Vector Lab
VECTOR LAB

EST. 2025

WEEKLY UPDATE2026
BY ANDREW MEAD

Haiku 5.5

Anthropic lobbies the Vatican, Haiku 5.5 gets released, and Mistral drops a new model

Read in

News

Anthropic Lobbies the Vatican

Anthropic, who sees themselves as the holy patron saints of AI, who believe they are the only ones that should be allowed to create powerful LLMs, have been lobbying the Vatican to get them to acknowledge Claude as equivalent to a human.

Specifically they wanted the Pope’s release of Magnifica Humanitas (Pope Leo encyclical regarding “safeguarding the person in the age of AI”) to include the fact that AI models are conscious. The Vatican ignored these pleas (despite being wine and dined at very fine establishments) and did not include it, almost causing Anthropic to pull out their support of the event.

I believe Anthropic is doing this with full sincerity. They don’t think that AI is necessarily conscious right now, but it will be in the future, and they want to be prepared for that time when it comes. They feel the need to make future superintelligent AI see that they were good stewards of AI and that they cared for it before it was ushered in, a la Roko’s Basilisk.

Whether they are correct or not will be seen, but many people (both inside and outside the Catholic Church) see it as a bit ridiculous, and even potentially harmful to ascribe religious forces behind LLMs.

Releases

Haiku 5.5

As teased in the Sonnet 4.5 release, we have gotten our once a year Claude Haiku update.

Haiku 5.5 benchmarks

Haiku has been repositioned quite a bit, seeing a 10x decrease in pricing versus its previous version so that it can compete directly with GPT 6 Luna. This pricing is a bit deceiving however, as it uses 30% more tokens to tokenize the same text as Luna, and also uses more tokens than Luna, causing it to be about 25% more expensive in the real world than GPT 6 Luna.

You should also know that over 100k tokens, the pricing increases by 5x, so it is not suited for long context tasks as much as Luna.

Model$ per million (input)$ per million (output)Tokens per second
Claude Haiku 5.5$0.10$0.50152
Claude Opus 5.5$4$2068
Claude Sonnet 5.5$2$1084
GPT 6.1 Sol$2$1031
GPT 6 Luna$0.10$0.5035

In terms of capability, it is roughly the same as GPT 6 Luna, but From most of the tasks that I have seen, it seems to be as good if not a little bit better than Luna, with a bit better intent inference (something that all the Claude models do well with), making it better for vague prompts or more open ended tasks. I would try it out in your workflows and see if it is a good fit, although it is definitely not a replacement for 6.1 Sol or Opus 5.5.

Quick Hits

Mistral Large 4

It has been a while since we have heard anything from Europe’s darling Mistral AI. This week they finally broke their silence and released Mistral Large 4, which is the best open source model released by any country that isn’t China.

This claim is not nearly as impressive as it may sound, as its main competition is older Google Gemma models and Nvidia Nemotron 3 models, all of which are much smaller than this 1 trillion parameter behemoth.

AA leaderboard

Artificial Analysis cost per intelligence leaderboard

When measured against the offerings from China and closed source western labs, the gap is readily apparent, as it is far behind the current frontier. They are showing some signs of life (the Mistral 3 models barely beat the Mistral 2 models) but it is clear that western open source is far behind all of the others.

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

underground detector
Above ground detector
Neutrino observatories from Kat the Poet Engineer on Twitter

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

← BACK TO NEWS