Open models reach two of the frontier's three tiers

You read that open models finally caught the frontier, and you drew the sensible conclusion. You’re no longer stuck with one AI vendor. If a lab raises its prices or yanks a model, there’s now a downloadable equivalent you can run yourself. The lock-in problem, more or less solved.

That conclusion held while “the frontier” meant one thing. As of last week, it doesn’t.

On July 9, OpenAI previewed its new model not as one thing but as three, sold at three prices. Luna at $1 per million tokens, Terra at $2.50, and Sol at $5. Sol isn’t a bigger Luna. It’s aimed at the hardest work: multi-hour agent tasks, the toughest coding, biology, cybersecurity. The developer Simon Willison, who wrote up the family the same day, and OpenAI’s own preview both describe Sol as a different class of model, not a faster one. It runs at 750 tokens per second on Cerebras, a specialized inference chip, and it burns enough reasoning tokens per task that its real cost can run two to three times the headline rate. So the frontier that used to be a single line is now a ladder with three rungs.

Here’s why that matters for the way you’ve been thinking about open models. An open-weights model is one whose full internals a lab publishes, so anyone can download it and run it on their own servers instead of calling an API. The reason people started saying the frontier was “caught” is GLM-5.2, an open model released under a permissive license by the Chinese lab Z.ai in June. It topped the Artificial Analysis Intelligence Index, a benchmark that blends many tests into one score, and placed second on a leading coding leaderboard. On the aggregate, it matched the best closed models. The freedom to walk away from any one vendor looked won.

Split the frontier into tiers and that win springs a leak, and the leak is in one exact place.

Open models give you a way out precisely on the work where being locked in costs you nothing, and no way out on the work where being locked in costs you the most.

Look at where the parity is real. GLM-5.2’s wins are concrete and they cluster in one place: 39% F1 on a cybersecurity task where Claude scored 32%, a jump on Terminal-Bench, a lead on a frontend design leaderboard, all at roughly a sixth of the closed-frontier price. That’s Luna and Terra territory. Routine generation, constrained retrieval, high-volume pipelines with good scaffolding around them, vertical tasks where a specialist model already wins. This is the work that’s commoditizing anyway. Three labs can do it, the price is falling, and switching between providers is a config change. You have real freedom to move here, and moving costs almost nothing because everyone’s roughly equal.

Now the top rung. Sol-class work is multi-hour orchestration, the reasoning-dense frontier tasks, the ceiling of what a product can attempt. That’s the tier where lock-in actually bites. It’s where you’d most want a fallback you control, for the day one vendor’s rate limits or politics get in your way. And going by what’s shipped, it’s the tier with no open equivalent to switch to. GLM-5.2 plausibly matches Luna or Terra. It very likely does not match Sol.

That inverts the whole pitch. The freedom to switch is most abundant on the cheap, crowded work and absent on the scarce, expensive work. You got the escape hatch, but it opens over the ground you’d never mind being stuck on.

There’s a quieter casualty in this too. The sensible build strategy that came out of the parity story was to use a general frontier model for breadth and open specialists for constrained tasks. That advice assumed “general frontier” was a fungible layer you could match. If the top of general frontier is Sol and Sol is single-vendor, then the breadth layer of that stack is a sole-source dependency dressed up as a choice. You have portability across your specialist tasks and none across the tier that sets your product’s ceiling.

The fair objection is that this is a timing artifact. Open models have trailed closed ones by roughly seven months, and the gap has been closing, not widening. On that read, today’s Sol is next winter’s download, and Sol-class parity is just a matter of waiting.

That objection measures one axis. The seven-month lag tracks catch-up over time, along a single line. A tiered frontier adds a second axis the lag can’t see: which rung. Open models can be closing the time gap to last year’s frontier while closed labs open a new rung above the one being matched. Catching up in time and staying a tier behind are fully compatible. If that’s the steady state, “parity” becomes a lap counter on a track with no finish, always real and always one rung short of the top.

I have to be straight about the soft spot in all this. No one has published a head-to-head of GLM-5.2 against Sol on multi-hour agent tasks specifically, and Sol itself is still a preview. The claim that open models match the lower tiers but not the top is the likely shape, not a proven one. If a benchmark cycle shows an open model matching Sol-class orchestration, not the aggregate index but the actual hard agent work, this collapses back into an ordinary lag and the old parity story absorbs it.

But the direction it collapses toward is the real question, and the two options are strategic opposites.

One is that closed labs keep adding rungs and open labs keep reaching the tier that just stopped being the frontier. Optionality stays real and stays capped below the top. Any product whose ceiling is Sol-class work carries a single-vendor dependency no open fallback can cover, and the honest move is to name a third category in your stack: the tasks with no portable equivalent, priced and governed as sole-source risk rather than filed under “we can always self-host.”

The other is that open labs never chase Sol at all. They own Terra-class economics instead, good-enough capability at a sixth of the cost, where most production volume actually lives. Then one rung below the top isn’t a lag. It’s a chosen position, defensible and profitable, and the missing top rung is a segment nobody’s contesting.

Either way, the reflex to treat open weights as blanket insurance is the thing to drop. Go through your own roadmap and mark which work sits on the top rung. That’s the work an open model can’t yet insure, and knowing which of your bets depend on it is worth more right now than knowing the aggregate score anyone topped this month.