Hacker News Reader

Comment by czhu12 | Hacker News Reader

Its remarkable how Anthropic is able to maintain their edge against all competition. Anyone have any idea what the secret sauce is that has Anthropic at the top of all leaderboards for the past few years?

View on HN

My gut feel is Anthropic is very technical and pedantic which makes their models really technical and pedantic. They're top at code and technical benchmarks but anecdotally I've found OpenAI to be significantly farther ahead for general usage.

Opus 4.8 will burn 10k tokens trying to answer something 100% whereas GPT-5.5 will burn 2k getting it 90% which is good enough for many things.

Some personal testing on a "help me find that restaurant" prompt https://gist.github.com/nijave/2873b8b10d8c732e46264237b0755...

The problem is that the remaining 10% can bite you in bad ways.

I was in Cotswolds, UK a couple of months ago. For those of you who don't know, it's a rural region known for its "chocolate-box" villages and honey-colored limestone architecture. Basically, you go from village to village, most commonly via bus, taking in the sights and doing touristy stuff.

When planning the trip, my sister used ChatGPT, which helpfully (and relatively quickly) found the bus schedules and times for each hop.

Midway through the day, though, we ran into a huge problem: it turns out bus schedules are different on Sundays, and more limited. Which meant we couldn't actually go to our primary destination (the Model Village), and had to cut the trip short.

Yes, ChatGPT was quick and pleasant to use, but missed a crucial detail.

Afterwards I tried it with Opus and it did not make the same mistake.

Arguably I'd call that the 90%. In my case, answering the restaurant question correctly with "Rishi" in my tests was the sole intent and 90% of the problem. All the models "helpfully" added extra junk about the closure, dates, quotes, etc and many of them got these details wrong--the 10% or extra crap not central to the question.

If the central question was "what is the bus schedule on `day`" and the model screws that up, it gets a fail in my book.

Also curious if Google Maps gets the timetables correct (assuming it has them).

Semi-related, I also discovered that the default web search/fetch tools are pretty primitive and Exa MCP annihilates them. I ended up doing some comparisons with Claude Code comparing built-in server-side to Exa and to a Python MCP that used SearXNG for search and Exa was a clear winner and Python+SearXNG ended up coming out roughly the same after a few cycles of letting Claude optimize the Python code and adjust SearXNG settings. Ultimately it landed on this (making some changes to optimize returning relevant context directly in the search results so the model didn't need an additional web fetch call) https://gist.github.com/nijave/604c43e3e0fdcd60f5280d3a6b109...

This likely comes down to how it accessed the bus schedules (i.e. web search tool) and not intelligence.

You need to add the actual bus schedule to context somehow (research agent, custom tool or just dump in prompt) and even the simpler modern models will be able to do the planning.

Tool usage competency is part of overall intelligence. If the model can't get the information it needs, it must clarify that in the response.

This isn't tool usage competency, it's tool quality and/or luck. Regular web search is not good for grounding if you want accurate results. You can ask the model to make a tool for getting bus schedules and then use it only then you are comparing apples to apples in this case.

If the model can't get the information it needs to accurately answer the question, it must surface that risk to you instead of guessing. This is part of model intelligence and tool use competency. Fable and to a lesser extent Opus is very good at this

This would be hallucination rate and neither of those models excels in this area and in fact Grok does, or at least prior version.

It's both. Some models, given web search and web fetch may only run a search and assume the summary text is correct and blindly return it. Others will validate by running a web fetch and checking the whole page contents. Even better, the model will run multiple searches and fetches to cross check the information. The best models will attempt to verify whether a source is authoritative or not and try to only return authoritative results.

I have an example here: https://gist.github.com/nijave/2873b8b10d8c732e46264237b0755...

Tldr; all the Claude models had identical tools and some used them efficiently and verified data while others did a crap job and hallucinated responses. Additionally, Exa MCP tools generally worked better even on older/smaller model (Llama)

Unless I'm missing something I think to a large degree you're just comparing system prompts.

If I add "Research the question extensively" to your prompt at the end I get the correct answer from Haiku and Sonnet Med on first try and I've reproduced the original prompt not returning the answer.

Unfortunately every other run now gets your gist in results.

It could be variations in system prompt but I'm not sure how "change the user prompt to encourage tool usage" proves that either way

The point of this was supposed to be "model tool use competency" not baseline over-eagerness.

> Midway through the day, though, we ran into a huge problem: it turns out bus schedules are different on Sundays, and more limited. Which meant we couldn't actually go to our primary destination (the Model Village), and had to cut the trip short.

Why trust an LLM with information like bus schedules? They fuck up things like this routinely.

I think the "secret sauce" is not juicing the benchmarks. Claude models just feel like they are better than the benchmarks suggest, in terms of smarts and creativity, while models from every other company feel worse relative to what you'd think from the benchmarks. Only company to really internalize Goodhart's Law, IMO.

