The flagship model is becoming a routing policy
OpenAI launched GPT-5.6 yesterday as a family: Sol for the frontier, Terra for balanced everyday work, and Luna for speed and cost. The headline benchmarks will get compared, screenshotted, and obsolete within a few weeks. The product shape is more interesting than the scores.
There is no longer one default model with a smaller budget version hiding under it. There is a compute ladder, and the application is expected to climb only as high as the task justifies.
That changes where the intelligence lives.
The model menu is an architecture
For a support reply, Luna may be enough. A document synthesis might start on Terra. A production incident spanning logs, code, and deployment history may justify Sol. OpenAI also introduced an ultra setting that coordinates multiple agents across parallel workstreams — effectively turning "use subagents" from an application pattern into a model option.
The frontier product is becoming a system, not a checkpoint.
This is good news for builders. Most workloads are mixtures: hundreds of cheap classification and extraction steps surrounding a few genuinely difficult decisions. Sending all of them to the largest model was never good engineering; it was what we did because routing was harder than overpaying.
Now routing is part of the job.
Difficulty is observable, eventually
The first routing policies will be crude:
- Start with the cheapest model that satisfies the required capabilities.
- Escalate on malformed output, low confidence, repeated tool errors, or a failed verifier.
- Reserve multi-agent execution for tasks that can actually be decomposed.
- Put a cost and time ceiling around the whole ladder.
The useful signals arrive during the trajectory. A request that sounds simple may reveal three conflicting data sources after the second tool call. A request that sounds ambitious may collapse into one database lookup. Routing once, before execution, leaves most of the information on the table.
I expect good agent runtimes to treat model choice like a database query planner: start with an estimate, observe the work, and change strategy when reality disagrees.
Benchmark the system you pay for
OpenAI says Sol delivers stronger performance per dollar and that ultra can accelerate complex work through parallel agents. Maybe. Vendor benchmarks tell me what is possible under their harness; they do not tell me whether my tools, prompts, latency budget, and error recovery produce the same curve.
The evaluation unit has to be the completed task:
- Did it succeed?
- How long did it take?
- How many model and tool calls did it need?
- What did success cost?
- How often does it succeed across repeated runs?
A smaller model that needs three retries is not cheaper. A frontier model that succeeds once but spends ten minutes orchestrating agents may not be better. Tokens are an ingredient; completed useful work is the product.
The release race will keep producing a new "smartest model." The durable advantage is building a runtime that knows when smart is worth paying for.