Snapshot 19 · Friday, 21 August 2026
What did your LLM cost a week ago? A month ago? Which provider is dropping prices fastest right now?
A daily snapshot of every priced LLM on arena.ai's leaderboard, committed to git so the history is auditable. 267 models priced today across 27 providers. Sibling pages: llm-rate for the use-case picker, bench for real workload runs.
Day one notice
Today is snapshot one. Price-drop diffs need at least two days of history, so the "biggest movers" section activates tomorrow. Until then, the snapshot below shows every priced model on the leaderboard right now. Subscribe to the GitHub repo if you want the JSON files as they land: data/pricing/.
First snapshot captured at 2026-08-21T13:56:45.730Z.
Sorted by blended cost (30% input / 70% output per million tokens), within each Arena rating band. Most production traffic is output-heavy, which is why blended cost matters more than headline input price.
Workhorse
rating 1300-1399 · the bulk-volume tier
| Model | Elo | In $/M | Out $/M | Blend | |
|---|---|---|---|---|---|
| 01 | sora-2 OpenAI |
1340 | $0.00 | $0.10 | $0.07 |
| 02 | solar-pro4 Upstage |
1376 | $0.03 | $0.12 | $0.09 |
| 03 | veo-3-fast-audio Google |
1347 | $0.00 | $0.15 | $0.10 |
| 04 | veo-3.1-fast-audio Google |
1361 | $0.00 | $0.15 | $0.10 |
| 05 | gpt-oss-120b OpenAI |
1352 | $0.03 | $0.17 | $0.13 |
| 06 | qwen3-30b-a3b-instruct-2507 Alibaba |
1383 | $0.05 | $0.19 | $0.15 |
| 07 | laguna-xs.2 Poolside |
1302 | $0.10 | $0.20 | $0.17 |
| 08 | sora-2-pro OpenAI |
1364 | $0.00 | $0.30 | $0.21 |
Mid-tier
rating 1400-1499 · the daily driver tier
| Model | Elo | In $/M | Out $/M | Blend | |
|---|---|---|---|---|---|
| 01 | mimo-v2.5 Xiaomi |
1434 | $0.14 | $0.28 | $0.24 |
| 02 | gemma-4-31b Google |
1451 | $0.14 | $0.40 | $0.32 |
| 03 | deepseek-v3.2 DeepSeek |
1425 | $0.27 | $0.40 | $0.36 |
| 04 | deepseek-v3.2-thinking DeepSeek |
1423 | $0.27 | $0.40 | $0.36 |
| 05 | deepseek-v3.2-exp DeepSeek |
1422 | $0.27 | $0.41 | $0.37 |
| 06 | deepseek-v3.2-exp-thinking DeepSeek |
1425 | $0.27 | $0.41 | $0.37 |
| 07 | hy3 Tencent |
1458 | $0.13 | $0.53 | $0.41 |
| 08 | grok-4-1-fast-reasoning SpaceXAI |
1431 | $0.20 | $0.50 | $0.41 |
Frontier
rating 1500+ · the "I need this right" tier
| Model | Elo | In $/M | Out $/M | Blend | |
|---|---|---|---|---|---|
| 01 | gpt-5.6-luna-xhigh (codex-harness) OpenAI |
1517 | $0.20 | $1.20 | $0.90 |
| 02 | deepseek-v4-flash-high DeepSeek |
1578 | $0.44 | $1.32 | $1.06 |
| 03 | qwen3.7-max-20260517 Alibaba |
1517 | $1.48 | $4.42 | $3.54 |
| 04 | gpt-5.6-terra-xhigh (codex-harness) OpenAI |
1520 | $2.00 | $12.00 | $9.00 |
| 05 | claude-opus-4-6-high Anthropic |
1505 | $5.00 | $25.00 | $19.00 |
| 06 | claude-opus-4-7-high Anthropic |
1502 | $5.00 | $25.00 | $19.00 |
| 07 | gpt-5.5-xhigh (codex-harness) OpenAI |
1508 | $5.00 | $30.00 | $22.50 |
| 08 | gpt-5.6-sol-xhigh (codex-harness) OpenAI |
1619 | $5.00 | $30.00 | $22.50 |
The snapshot job runs once a day, hits arena.ai's leaderboard (which embeds input and output pricing per million tokens directly in the page payload), and commits a dated JSON file to data/pricing/. This page reads the latest two snapshots and computes the deltas. Blended cost is a 30/70 input/output mix, which approximates a typical chat or agent workload where outputs are 2-3x the inputs.
Caveats: arena.ai's prices reflect what their tracking system has captured, not necessarily every provider's most recent rate card. When you see a drop here, confirm against the provider's own pricing page before re-pricing your stack. Open-weights models priced at $0 are filtered out (they're not API products in the same sense).