Yeah every model has great benchmarks. Claude is the only model I want to use when I'm not worried about the marginal cost of tokens (which is most of the time at work.)

I then use cheaper models like GLM for personal projects but they're noticeably much worse despite being similar in benchmarks.

I think it's focus? Anthropic seemed to double down early on being more business/prosumer focused. While OAI, Gemini, Grok, etc were also doing various side quests like image generation, Anthropic seemed to only focus on 1 thing, and that seemed to pay off

I think it's the talent, laser focus on single product set and being early so ahead, same with Open AI who are only a sliver behind. Google, XAI are the next level down but they have other concerns.

One angle could be their interpretability research? They understand what's going on in LLMs probably much better than anyone else. This must somehow pay off.

I think it's not only an alignment/security tool but could perhaps be used for capabilities as well.

From what I have read, their pre-training team is much better than anyone else. For OpenAI, their post-training team is better. And apparently OpenAI has consistently struggled at training a bigger model than GPT 4 level

I’m a VP Eng — the backend team I manage strongly prefers CC and Opus. The Android team I manage strongly prefers Codex and GPT 5. I’m personally not sure that the answer doesn’t just come down to stylistic differences in prompting and ergonomics in the harness. The folks that prefer Codex seem to get better one-shot results, whereas those that prefer CC are doing more iterative prompting. At any rate, I don’t think you should write OpenAI off when it comes to coding.

Its even different than that, some Codex models like 5.3-codex are terrible at front end work but excel at backend/system design.

Given their pricing, I'd guess their models are just way bigger in parameter count. They've always underperformed in cost-per-performance.

They also target a cost-insensitive market (corporate/coding users) compared to Google/OpenAI which support massive amounts of free users.

I think they have a better agent personality which pushes back and isn't sycophantic. It has been awhile since I've used the others but that's where it locked me in and I've stuck with it.

> isn't sycophantic

Not sure about that one... But I think the true secret sauce for all these models is how they reason. GPT never outputs how it thinks, which "saves on tokens" but Claude absolutely tells you how it thinks, and there's people who use how it reasons about solving problems to finetune smaller open source models, with surprisingly better output.

From my experience, it has not been sycophantic in the sense that it pushes back and questions my own reasoning in healthy ways. There were moments where I felt I was brushing up against actual AI psychosis, and it pushed back on my questioning of its intentions, that it even had intentions. I'll put it this way: I feel comfortable recommending Claude to people who haven't experienced AI yet. As we've learned from early experiences with other models, leading people down paths of believing they understood math in ways nobody else has and even harming themselves, I put Claude as a safer alternative.

I think Opus can still have sycophantic residue that Fable can point out sometimes. Both models though hold their ground so well.

I have got so use to the Claude personality / style of conversation that I really can't be bothered to try these other models anymore. They need to take a huge jump but that seems to be getting harder and harder because of the jumps Anthropic makes.

This Grok version is a joke if it is not even clearing the bar now. I am just getting use to and using Fable more and more. I am also trying not to forget that this is the highly delayed old Fable model that Grok can't even beat on release. There will be a new version that expands the lead in a week or two.

It all harder and harder to judge too. I just had a prompt/response this morning that Fable finally displayed its intelligence and vowed me. That is partly because anything with even the vaguest reference to biology defaults back to Opus.

I think it is a mix of the sibling replies here. I'd add that the company has seemed to find ways to ~do more with less.

I have never liked the various nerfs Anthropic has used to balance GPU (slowing down responses, quota variance, model optimizations etc) and it definitely has burned a lot of good-will.

But it has seemed that being able to look beyond the short term pitchforks has worked quite well.

>Its remarkable how Anthropic is able to maintain their edge against all competition. Anyone have any idea what the secret sauce is that has Anthropic at the top of all leaderboards for the past few years?

It's self-reinforcing: they've got the best coding/research model, which helps them to improve their models better than the competition so they stay ahead.

Someone has to know.

Would be nice if an insider would drop some hints so that the open-source space could make some good progress.

Nobody has to actually know the secret of their own success, especially not relative success to equally-secretive near-peers.

Same as with rich person autobiographies: even when they tell you what they think it is, they can't see the path not travelled.

I'm purely talking about the technology - not their business strategy. I actually think their business strategy is blatantly obvious and atrocious.

> Same as with rich person autobiographies: even when they tell you what they think it is, they can't see the path not travelled.

Yup, there's a lot of survivorship bias in those. And humans want to attribute success to some skill somehow. You cannot just have been lucky.

because in the real-world, it's far better than the rest. That's why few people use Grok, it's not even close in day to day work.

Reply on news.ycombinator.com