For a year, the honest reason founders paid flagship model rates on their hardest agent paths was simple: only a flagship finished the job. Long, multi-step professional tasks — the kind where an agent has to hold context across a dozen tool calls without losing the plot — were where the cheap tiers fell apart. So you routed those paths to the most expensive model you had and told yourself it was worth it.
On July 9, OpenAI shipped numbers that quietly retire that reasoning.
The bar, and who just cleared it#
The benchmark is Agents' Last Exam — an evaluation of long-running professional workflows across 55 fields. It is worth caring about for one specific reason: long-horizon completion is exactly where real agent products die. The demo finishes. The production task, three steps deeper, does not. A score here is a proxy for "does this thing actually complete a multi-step job," which is the question your users are really asking.
Here is where the new GPT-5.6 family landed, against the prior frontier as the reference point:
- GPT-5.6 Sol (flagship): 52.7, and 53.6 at max reasoning
- GPT-5.6 Terra (balanced): 50.4
- GPT-5.6 Luna (cheap): 50.3
- GPT-5.5 (last generation's mid-tier): 46.9
- Claude Fable 5 (the prior frontier): 40.5
Read the top line and you'd say "Sol sets a new high." True, and not the story. The story is three rows down: Luna and Terra — the cheap and balanced tiers — clear last year's frontier by roughly ten points on the axis where agents fail. And per OpenAI's own estimate, they do it at around one-sixteenth the cost of running Fable 5.
The premium you paid for a flagship "because only it finishes the job" no longer describes this benchmark.
What actually changed#
Not "Sol is good." What changed is the cost-to-completion ratio for long-horizon work.
Twelve months ago, finishing a hard professional task was frontier-only, and frontier meant flagship pricing — Fable 5 lists at $10 in / $50 out per million tokens. Today, a model at $1 in / $6 out (Luna) posts a higher completion score than that flagship did. The "frontier tax" — the surcharge you accept because capability and price were welded together at the top of the ladder — just came unwelded on this eval. Capability for multi-step completion slid down the price ladder faster than most routing configs have been updated to notice.
That matches the macro picture. This is the same 48 hours that gave founders three sub-frontier models undercutting the flagships and reset the intra-vendor tier ladder. The price war isn't just about cheaper tokens; it's about cheaper completion.
The caveats, stated plainly#
This is one benchmark, from the model's own vendor, published the same day the models shipped. "Estimated cost" is doing real work in the 1/16 figure. "Professional workflows across 55 fields" is a broad average that hides per-field variance — your domain may not track the mean. Benchmarks get gamed, and self-reported ones most of all. If you migrate production traffic on the strength of a launch-day tweet, that's on you, not on the number.
So don't read the decimal. Read the direction — and the direction is not subtle. The capability required to complete a long-horizon professional task is falling down the price ladder quarter over quarter. Whatever routing assumptions you set three months ago are probably stale.
What to actually do#
The wrong takeaway is "switch everything to Luna." The right one is: stop assuming your hardest agent path needs your most expensive model, and go measure.
Concretely:
- Re-run your own long-horizon eval across a cheap, a mid, and a frontier tier — your real multi-step tasks, your tools, your context. Vendor benchmarks are directional; yours is decisive.
- Price by completion, not by token. Cost per finished task = output tokens the model emits to complete it × the output rate, plus cache-adjusted input. A chattier cheap model can lose to a terser mid one. Log output-tokens-per-task per tier and multiply.
- Route by the number, not the name. Let measured completion rate pick the model for each path.
- Reserve the frontier for the paths that still fail without it. Sol beats Fable 5 by 13.1 points at max reasoning — the flagship gap is real, just narrow, and now the exception rather than the default. This is the same logic that made Sonnet 5 the cheaper default for most agent work; GPT-5.6's cheap tiers extend it to long-horizon tasks.
The benchmark that most agents were failing 97% of the time a few months ago is the same one where a dollar-per-million-token model now clears the prior frontier. That's not a reason to trust the score. It's a reason to re-price your stack — because the model you're overpaying for the privilege of "it's the only one that finishes" may no longer be the only one that finishes.
