The Mix Index · v0.1 · cost layer · updated July 25, 2026
Stop picking a model.
Price the pair.
Every benchmark site ranks single models. But that's not how serious work runs anymore: a strong model plans and reviews while a cheaper one executes — Anthropic's own published agent patterns work exactly this way. Nobody maintains the price sheet for those pairs. This is it.
v0.1 is the cost layer with every assumption in the open. The quality layer (measured evals per pair, not vendor claims) is v0.2 — we won't print numbers we haven't run.
| Planner × Executor | Feature build | Research run | vs all-Fable | When you'd pick it |
|---|---|---|---|---|
| Fable 5 solo usage credits (Pro) / in-plan on Max |
$2.00 | $2.23 | — | The baseline everyone imagines: flagship does everything. |
| Fable 5 × Sonnet 5 promo pricing through Aug 31 |
$0.92 | $0.87 | −58% | The classic split-stack: flagship judgment, cheap typing. Still the reference mix. |
| Fable 5 × Haiku 4.5 | $0.78 | $0.70 | −65% | Max-frugal Anthropic stack for high-volume, low-risk work. |
| Fable 5 × GPT-5.6 Luna | $0.80 | $0.71 | −64% | Cross-vendor floor: cheapest capable executor under a Fable plan. |
| Opus 5 solo launched Jul 24; in-plan on Pro+ |
$1.00 | $1.11 | −50% | The new Pro-plan default: one strong model, no credits. |
| Opus 5 × Sonnet 5 promo pricing through Aug 31 |
$0.59 | $0.60 | −72% | Post-Jul-24 sweet spot: near-flagship planning at half rate. |
| Opus 5 × Haiku 4.5 | $0.46 | $0.43 | −79% | Budget Anthropic stack now that Opus 5 plans well enough. |
| Opus 5 × Kimi K3 open weights due Jul 27 |
$0.73 | $0.78 | −64% | Open-weight executor with a 1M context for long-haul runs. |
| GPT-5.6 Sol × GPT-5.6 Terra | $0.73 | $0.74 | −65% | The all-OpenAI mirror of the split-stack. |
| Fable 5 × Opus 5 launched Jul 24; in-plan on Pro+ |
$1.32 | $1.38 | −36% | All-in Anthropic quality: Fable judges, Opus 5 executes. |
The assumptions, in the open
Feature build = plan (12k in / 3k out) + review (25k in / 2.5k out) on the planner; execution (60k in / 15k out) on the executor. Research run = frame (8k/2k) + adjudicate (20k/3k) on the planner; synthesis (120k in / 10k out) on the executor. Straight API list prices, no batch or cache discounts (both usually help the executor most). Change the token mix and the ranking barely moves; the spread between rows is what matters.
Why there's no quality column yet
Because we haven't measured it, and vendor numbers aren't measurements. The two anchors that exist are Anthropic's own (self-reported, on their benchmarks): an advisor pattern — Sonnet executes, Fable 5 advises — scoring ≈92% of flagship quality at 63% of the cost, and an orchestrator pattern at ≈96% at 46%. Directionally they say what this whole page says: intelligence is worth the most at decision points, and volume work doesn't need flagship tokens. v0.2 replaces anchors with our own task-level evals.
This table changes every time a vendor moves
Three price events in the last two weeks alone (the Jul-18 split, Opus 5, Kimi K3). The Field Guide re-runs this math weekly with worked workflows on top — $12/mo or $79/yr, read a full free issue first. Prefer one-time? The $29 Starter Pack ships 8 runbooks with the same honest math.
Sources: Anthropic pricing (Fable 5 $10/$50; Opus 5 $5/$25, launched Jul 24; Opus 4.8 $5/$25; Sonnet 5 $2/$10 promo through Aug 31; Haiku 4.5 $1/$5), OpenAI (Sol $5/$30, Terra $2.50/$15, Luna $1/$6), Moonshot (Kimi K3 $3/$15), xAI (Grok 4.5 $2/$6). List prices per million tokens as published July 25, 2026; check vendor pages before budgeting. Independent site — no vendor relationships.