I have a confession that sounds backwards: the smartest model is usually not the one my agents run.
Quick background so you know where this is coming from. I run a fleet of Hermes agents on my own Mac minis, around the clock. They pull news feeds every few hours, they publish and check this site, they watch my other projects, they message me on Telegram when something breaks. That is real agent work: thousands of small jobs a day, most of them routine, all of them burning tokens. And after a few weeks of running the newest releases side by side, two models have earned the everyday seats: Grok 4.5 and GPT-5.6 Luna.
What agent work actually needs
Benchmarks measure a model answering one hard question really well. An agent fleet is the opposite problem. It is a model answering ten thousand easy-to-medium questions fast, cheap, and without falling over. When an agent runs all day, three things matter way more than a leaderboard crown:
Speed, because a slow model turns a 2-minute job into a 20-minute job and the queue backs up. Cost, because at fleet volume the token bill is the whole ballgame. And context, because agents drag logs and files and feeds around with them everywhere they go.
Smarts still matter. But there is a floor of smart that most agent jobs need, and once a model clears that floor, every extra point of intelligence you are paying for is mostly decoration.
Grok 4.5: the fast workhorse
xAI shipped Grok 4.5 on July 8 at $2 per million input tokens and $6 out. Musk's pitch was that it is "roughly comparable to Opus 4.7, but much faster", and while I always take vendor framing with salt, the "much faster" part matches what I see. It finishes agent tasks in fewer steps and does not meander.
Let me be honest about the scoreboard too, because I do not do hype here. On the hard coding evals, independent testing puts Grok 4.5 below Claude Fable 5 and GPT-5.5, with numbers like 53% on DeepSWE 1.1 against Fable 5's 70%. If I am doing surgery on a gnarly codebase, Grok 4.5 is not the model I reach for. But my agents are not doing surgery at 3am. They are parsing feeds, drafting summaries, running checks, and filing reports, and for that work Grok 4.5 clears the smart floor easily at a fraction of the flagship price. It has been quick, steady, and cheap enough that I stopped thinking about it. That is the whole point.
GPT-5.6 Luna: the volume king
One thing worth being precise about, because I keep seeing it written wrong: Luna is not OpenAI's flagship. The GPT-5.6 family launched in July as three tiers, Sol at the top, Terra in the middle, and Luna as the fast tier. Sol is the one with the frontier benchmark scores.
Luna is the one I actually deploy. It runs $1 per million in, $6 out, with a context window around a million tokens, and that combination is kind of absurd for agent work. A million tokens of context means my agents can carry an entire day of feed data or a whole log history into one call without any chunking gymnastics. At that price I do not have to ration it. OpenAI claims Luna punches way above its price tier, and while I would attribute those claims rather than swear by them, my own boring test is simpler: I have not caught it botching the jobs I give it, and the bill at the end of the week makes me smile.
How I actually split the work
My setup is not complicated. The routine lanes, feeds, summaries, site checks, message triage, run on Luna and Grok 4.5, roughly by which shape of job each one has been better at. Luna gets the big-context grinds. Grok 4.5 gets the quick agentic runs where step efficiency matters.
And when a job really matters, judgment calls, tricky code, anything customer-facing, I still escalate to a frontier model and pay the toll. The fast models did not replace the smart ones. They took the 90% of the workload that never needed the smart ones in the first place.
If you run agents yourself, that split is the actual lesson, more than any specific model name. Match the model to the job, not to the leaderboard. Model names will rotate again by fall, and there are already rumors about what Anthropic ships next. The principle stays.
One last thing since we are talking about agents running unattended all day: whatever models you pick, they are only as safe as the credentials you hand them. I wrote about running agents 24/7 without leaking keys and why my keys stay local, and both apply double at fleet scale.
Fast, cheap, and good enough turns out to be the winning combo for agents. Not the headline benchmarks. The models doing my work while I sleep are the two nobody brags about at dinner parties, and honestly, that suits me fine.