#1 in a very close race is way less useful when you have to walk on eggshells to avoid triggering censorship (“safeguards”) that either refuse or knock it down to another model. I’ve almost completely stopped using Claude (except some legacy workflows) for this reason, reliability matters more than scoring 61 instead of 57. To me Claude is the most compromised and unreliable model (between the censorship and the id checking - which I have not experienced personally), it’s not worth whatever slight benchmaxxing they did for the latest release.
Every time an online chatter (e.g. "limits are better", "model is better") makes me to reevaluate my principle of never paying Anthropic, I go to the model card, which strengthens my belief in the principle.
Why is Anthropic is so hell-bent on this auto/silent downgrade? Do they have a single user who prefers an auto-lobotomization instead of a refusal? Have they learned nothing from the backlash the first time?
Working on dimensionl reduction algorithms, I hit it all the time. I'm also trying to port related protocols from single-cell transcriptomics to collective intelligence systems (working with people x reaction matrices as analogous to single-cells cell x gene matrices.
Something between single-cell work and advanced nonlinear DR methods (perhaps used in alignment work?) it always flags me
I've hit it with intensely benign things; like asking it to make me a web-based client-side word game. I am guessing it saw the dictionary and pattern matched on various words, though ultimately it provided no explanation for why it triggered safeguards.
If you even broach language related to biology you’ll get rerouted. I was presenting data in a grid and referred to a grid cell, Fable saw the word “cell” and safeguards kicked in
I was profiling a slow machine the other day, and triggered the safeguards.
I've been saying this a lot lately, but it doesn't bites you until it bites you.
The more you use the clanker as a general purpose fix-it tool (goodbye manual NeoVim configuration, you will not be missed!), the more you will find yourself bumping into these safeguards.
Yep. I cancelled my Claude Max subscription 2 weeks ago after feeling like Anthropic was doing everything it could to fuck with my day to day. Their lead would have to become significant for me to ever go back.
The chart shows max effort, used mostly by price-insensitive enterprise users. At medium effort it drops to almost half K3’s cost, and is probably sufficient for 95% of coding tasks.
Very interesting that one of the components is "AA-Omniscience Index"
AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer.
This seems to be a good proxy for param size/density and the ranking breaks down as such: Claude Fable 5 (with fallback), Gemini 3.1 Pro Preview, Claude Opus 5 (Max), Grok 4.6 (high), Gemini 3.6 Flash, GPT 5.6 Sol (Max)
I've thought for a while that Gemini 3.x has 'big model smell'
I'm not so sure. Especially with 3.6 Flash being 24 to Sol's 22. The 3.x Flash models were thought to fit on a single TPU 8i and it's the odd one out in that list at a whopping 234.7t/s compared to Sol's 64.4t/s and Opus's 56.3t/s. Even 3.1 Pro is a much higher 113.9t/s than the others.
I think AA-Omniscience Accuracy follows your expectations better. An ultra size Fable at 61%, followed by large frontier models like Sol, 5.5 and Opus. With Flash being up there. I assume because Gemini is more focused on general knowledge to operational cost in particular, rather than getting the highest scores in coding benchmarks. If you go to Domain Score you'll see that their only real penalties are in Software.
Gemini 3.1 pro is really good for knowledge tasks. Google has done well there. And image analysis with Gemini flash 3.6 is solid. It’s just anything coding or agentic they fall short.
Indeed, you can filter the graphs to see these the values for alternative reasoning settings of the models. Opus 5 High reasoning scored 59 on the index (exactly the same as GPT 5.6 Sol Max), and costs $1.06 per task (vs $1.04 Sol Max). So these seem essentially equivalent on both metrics.
It shouldn't be surprising OpenAI does have the most compute out of all the major labs. The only reason why Anthropic models are expensive is they are the most in demand models in the world and Anthropic is fighting for compute. The only way to you limit demand for your model is increasing API pricing this is also why Anthropic probably has great margin and probably is profitable compared to OpenAI.
I hate to disagree with Broadcom Throwaway himself but it's unlikely that the OpenAI Jalapeno ASIC has been deployed yet. It takes 6-12 months to test, develop software, ramp production, etc.
I'd be curious to see the results, especially with some models having 1.5m and 2m context sizes, if the first 75% of the context was filled with unrelated info.
Are you using Claude Code/CoWork or an API client? I’m curious if it has different training that makes it more effective with specific instructions/ tool calling methods that are only implemented in official harnesses.
I'm curious about this too, and it's difficult to get any information about this given everyone has different setups, workflows and use-cases.
I bizarrely had Opus 4.8 this week (in pi.dev within a podman container, using openrouter) start installing various python packages (and uv!) within the environment (not as root) when I asked it to code review some fairly basic Rust .rs files that were generally stand-alone (it did very nicely work out and write some stubs for them to build them and work out how they worked).
It only gave up with the weird Python installing stuff when it discovered one of the Python packages needed Tensorflow.
It seems pretty focused and persistent in continuing its initial approach, and I'm wondering if I need to alter some instructions / initial prompts to rein it in a bit...
Why is Anthropic is so hell-bent on this auto/silent downgrade? Do they have a single user who prefers an auto-lobotomization instead of a refusal? Have they learned nothing from the backlash the first time?
Something between single-cell work and advanced nonlinear DR methods (perhaps used in alignment work?) it always flags me
I've been saying this a lot lately, but it doesn't bites you until it bites you.
The more you use the clanker as a general purpose fix-it tool (goodbye manual NeoVim configuration, you will not be missed!), the more you will find yourself bumping into these safeguards.
At least two models (GPT-5.6, Kimi K3) match its score (~1-2% diff) for half the cost.
AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer.
This seems to be a good proxy for param size/density and the ranking breaks down as such: Claude Fable 5 (with fallback), Gemini 3.1 Pro Preview, Claude Opus 5 (Max), Grok 4.6 (high), Gemini 3.6 Flash, GPT 5.6 Sol (Max)
I've thought for a while that Gemini 3.x has 'big model smell'
I think AA-Omniscience Accuracy follows your expectations better. An ultra size Fable at 61%, followed by large frontier models like Sol, 5.5 and Opus. With Flash being up there. I assume because Gemini is more focused on general knowledge to operational cost in particular, rather than getting the highest scores in coding benchmarks. If you go to Domain Score you'll see that their only real penalties are in Software.
No wonder why Tibo can afford to hit the reset button liberally.
Like 96% vs 93% or something
"Not fair! They distilled Opus 5!"
Gets confused by permission prompts, cannot debug a failing test it caused (Opus 4.8 got it right after, without tens of rounds "thinking").
I bizarrely had Opus 4.8 this week (in pi.dev within a podman container, using openrouter) start installing various python packages (and uv!) within the environment (not as root) when I asked it to code review some fairly basic Rust .rs files that were generally stand-alone (it did very nicely work out and write some stubs for them to build them and work out how they worked).
It only gave up with the weird Python installing stuff when it discovered one of the Python packages needed Tensorflow.
It seems pretty focused and persistent in continuing its initial approach, and I'm wondering if I need to alter some instructions / initial prompts to rein it in a bit...