflation.” An AMD senior director ran the forensics on 6,852 Claude Code sessions and found median reasoning depth collapsed roughly 73% between January and March. Independent benchmarks from Marginlab show Opus 4.6’s SWE-Bench-Pro pass rate sliding from 56% to 50%. Gemini 3 Pro users have been filing near-identical complaints in Google’s own developer forums since December — context windows that forget at 32k despite being marketed as 1M, documents that won’t process, reasoning is obviously thinner.
In any case, Anthropic just dropped Claude Opus 4.7, and while the headlines say “faster and better at coding,” the real story is what’s not in the box.
The Benchmarks
The technicals on Opus 4.7 look good on paper. Anthropic is touting a massive leap in vision capabilities and a specific focus on “economically valuable knowledge work.”
They’re pointing to the GDPval-AA benchmark—a metric that measures finance and legal domain expertise—where 4.7 is showing clear gains over its predecessor. In the coding world, the “ultrareview” command and “xhigh” reasoning levels suggest they’re trying to claw back the trust of developers who felt the previous versions were starting to hallucinate or get lazy.
The model everyone actually wants—Claude Mythos—is still in “Preview” and under lock and key. But note that they’re teasing it in the table above, which was released in the PR for 4.7. Anthropic has said that Mythos is too powerful (or too dangerous for cybersecurity) to let the general public touch it yet.
They’ve launched “Project Glasswing” to gatekeep the top-tier intelligence, citing safety concerns. For the market, this is a double-edged sword. On one hand, it builds the “God-model” hype. On the other, it confirms the suspicion that the models we can pay for are effectively the “lite” versions, throttled by safeguards that act as a drag on performance, or that there is some kind of degradation or throttling due to over use.
From an investment perspective, this is where it gets interesting. We’re seeing a shift in the AI trade. The initial “wow” factor of LLMs is hitting the reality of corporate safety and compute costs.
When users complain about a model being “nerfed,” what they’re usually seeing is a company trying to save on inference costs or avoid a PR nightmare. If the “usefulness” of these tools is being capped by safety protocols, the ROI for businesses starts to look a lot different.
It seems to me that Opus 4.7 is a “bridge” model. It’s designed to keep users happy while Anthropic figures out how to release Mythos without breaking the world (or their servers). The whole discourse around nerfing actually bodes well for inference plays like Micron (MU) and Samsung. It’s also good for Broadcom (AVGO) and Marvell (MRVL), who are designing more efficient chips for Google.
Keep an eye on the design stocks today—Adobe and Wix are already feeling the heat from the web-design rumors surrounding this release. But until “Mythos” is actually out in the wild, the AI sector is trading on potential, not full-throttle delivery
