Hypertask · model exploration
The picker shows 7 options. They’re only 3 gateway models with a reasoning-effort switch. Here’s the true per-token price of each, why “GPT 5.5 Instant” is the most expensive tier you can pick, and how the new GPT-5.6 line (Luna / Terra / Sol) compares.
Instant and Thinking are the same model at the same rate. In aiModelOptions.ts each pair points at one gateway model and just flips reasoningEffort low↔high (Anthropic: thinking: disabled↔adaptive). Effort doesn’t change the rate card. “Instant” is cheaper only per turn, because low effort emits fewer tokens, never a cheaper price.
Your current default, GPT 5.5 Instant, is the priciest input rate on the board $5 / $30 — tied with Opus and Sol. The word “Instant” makes it feel light; the meter says otherwise.
GPT-5.6 has no “Instant” variant. It ships as three price tiers: Luna $1 / $6, Terra $2.50 / $15, Sol $5 / $30. Against GPT 5.5, Luna is 80% cheaper both ways with 1.05M context, Terra is 50% cheaper, and Sol is identical — Sol buys nothing on price.
The move: add GPT-5.6 Luna as a selectable option and steer the cheap-and-fast slot there instead of GPT 5.5 Instant. Keep the new-user default on a Claude model until Luna wins a real tool-calling eval — don’t swap the default on a release note.
| Picker option | Gateway model | Effort | Input | Output | Context | Status |
|---|
cheap · mid · premium · in picker live today · propose not wired, worth adding. Instant/Thinking rows share a model and a price — the only difference is reasoning effort.
| 5.6 tier | Input | vs 5.5 in | Output | vs 5.5 out | Context |
|---|
Baseline is gpt-5.5 at $5 / $30 — the model behind both GPT 5.5 Instant and GPT 5.5 Thinking in the picker.
One request = input × input rate + output × output rate. Thinking multiplies output tokens, not the rate. Prompt-cache reads (repeated system prompt + tools) would pull every bar down further.
| Plan | AI budget / mo | Turns on GPT 5.5 Instant | on Luna | on Sonnet 5 | on Opus Thinking |
|---|---|---|---|---|---|
| Pro$25 / mo | $5.00 | ~90 (3/day) | ~450 | ~250 | ~60 |
| Freetrial allowance | $0.50 | ~9 | ~45 | ~25 | — |
| BYOK$10 / mo · own key | Unlimited — requests bill to the member's own API key. Also the upsell when credits run out. | ||||
Budget ≈ 20% of plan price as AI COGS. Turn = 5k input + 1k output tokens, uncached; prompt-cache reads on the repeated system prompt + tools would stretch these 2–3×. Meter in dollars (gateway usage × price table), display as a credit bar — token counts aren't comparable across models. The spread is the argument for Luna: the same $5 buys 5× the turns.
At 5.1 there were instant / thinking SKUs, both $1.25/$10. At 5.6 that split is gone — you pick a price tier: Luna, Terra, Sol.
In Hypertask, “GPT 5.5 Instant/Thinking” are both gpt-5.5; only reasoningEffort changes (low vs high, per HTPR-3970 — “minimal” returned empty completions).
One rate per model. Sonnet 5 is $2/$10 and Opus 4.8 is $5/$25 whether thinking is disabled or adaptive.
Thinking just adds reasoning output tokens, so the turn costs more because it’s longer — same rate card.