The Cost of Intelligence
LLM API PRICING · 2023–2026 DATA · PROJECTED TO 2030
setTab("curves")}>Capability Curves
setTab("scatter")}>All Models
setTab("compression")}>Cost Compression
setTab("tiers")}>Tier Detail
setPt("blended")}>Blended (3:1)
setPt("input")}>Input
setPt("output")}>Output
{(tab==="curves"||tab==="scatter")&&
}
{(tab==="curves"||tab==="scatter")&&
}
TIER:
{TIERS.map(t=>tog(selT,setSelT,t.id)}>{t.short})}
TYPE:
tog(selK,setSelK,"closed")} color="#f59e0b">Closed
tog(selK,setSelK,"open")} color="#22c55e">Open
EVENT:
tog(selE,setSelE,"launch")} color="#10a37f">Launches
tog(selE,setSelE,"cut")} color="#f59e0b">Price Cuts
{(tab==="curves"||tab==="scatter")&&(
{tab==="curves"&&
Solid = best-fit · Dashed = projected to 2030 · ⭐ laptop / 📱 phone edge milestones · Hover tier labels to isolate
}
{tab==="curves"&&(
<>
{TIERS.filter(t=>selT.has(t.id)).map(t=>{
const cnt=(tierDat[t.id]||[]).length,tr=tierTrends[t.id];
const hl=tr&&tr.fit.b<0?(-Math.LN2/tr.fit.b*12).toFixed(0):null;
return(
setHov(t.id)} onMouseLeave={()=>setHov(null)} style={{padding:"6px 14px",borderRadius:8,cursor:"pointer",border:`2px solid ${TC[t.id-1]}`,background:hov===t.id?`${TC[t.id-1]}22`:"transparent",display:"flex",alignItems:"center",gap:8}}>
{t.short}
{t.desc}
({cnt})
{hl&&· t½≈{hl}mo}
);
})}
{visP.map(p=>{p})}
|
● Closed ◯ Open ◆ Cut ── Fit ╌╌ Projected ⭐ On laptop 📱 On phone
)}
{tab==="scatter"&&
{TIERS.filter(t=>selT.has(t.id)).map(t=>● {t.short})}
| ◆ Cut · Ring = Open
}
)}
{tab==="compression"&&(
First price vs cheapest today at each capability level
{TIERS.map((t,i)=>{
const s=compStats[t.id];if(!s)return null;
const tr=tierTrends[t.id],hl=tr&&tr.fit.b<0?(-Math.LN2/tr.fit.b*12).toFixed(0):null;
return(
{s.ratio.toFixed(0)}x
cheaper
First:{s.fn_} ${s.fp.toFixed(2)}
{s.ccp!=null&&
Best closed:{s.ccn} ${s.ccp.toFixed(2)}
}
{s.cop!=null&&
Best open:{s.con} ${s.cop.toFixed(2)}
}
{hl&&
Price half-life: ~{hl} months
}
);
})}
)}
{tab==="tiers"&&(
{TIERS.filter(t=>selT.has(t.id)).map(t=>{
const tm=models.filter(m=>m.tier===t.id&&selK.has(m.type)&&selE.has(m.event)).map(m=>({...m,price:gp(m)})).sort((a,b)=>a.date-b.date);
return(
{t.name}
{t.desc} · {t.bm}
{tm.map((m,i)=>(
{m.name}
{m.event==="cut"?"CUT":"NEW"}
{m.type}
{m.provider} · {dL(m.date)}
${m.price.toFixed(2)} · ${m.input.toFixed(2)}/${m.output.toFixed(2)}
))}
);
})}
)}
Key Insights
{[
{icon:"📉",t:"Every tier follows exponential decay",d:"Half-lives 6–12 months. Capable tier prices halve roughly every 6–8 months."},
{icon:"🔄",t:"Labs keep shipping at every tier",d:"GPT-5 Nano, Gemma 3N, Grok 4.1 Fast — fresh entries across all tiers through 2025-2026. xAI especially aggressive on price."},
{icon:"🆕",t:"New entrants smash the price floor",d:"Grok 4.1 Fast at $0.20/$0.50 delivers Pro-tier reasoning. GPT-OSS-120B at $0.04/$0.19 — Advanced-tier at Basic prices."},
{icon:"✂️",t:"Two compression mechanisms",d:"New models entering a tier AND price cuts on existing models. Both push the floor down."},
{icon:"💡",t:"Architecture > parameters",d:"Qwen3.5-35B (3B active) beats prior 235B flagship. Hybrid attention + MoE + RL moves the frontier without scaling params."},
{icon:"📱",t:"Cloud → Edge lag is shrinking fast",d:"Basic tier took 4mo to reach laptops, 9mo to phones. Qwen3.5-35B (3B active) now runs Advanced-tier on a 21GB Mac — frontier on a regular laptop."},
].map((item,i)=>(
{item.icon}
{item.t}
{item.d}
))}
METHODOLOGY
Tiers: GPQA Diamond, SWE-bench, MMLU composites. Prices: Launch = provider announcement. Cut = pricing page change. Trendlines: Least-squares exponential fit on log-price vs time. Half-life = ln(2)/|b|. Blended: (3×input + output)/4. Edge: ⭐ = runs on MacBook Pro (Apple Silicon), 📱 = runs on iPhone/Android. Prices via OpenRouter. X-axis icons mark first time each tier reached that device.
Sources: Onyx · Artificial Analysis · Arena · Vals.ai · Provider docs · Feb 2026
);
}