Added September 5: Astra now has five measured thinking levels: low, medium, high, xhigh, and max. Its numbers come from the September 4 archive of the same benchmark used here. All other models keep their September 2 values. Ultra has no comparable measured point.
Updated September 2: Fable 5.1's five measured effort levels are now on the chart after its September 1 launch. The “up to 45% cheaper” line is a best-case estimate for highly agentic work that repeatedly reads cached context. Anthropic estimates about 25% savings for typical workloads; input and output remain $10 and $50 per million tokens.
I like this chart because I can understand it in about five seconds. I do not have to read a hundred posts about why one model is supposedly destroying another model this week.
Most connected dots are different amounts of thinking. The higher the dot, the smarter it tested. The farther right it goes, the more it costs.
What I see
Fable 5.1 changes the top of the chart. At high effort it essentially ties Opus 5 xhigh while costing about 21% less per task. At max it becomes the smartest measured point, 2.6 points above Opus 5 max, but it also costs about 58% more per task. The best value is not automatically the highest dot.
Why I retest instead of re-reading
Fable's savings come from reuse, not cheaper words. Input and output still cost $10 and $50 per million tokens. Cache reads dropped 75%, from $1 to $0.25, so workflows that keep reusing the same long context benefit most. Anthropic estimates about 25% savings for typical work and up to about 45% for highly agentic work.
That is the practical argument for keeping operations model agnostic. If switching models is a config change instead of a rewrite, a pricing special turns into free capacity to get real work done. If it is a rewrite, the special expires before the argument about it finishes. Making the chart is the easy part, using it is the point.
What changed?
I added Fable 5.1's full five-level measured curve and rechecked every existing model on the same Artificial Analysis snapshot. Fable now sets the capability ceiling, while its high setting looks like the more practical value point.
I also made the chart interactive. Hover over a model to bring its line forward. Click one to keep it highlighted while you look at the details.
The one gray Gemini 3.1 Flash-Lite dot stays because it gives me a useful comparison for AI Health Export.
Want to see the methodology?
Where these numbers come from, what counts as measured versus estimated, and a running log of what changed on every update, including the one where a model's cost moved 12% with no price change behind it. How the model chart is built →
Why am I looking at this?
I am thinking about Luna and Gemini for AI Health Export after version 3.0. I am not changing anything yet. I just want to live with the tradeoff for a while and see what starts to make sense.
Technically, I am already using Luna in one real place: Hermes, my personal AI agent. That is part of why I keep looking at it.
The chart does not need to answer everything. Later, I can try both models inside the real app and see which one feels faster and gives better answers. For now, this gets me 90% of the way there.
The bigger idea
I never want a project trapped inside one AI model, whether it is an app, a website, an internal tool, or something I am building for a client. So I keep the important parts in a set of mostly Markdown documents: the project history, decisions, instructions, rules, and step-by-step playbooks that any AI agent can read. That is the basic idea behind AgentOps, a system I built to keep my work organized and portable. Claude, GPT, Gemini, or whatever comes next can pick up those same documents and understand how the project works. If something better comes along, I can switch models without starting over. I think everyone doing serious work with AI should build some version of this.
If your own AI setup feels more confusing than helpful, that is exactly the kind of mess I like untangling.