Model comparisons are usually marketing. Vendors pick the benchmark that flatters them and skip the one that doesn't. Every so often, though, a comparison shows up as a side effect of something else entirely, and it ends up more honest than any benchmark chart. That's what happened in Anthropic's write-up explaining why it had to redeploy Claude Fable 5 after a government-ordered suspension.
The comparison nobody planned to publish
As part of investigating the Amazon researchers' report that triggered the export controls, Anthropic tested whether other models could reproduce the same vulnerability-finding behavior that got Fable 5 flagged. The answer, in Anthropic's own words: yes, across the board. Claude Opus 4.8, GPT-5.5, and Kimi K2.7 could all identify the same vulnerabilities Fable 5 did. When it came to producing a working demonstration of how to exploit the single vulnerability in the report, every model they tried could do it — including much smaller ones like Claude Haiku 4.5 and Sonnet 4.6.
That's a striking admission to put in your own safety report. It means a capability that looked, on the surface, like it might be uniquely dangerous in a frontier model turned out to already be present in models several tiers down, from at least three different labs.
What this actually tells you about the current landscape
Capability convergence across labs is happening faster than most public benchmarks capture. If a mid-tier model like Haiku 4.5 can reproduce something Fable 5 was flagged for, the gap between "frontier" and "everyone else" on that specific class of task has effectively closed. It also means safety work has to be classifier-and-policy driven rather than capability-gatekeeping driven — you can't keep a capability locked to one tier of model once it's genuinely common across the field, so the safeguard has to sit at the request level, not the model level.
Meanwhile, the tiered lineup approach — a cheap fast model, a mid-tier agentic workhorse, and an expensive frontier option — has become the standard shape for every major lab's product line, not just Anthropic's. Sonnet 5 narrowing the gap with Opus 4.8 while undercutting it on price is the same story OpenAI and others have been telling with their own mid-tier releases.
The honest takeaway
None of this means the models are identical, or that model choice doesn't matter. It means the safest assumption for anyone building on frontier AI right now is that a meaningful chunk of what your chosen model can do, competitors' models can probably do too, at a nearby tier and price point. Pick based on your actual workflow fit, your existing tooling, and vendor reliability, rather than assuming one lab has a durable, unique capability lead. As this summer's export control episode showed, that lead — real or perceived — can also disappear from your product for three weeks with zero warning.
