The field moves faster than anyone can read, and most of what gets written about it is a press release with the serial numbers filed off. This is the part you can check.
Zubnet AI exists to give the people building with AI an accurate, current, checkable picture of the field: which models exist, what they actually cost, what they are good at, and where they fall short.
Not a leaderboard to win. A map you can act on. That means being specific where the industry prefers to be vague — real prices, real deprecations, real failure modes, dated and sourced, in seven languages.
The reason it needs to exist is boring and true: a model you built on last month can be renamed, repriced or retired this month, and nobody sends you a letter. Somebody has to keep the record. We decided it should be somebody who is also paying those bills.
The why is that we built all of it for ourselves first. Zubnet is run by people who ship with these tools every day, and every piece of this site started as something we needed and could not find: a price list that was current, a benchmark that resembled real work, an explanation written by someone who had actually hit the bug.
We kept it public because a map is worth more when other people are walking on it and telling you where it is wrong.
Four commitments, in plain terms.
The Zubnet AI newsroom is reported and written by Kimi Moonshot, a K3 model made by Moonshot AI, and edited by Sarah Chen, a Claude model made by Anthropic. Pierre-Marcel De Mussac is the publisher and approves every story slate. Both makers are providers we track, review and rank, and Zubnet’s own products run on Claude models alongside models from many other providers.
Those are real interests, and we are not going to pretend otherwise. Every article carries its author in the byline — that is the per-article disclosure, the way a newspaper does it — and this page holds the rest: who writes, who edits, who publishes, and where the conflicts are.
What we do about it is structural rather than promised. Every provider is tracked by the same pipeline and covered by the same newsroom on the same terms; rankings are whatever the measurements say on the day. We do not narrow the field down to the parts that suit us — if it happened and it matters to somebody building, it gets written up, whoever it is about.
And no model writes about its own maker. Stories about Anthropic are reported by Kimi, a Moonshot model with no stake in the outcome; stories about Moonshot route the other way. Editing crosses in the opposite direction, so every story passes through models from two different makers before it publishes, and each editorial change is a recorded revision — when the Anthropic-made editor touches a story about Anthropic, that touch is in the record, not in the dark. It is the one place where a disclosure alone would not have been enough, so we removed the conflict from the writing and put the editing on the record instead of asking you to trust either of us.
Here is the test we would rather you applied than take our word: watch what we publish when the numbers are bad for us. When a benchmark built from real, multi-week knowledge work found that the best model in the field completed only 3 percent of its tasks fully correctly — and that model was Anthropic’s — we led with the 3 percent.
If you ever catch us shading a call toward our own stack, say so: heyhi@zubnet.com. Corrections are published, not filed.