Welcome to the first post in a weekly series that takes the main AI model leaderboards and explains them in plain language for people who do not follow this world every day. Each week we pick the charts that are most relevant right now and, for each one, cover what it measures, what it does and does not capture, and what the numbers actually mean.
This week three charts stood out, and together they answer three different questions: which model is the smartest, which models are actually getting used, and which models are the hardest to trick. We chose them because the race at the top of the capability charts is unusually tight, real-world usage just saw a clear shift, and security is the dimension that is easiest to forget when picking a model. No single model wins all three.
All figures are a snapshot from 16 June 2026, and each chart links to the live version so you can check it yourself.
The smartest models
The first chart is the Artificial Analysis Intelligence Index, one of the most widely cited measures of raw model capability. It is a single score built from nine separate tests covering reasoning, coding, knowledge, and following instructions. A higher number means the model did better across that whole set. We show one entry per model, matching the headline chart on their site. Keep in mind that a benchmark average is a useful proxy, not a promise of how a model will do on your specific task, and Artificial Analysis refreshes the index over time, so treat the scores as a current snapshot rather than a fixed grade.
Source: Artificial Analysis Intelligence Index. Data as of 16 June 2026.
The race at the top is close. Claude Fable 5 leads at 60, ahead of Claude Opus 4.8 at 56 and OpenAI’s GPT-5.5 at 55. Below them the table is tightly packed, with models from Anthropic, OpenAI, Google, Alibaba, and MiniMax sitting within a few points of each other. Google’s Gemini 3.5 Flash is the one to watch: it has reached the top tier at 50 and is much cheaper to run than the models above it. And MiniMax-M3, which tops the usage chart below, is the highest-scoring open-weight model here at 44, a reminder of how close the open models have come to the leaders.
What gets used through OpenRouter
The second chart comes from OpenRouter, a service that sits between apps and dozens of AI models and routes each request to whichever model the developer picked. Usage is measured in tokens. A token is just a chunk of text, roughly a short word or part of a word, and it is the unit these models read and write. More tokens means more use.
It is worth being clear about what this does and does not show. OpenRouter only counts traffic that passes through OpenRouter. It does not see usage that goes straight to a vendor’s own API, such as calling OpenAI, Anthropic, or Google directly, which is almost certainly the larger share of all AI use, especially inside big companies. So read this less as total market share and more as a window into one active slice: independent developers and the apps built on OpenRouter, including coding tools and chat and roleplay products, where switching between models to balance quality and price is normal.
Source: OpenRouter Rankings (top models, weekly). Data as of 16 June 2026.
Within that slice, MiniMax M3 is the most-used model this week, with 4.56 trillion tokens routed in seven days and growth of about 58 percent week over week, enough to put it just ahead of DeepSeek V4 Flash. The top of the list is dominated by lower-cost models, many of them open and mostly from Chinese labs. Anthropic’s Claude is the most-used of the big Western names here, and Claude Opus 4.7 climbed sharply this week. The pattern is clear: for a lot of the work that runs through OpenRouter, teams care less about topping a benchmark and more about a good-enough answer at a low price.
The hardest to trick
The third chart looks at something the first two ignore: security. It comes from F5 Labs, whose CASI score, short for Comprehensive AI Security Index, measures how well a model resists attacks that try to make it misbehave. Two common attacks are jailbreaks, where someone talks the model into ignoring its own rules, and prompt injection, where hidden instructions are slipped into content the model reads. A higher score means the model held up better against F5’s security team. This is one security team’s tests on a limited set of models, so a model being absent is not a clean bill of health, just untested here. The board is updated monthly, and these scores are from F5’s early-June run, the freshest security read available.
Source: F5 Labs CASI Leaderboard. Updated 3 June 2026.
Anthropic’s Claude models hold the top three spots, with Claude Haiku 4.5 in front at 93.6, followed by Claude Opus 4.8 and Claude Sonnet 4.6. Here is the connection worth drawing across the charts: most of the models topping the OpenRouter usage list, such as MiniMax M3 and DeepSeek, are not on F5’s security board at all, and the cheaper Qwen models that are on it sit well below Claude. Popularity and low price clearly drive what gets used, but they tell you nothing about how easily a model can be tricked. If you are putting AI into a product, security is worth checking on its own.
What it adds up to
Three charts, three different leaders. The smartest model this week (Claude Fable 5), the most used on OpenRouter (MiniMax M3), and the hardest to trick (Claude Haiku 4.5) are three different models from two different makers. The standout move was MiniMax climbing to the top of usage, a sign of how fast adoption can shift toward cheaper options, at least among the developers who route through OpenRouter. The broader point: there is no single best model, only the best model for what you are trying to do, at the price and the risk level you can accept.
We will be back next week with whichever charts matter most then, and why. If there is a chart or a question you would like us to cover, let us know.
Share this article
