Context:
https://m1astra-mythos.pages.dev/
Any model described as being a size beyond Opus, which is expected to be called Claude Mythos or Capybara but might be called any other name, will be considered for the purpose of this market
The first version of Claude Mythos publicly available in the api is used for this market's resolution. We will use the "most standard price" if there are multiple. eg, if price above 200k/Mtok increases, we resolve based on price before this price increase. if a /fast version of the model is twice as expensive, we use the normal version. if different reasoning efforts are different prices, we use the Claude Code default for claude max 5x users. If a cheaper rate is offered for cybersecurity companies, or for students, the most "typical" rate will be used for resolution.
If no such claude version is released and available in the API before 2027, this market resolves N/A
1,000Pessoas também estão operando
"Claude Mythos Preview will be available to participants at $25/$125 per million input/output tokens (participants can access the model on the Claude API, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry)."
@Bayesian doesn't the market resolve based on this? There's only one price available for the model in the Claude API (and based on Anthropic's announcement they don't plan to release the model publicly outside of the glasswing program).
@Bayesian I see, I think the argument for resolving based on the 25/Mtok price is straightforward:
That's the price the model is available at in the Claude API
It is the "most typical" price
The market doesn't mention any other requirements which are relevant here
Some things that might square this situation more with what you were thinking when you created the market:
Anthropic is rolling out the model to about 50 large customers based on their post. That almost certainly includes most or all of their current largest API customers, and it seems highly likely that those customers will use more than the $100M of credits they're giving away and continue using the model at the $25/Mtok rate.
Even if this is ostensibly a cybersecurity program it seems unlikely that these large enterprise customers will be strictly limiting themselves to cyber use cases. For all intents and purposes this mostly seems like a normal API release of the model which just happens to be limited to large customers.
They mention that they won't be releasing the model publicly, so outside of using the $25/Mtok rate there likely won't be any way to resolve this market (and it would n/a)
@2b3o4o just so we’re on the same page, did you mean $125? $25/$125 per million input/output tokens means that the output price is $125/Mtok?
But yeah basically this is not my sense of the meaning of ‘available’
@Bayesian Oh... yeah I misread the market ☠️
I think the arguments I made still apply but since this now means I would be losing money with the resolution I will stop here and happily accept your definition of 'available' lol
@2b3o4o It's not clear to me that they won't be releasing Mythos publicly? They say they don't plan to release Mythos Preview, but this could mean they want to make it available for patching up cybersecurity issues, want to release a Mythos model eventually but are not finished with it, and think the preview model is good enough to protect against the issues they anticipate with the finished model. It's possible there is info I haven't seen disproving this interpretation, but I think there's still a reasonable chance that there will be a public release by 2027.
@speck that's true. I guess you're responding to me claiming that there's no way to resolve if we don't use the pricing they announced here? That probably was incorrect but I do think we'll be in a weird situation if an eventual public release comes with different pricing
@2b3o4o Exactly, and yeah I'm expecting resolution to be at least a little weird if it happens for that reason.