LLM API Pricing Comparison — per-token prices over time, updated monthly
What a million tokens costs, tracked over time. Two data sources, both dated and verifiable: historical months reconstructed from Wayback Machine snapshots of Artificial Analysis model pages (per-model snapshot timestamps preserved in the underlying data), and the same live monthly collection that feeds the Pareto Frontier report from July 2026 on. Where no snapshot exists, the chart shows a gap — never an estimate; the archive’s coverage of AA starts February 2026. Superseded models (dashed) stay on the chart, because model churn is half the price story.
The numbers
Section titled “The numbers”Sorted by the metric that actually matters — dollars of output per Intelligence Index point (lower = more intelligence per dollar):
| Model | Output $/M | Input $/M | AA Index | $ per index point | Weights |
|---|---|---|---|---|---|
| DeepSeek V4 Flash | $0.28 | $0.14 | 40 | $0.007 | open |
| MiniMax M3 | $1.20 | $0.30 | 44 | $0.027 | open |
| GLM 5.2 | $4.40 | $1.40 | 51 | $0.086 | open |
| Kimi K2.6 | $4.00 | $0.95 | 43 | $0.093 | open |
| Kimi K2.7 | $4.00 | $0.95 | 42 | $0.095 | open |
| Gemini 3.5 Flash | $9.00 | $1.50 | 50 | $0.180 | closed |
| Sonnet 5 | $15.00 | $3.00 | 53 | $0.283 | closed |
| GPT-5.4 | $15.00 | $2.50 | 51 | $0.294 | closed |
| Opus 4.8 | $25.00 | $5.00 | 56 | $0.446 | closed |
| GPT-5.5 | $30.00 | $5.00 | 55 | $0.545 | closed |
| Claude Fable 5 | $50.00 | $10.00 | 60 | $0.833 | closed |
| MiMo V2.5 | $0.28 | $0.105 | — | — | open |
Analysis
Section titled “Analysis”- Every open model buys intelligence cheaper than every closed model. The most efficient closed option (Gemini 3.5 Flash, $0.18 per index point) still costs about twice the least efficient open one (Kimi K2.7, $0.095) — and 26× DeepSeek V4 Flash’s $0.007.
- The premium is at the top. From GLM 5.2 to Fable 5 the index rises 18% (51 → 60) while the price per point rises 10× ($0.086 → $0.833). You don’t pay for intelligence — you pay for the last few points of it.
- What to watch as editions accumulate: open-weight prices trend down because any host can serve the weights and competition compresses margins to serving cost; closed prices only move when the sole vendor decides. A year ago the best open model scored 22 on this index — today’s 51 came with no price increase at the open tier. If that dynamic holds, the lines above will keep diverging.
Break-even: when does flat-rate beat the meter?
Section titled “Break-even: when does flat-rate beat the meter?”The per-token prices above stop mattering past a usage threshold — here it is, computed from the table (blended at the 80% input / 20% output mix typical of real workloads):
| Flat subscription | vs paying per token for | Blended $/M | Break-even: tokens/month | In agent-hours¹ |
|---|---|---|---|---|
| Frontier, from $44.20/mo | GLM 5.2 itself (open, list price) | $2.00 | 22.1M | ~11 h/month |
| Frontier, from $44.20/mo | GPT-5.4 — same index, closed | $5.00 | 8.8M | ~4 h/month |
| Frontier, from $44.20/mo | GPT-5.5 | $10.00 | 4.4M | ~2 h/month |
| Core, from $5.94/mo | DeepSeek V4 Flash itself (open, list price) | $0.17 | 35.4M | ~17 h/month |
| Core, from $5.94/mo | GPT-5.4 mini — same index, closed | $1.50 | 4.0M | ~2 h/month |
¹ We measured a coding agent at ~2.1M tokens/hour — so a single working day of agent traffic per month clears every break-even on this table against closed per-token pricing, and 2–3 days clear it even against the open models’ own list prices. Below those volumes, the meter is genuinely cheaper; that’s the honest threshold.
The same thing as a picture — monthly cost against monthly usage; where a per-token diagonal crosses the flat line, the subscription starts winning:
One structural note: the flat side of this comparison is a reserved daily time block serving one request at a time per key — you’re buying capacity, not metered volume, so past break-even the marginal token costs zero for the rest of the month. The full framework for that decision is your AI bill should scale with users, not usage.
Caveats
Section titled “Caveats”Prices are vendor list prices for the reasoning variants Artificial Analysis evaluates; open-weight models are often cheaper on aggregators, so the open rows are conservative. ”$ per index point” divides output price by a composite index — it’s a comparison heuristic, not a claim that index points are linear in value. MiMo V2.5, which we serve, appears with its current OpenRouter price but no index or history — it has no Artificial Analysis entry yet, so those cells stay empty rather than estimated.
Changelog
Section titled “Changelog”- 2026-07-06 (backfill) — historical prices February–June 2026 reconstructed from Wayback Machine snapshots of Artificial Analysis: 19 models including superseded ones (Opus 4.6→4.7 handover, Kimi K2.5, GLM 5.1, MiniMax M2, Gemini 3 Pro/Flash). Notable finds in the record: DeepSeek V3.2’s output price rose from $0.42 to $1.60/M in May — open-weight prices mostly fall, but not always — and Kimi K2.5 wobbled $3.00 → $2.85 → $3.00.
- 2026-07-06 (first edition) — baseline prices ingested from the 2026-07 Pareto collection. Range on the board: $0.28/M (DeepSeek V4 Flash) to $50/M (Claude Fable 5) per million output tokens — a 178× spread.
On our pools these per-token prices stop applying at all: flat monthly, no token caps during your reserved hours — Frontier from $44.20/mo, Core from $5.94/mo. See the pools or read why your AI bill should scale with users, not usage.