GPT-5.6 Luna is now the default model — and we caught a pricing bug before it shipped
As of July 31, 2026, new Agent Planners organizations default their orchestrator, analytics, optimization and creative agents to GPT-5.6 Luna instead of Claude Sonnet 4.6 — a rare change that improves quality and cost at once: $0.20/$1.20 per 1M tokens versus $3.00/$15.00, on a 1.05M-token window versus 1M. While shipping the change, we found that our metering fallback would have billed Luna at 30x its real price if the live gateway were ever briefly unreachable, and fixed it in the same change — before it could charge anyone incorrectly.
What changed
New organizations' “Quality-first” preset — the default every org starts on — now runs GPT-5.6 Luna on the four agents where model quality matters most: the orchestrator (planning), analytics, optimization and creative. Safety stays on a fast, cheap model (structured yes/no checks don't need a frontier model), and research runs a long-context model suited to synthesis.
This only affects organizations created from now on. If you already picked models on the Agents page, or your org was created before this change, nothing about your setup moves — the new default is seeded only at org creation, never migrated onto an existing one.
The pricing case for switching the default
| Claude Sonnet 4.6 (previous default) | GPT-5.6 Luna (new default) | |
|---|---|---|
| Price per 1M tokens | $3.00 input / $15.00 output | $0.20 input / $1.20 output |
| Context window | 1,000,000 tokens | 1,050,000 tokens |
| Credit rate / 1K tokens | 12 in · 60 out | 1 in · 6 out |
The bug we found while shipping this
Metering has two layers: the live gateway price, checked first on every call, and a static fallback table used only if that call is briefly unreachable. The fallback matches models by pattern — and Luna's id contains “gpt-5”, which the fallback's generic frontier-model rule also matches, at 30 credits per 1K input tokens and 120 per 1K output.
Luna's real rate is 1 and 6. If the fallback had ever fired for an org defaulted onto Luna, it would have billed a $0.20-per-million-token model as if it cost $6.00 per million — a 30× overcharge, on the path every new organization now takes by default.
We caught it in the same change that introduced the new default, added a specific fallback rule for Luna (and its sibling Terra) ahead of the generic pattern, and shipped both together. No organization was ever billed the wrong rate — the live gateway price was correct throughout, and the fallback is now correct too.
It wasn't just Luna
Finding that gap prompted a full audit of the fallback table against real gateway prices. The result: 29 of the 43 selectable models had drifted from their real price by more than 1.5× on the output rate — in both directions. GLM-4.7 Flash was being over-billed roughly 10×; Gemini 3.5 Flash — the safety agent's model in every preset, including Cost-first — was under-billed roughly 11×.
The fix wasn't re-tuning 29 numbers by hand, since a hand-tuned table is exactly what drifts. It's a daily cron that re-derives every model's real price from the live catalog into a persisted snapshot, which both the billing fallback and the public model pricing page now read — a number recomputed from the source every day can't silently go stale the way a hand-typed one did.
The same audit habit caught one more gap while curating Anthropic's newest models the following day — see Claude Fable 5 and Opus 5 are now in Agent Planners.
What this means if you're starting out
- New organizations get a strong planner and executor on GPT-5.6 Luna by default, at a fraction of the previous default's per-token cost — first-run quality without a first-run credit hit.
- You can change any agent's model any time on the Agents page — the default is a starting point, not a constraint.
- See exactly what every model costs, always current, on the model pricing & credits ranking.
Frequently asked questions
- What is the default AI model in Agent Planners?
- As of July 31, 2026, new organizations default their orchestrator, analytics, optimization and creative agents to GPT-5.6 Luna ($0.20/$1.20 per 1M tokens, 1.05M-token context). Safety runs a cheap fast model; research runs a model suited to long-context synthesis. Existing organizations are unaffected — the default only seeds at org creation.
- Was anyone overcharged for GPT-5.6 Luna?
- No. The bug was in the metering FALLBACK, used only if the live gateway price is briefly unreachable — and it was found and fixed in the same change that introduced Luna as the default, before it could bill anyone at the wrong rate. The live gateway price, which meters usage the vast majority of the time, was correct throughout.
- Why is GPT-5.6 Luna so much cheaper than the model it replaced?
- Different providers price frontier and near-frontier models very differently. Luna prices at $0.20/$1.20 per 1M tokens versus Claude Sonnet 4.6's $3.00/$15.00 — about 13x cheaper per output token — while running on a slightly larger 1.05M-token context window and passing the same functional checks (structured output, tool calls, CJK).
- How do I see current AI model pricing on Agent Planners?
- The model pricing & credits ranking page lists every selectable model's credit rate, ranked cheapest to priciest, generated from the same daily-refreshed real gateway prices that back billing — never a stale hand-typed number.
- Can I keep using a different model as my default?
- Yes. The new default only applies to organizations created from now on, and even then it's just a starting point — change any agent's model any time on the Agents page.