When will Anthropic release Claude 4 and make it generally available?
When Anthropic released Claude 3.5 Sonnet on June 20, they stated that "Our aim is to substantially improve the tradeoff curve between intelligence, speed, and cost every few months." Since the tradeoff curve is made up of different model sizes, this could be read as Anthropic planning to release a new model family every few months. However, this is not clear, since a delayed release within the same model family can also improve the tradeoff curve.
For the purposes of this question, Claude 4 must be a frontier model that Anthropic claims to be more intelligent or powerful than any previous Anthropic model.
For example, if they release Claude 4 Sonnet first, as they did with 3.5, and 4 Sonnet is superior to their previous best model (e.g. 3.5 Opus), that resolves this question YES. However, if they release 4 Haiku first, and it does not surpass 3.5 Opus, that does not resolve this question. I'll defer to Anthropic's characterizations of their models for this purpose.
If Anthropic abandons their number-based naming scheme, I'll resolve it to when "Claude 4 in spirit" comes out, if it's clear when such a model is released. But if it's not clear, it will resolve to the last time option.
Claude 3.5 Sonnet was released 6/20/24.
Claude 3 was 3/14/24.
Claude 2 was 7/11/23.
Anthropic has promised Claude 3.5 Opus/Haiku before the end of the year (link below).
That they could have Claude 4 by 9/30 (end of 3Q) is insane. They can get to the next gen model in 6 months (3/14-9/30)?
You’d figure if they could so easily, they would have released Opus/Haiku with Sonnet at the June release.
Even 4Q would be impressive. But then they squeeze their Haiku/Opus releases. Opus 3.5 figures to still be tops on the market and “clearly” better than current market leader Sonnet 3.5. And Claude 4 has to be unambiguously better than Opus 3.5. Why even bother was Opus 3.5 if Claude 4 is right around the corner? Maybe pull a switcheroo and release Claude 4, say October 1, then backfill and release Opus/Hakiu 3.5 by Jan 1? Possible, I suppose
It’d be interesting if we saw evidence it was getting easier to scale to the next-gen model. However, OpenAI treading water in its releases (a bunch of “4” models not even clearly better - at the general level - than each other) seems to indicate the next-gen difficulty is staying the same or slightly harder. Or maybe Anthropic has not only erased OpenAI’s first-mover advantage, but has created one of their own in such a short time?!
To me, a Claude 4 release of March/April 2024 seems most probable, +/- 1 month.
Getting it Feb (or sooner) or May (or later) definitely tells us a lot about how hard scaling to the next level is going.
My median is ~January because Anthropic said they would advance the pareto frontier every few (3-4?) months. And given growth in chip sales my guess is that Sonnet is a much smaller compute scale than what Anthropic can muster in the next few months.