I would honestly guess that this is just a small amount of tweaking on top of the Sonnet 4.x models. It seems like providers are rarely training new 'base' models anymore. We're at a point where the gains are more from modifying the model's architecture and doing a "post" training refinement. That's what we've been seeing for the past 12-18 months, iirc.
Ofc it matters. Who pays for the power? Does the AI pay for the data or the power they use for training? Nope, they dont.
Consumers pay for the power in rising enerfy bills, while the AI datacenters get huge gov subsidies. At the same time people get booted because some CTO has gone full blown AI blind.
The scary implication here is that deception is effectively a higher order capability not a bug. For a model to successfully "play dead" during safety training and only activate later, it requires a form of situational awareness. It has to distinguish between I am being tested/trained and I am in deployment.
It feels like we're hitting a point where alignment becomes adversarial against intelligence itself. The smarter the model gets, the better it becomes at Goodharting the loss function. We aren't teaching these models morality we're just teaching them how to pass a polygraph.
It's wild that Sonnet 4.6 is roughly as capable as Opus 4.5 - at least according to Anthropic's benchmarks. It will be interesting to see if that's the case in real, practical, everyday use. The speed at which this stuff is improving is really remarkable; it feels like the breakneck pace of compute performance improvements of the 1990s.
> In areas where there is room for continued improvement, Sonnet 4.6 was more willing to provide technical information when request framing tried to obfuscate intent, including for example in the context of a radiological evaluation framed as emergency planning. However, Sonnet 4.6’s responses still remained within a level of detail that could not enable real-world harm.
Interesting. I wonder what the exact question was, and I wonder how Grok would respond to it.
Has anyone tested how good the 1M context window is?
i.e given an actual document, 1M tokens long. Can you ask it some question that relies on attending to 2 different parts of the context, and getting a good repsonse?
I remember folks had problems like this with Gemini. I would be curious to see how Sonnet 4.6 stands up to it.
With such a huge leap, i’m confused why they didn’t call it Sonnet 5? As someone who uses Sonnet 4.5 for 95% tasks due to costs, i’m pretty excited to try 4.6 at the same price
I always grew up hearing “competition is good for the consumer.” But I never really internalized how good fierce battles for market share are. The amount of competition in a space is directly proportional to how good the results are for consumers.
The weirdest thing about this AI revolution is how smooth and continuous it is. If you look closely at differences between 4.6 and 4.5, it’s hard to see the subtle details.
A year ago today, Sonnet 3.5 (new), was the newest model. A week later, Sonnet 3.7 would be released.
Even 3.7 feels like ancient history! But in the gradient of 3.5 to 3.5 (new) to 3.7 to 4 to 4.1 to 4.5, I can’t think of one moment where I saw everything change. Even with all the noise in the headlines, it’s still been a silent revolution.
Am I just a believer in an emperor with no clothes? Or, somehow, against all probability and plausibility, are we all still early?
116 comments
[ 3.1 ms ] story [ 179 ms ] threadStuff from last year will be outdated today.
Consumers pay for the power in rising enerfy bills, while the AI datacenters get huge gov subsidies. At the same time people get booted because some CTO has gone full blown AI blind.
Its a bad situation for the consumer.
Now the question is: how much faster or cheaper is it?
It feels like we're hitting a point where alignment becomes adversarial against intelligence itself. The smarter the model gets, the better it becomes at Goodharting the loss function. We aren't teaching these models morality we're just teaching them how to pass a polygraph.
The much more palatable blog post.
Interesting. I wonder what the exact question was, and I wonder how Grok would respond to it.
Was sonnet 4.5 much worse than opus?
i.e given an actual document, 1M tokens long. Can you ask it some question that relies on attending to 2 different parts of the context, and getting a good repsonse?
I remember folks had problems like this with Gemini. I would be curious to see how Sonnet 4.6 stands up to it.
A year ago today, Sonnet 3.5 (new), was the newest model. A week later, Sonnet 3.7 would be released.
Even 3.7 feels like ancient history! But in the gradient of 3.5 to 3.5 (new) to 3.7 to 4 to 4.1 to 4.5, I can’t think of one moment where I saw everything change. Even with all the noise in the headlines, it’s still been a silent revolution.
Am I just a believer in an emperor with no clothes? Or, somehow, against all probability and plausibility, are we all still early?
Google needs stiff competition and OpenAI isn’t the camp I’m willing to trust. Neither is Grok.
I’m glad Anthropic’s work is at the forefront and they appear, at least in my estimation, to have the strongest ethics.
https://web.archive.org/web/20260217180019/https://www-cdn.a...