OpenAI moved GPT-5.6 to general availability on July 9, 2026, and did something quietly telling: instead of one flagship, it shipped three. Luna, Terra, and Sol — smallest to largest — at $1/$6, $2.50/$15, and $5/$30 per million input/output tokens. On paper Sol undercuts Claude Fable 5 ($10/$50) and matches Claude Opus ($5/$25). But as Simon Willison points out, the sticker price barely matters anymore.

The token is the wrong unit

Here’s the number that should reframe the whole comparison: Sol reportedly reaches 62.6% on OSWorld 2.0 while spending 85% fewer output tokens than Opus 4.8. Output tokens are the expensive half of every bill. A model that costs the same per token but emits a fraction of them is, in practice, dramatically cheaper — and faster. Price-per-million has become a decoy stat. The metric that matters is cost-per-completed-task, and that depends on how tersely a model reasons its way to an answer.

This is why the three-tier split is smart product design rather than menu padding. Luna isn’t a crippled flagship; it’s a deliberate floor for high-volume, low-stakes calls where you’d never want Sol’s deliberation. OpenAI is betting developers will route work across tiers the way you’d pick EC2 instance sizes — and that routing, not raw capability, is where the money leaks.

Programmatic tool calling is the quieter headline

MarkTechPost flags the change that will actually alter how people build: Programmatic Tool Calling in the Responses API. Sol also tops the Artificial Analysis Coding Agent Index at 80, 2.8 points above Fable 5. Both signals point the same direction — this release is aimed squarely at agents, not chatbots.

Programmatic tool calling lets the model orchestrate tools as code rather than emitting one JSON call, waiting, and repeating. For agentic loops, that collapses round-trips, and round-trips are where latency and token waste compound. Pair that with an 85%-leaner output profile and you can see the thesis: OpenAI wants to win the agent workload specifically, where efficiency multiplies across dozens of chained steps.

What to actually watch

Benchmark leads of a few points rarely survive contact with real workloads, and a single vendor’s token-efficiency claim deserves independent replication before anyone reprices their infrastructure around it. The Coding Agent Index gap over Fable 5 is real but narrow.

The strategic read is clearer than the leaderboard. The frontier fight has moved from “who has the biggest model” to “who finishes the task with the fewest tokens and the fewest round-trips.” GPT-5.6’s three tiers and programmatic tool calling are both answers to that question. If the efficiency numbers hold up outside OpenAI’s own benchmarks, the interesting competition in 2026 won’t be about capability ceilings — it’ll be about how cheaply you can hit them.