Open Source AI Models Cost Far Less Than Closed Counterparts on Output Tokens

If we project this out, and the next Fable costs nearly 2x Claude Fable 5 per token, very few people will use it.

Which leads me to wonder once again whether we will really need all these data centers.

this won’t end well

@GaryMarcus Crazy thinking that Qwen, Minimax m3 are at most 10 points behind on aggregate benches like artificial analysis, but that much further ahead in bang for buck.

There is better examples out there but for instance on FrontierCode. Fable 5 at medium effort is roughly $7 per task with a score around 18%. Opus 4.8 at max effort costs more, around $10.50 per task and only scores about 11.5%

So in that comparison Fable 5 is both cheaper and substantially better as it doesn't need as many tokens.

@suavecito585 but are those increases small linear increases? giant jumps?

the data i presented on cost were shown in WSJ and I presume legit

@GaryMarcus There’s also token efficiency sir this model in some cases only uses 1/3 reasoning tokens compared to Opus 4.8 as it doesn’t need as much CoT. So that factors in a lot of the cost to actually use it

@JasonBotterill data on this?

@MoonlitMonkey69 @GaryMarcus bang for buck is the only metric that matters once you've been paying for these for a year. m2.7 voice + the right harness punches way above its price tier

@GaryMarcus mythos class pricing is the real singularity

@GaryMarcus Anthropic burns cash at ~$5b/yr while training models that cost $100m per run. their fear wasn't a bug. it was the only rational response to the ledger

@GaryMarcus Your just a fucken sad hater

@GaryMarcus wait so the input cost curve is flat but output is vertical?

@GaryMarcus It was never going to end well, not with these players.

@GaryMarcus Guess they figured out how to exponentially inflate their budget too.

@GaryMarcus that's...not true. it leads in practically every evaluation and many of them by a long shot.

@GaryMarcus $5B annual burn. the data center will be empty by the time the model actually works. cope.

@GaryMarcus It's time to nationalize them. If only government and big companies can afford it, we have nothing to lose nationalizing them.

@GaryMarcus I mean Qwen3.6 35B A3B roughly matches GPT-5's performance yet it is ~10x cheaper than GPT-5. That means it took ~8 months for the cost of that level of capability to fall 10x

@GaryMarcus surely it just plays out like this?

@GaryMarcus https://gizmodo.com/palantir-ceo-says-bernie-sanders-will-get-regret-only-wanting-50-public-ownership-of-ai-companies-2000770102