LLM price barometer
Changes of 1% or more in the price we quote, from 14 Nov to 21 Nov, biggest first. Percentages are on the blended price (30% input, 70% output). Data snapshot: 2026-11-21.
7 price changes (4 up, 3 down) among 306 models compared. 24 other moves are flagged as a change of quoted host or an outlier listing.
Price changes (7)
Price increases (4)
- +33.3% Ling 3.1 Flash (inclusionai): blended price $0.108 to $0.144 per 1M tokens.
- +10.0% Mistral Medium 3.1 (Mistral): blended price $1.52 to $1.672 per 1M tokens.
- +10.0% Mistral Medium 3 (Mistral): blended price $1.52 to $1.672 per 1M tokens.
- +1.6% Particle-1 (particle-ai): blended price $0.365 to $0.371 per 1M tokens; input +40.0%, output 0.0%.
Price decreases (3)
- -40.0% GLM 5.3 Prime (Zhipu): blended price $7.00 to $4.20 per 1M tokens.
- -1.8% DeepSeek V4.1 Flash (DeepSeek): blended price $0.1572 to $0.1544 per 1M tokens.
- -1.5% Llama 3.2 3B Instruct (Meta): blended price $0.2498 to $0.246 per 1M tokens; input -1.8%, output -1.5%.
Quoted host changes and outliers (24)
Many models are sold by several hosts at different rates, and we quote one of them. These moves are flagged as a change of quoted host: the new rate matches a different host from the one we quoted at the start of the window. Rows marked Outlier listing involve a quoted rate far from the model's typical host rate.
Price increases (7)
- +121.2% MiniMax M2.7 (MiniMax): blended price $0.651 to $1.44 per 1M tokens; input +185.7%, output +114.3%.
- +89.9% MiMo-V2.5-Pro (Xiaomi): blended price $0.7395 to $1.404 per 1M tokens; input +10.3%, output +106.9%.
- +28.7% DeepSeek V4.1 Flash 1017 (DeepSeek): blended price $0.3614 to $0.465 per 1M tokens; input +294.7%, output +20.0%.
- +26.0% Gemma 3 27B (Google): blended price $0.269 to $0.339 per 1M tokens; input 0.0%, output +28.6%.
- +24.2% Muse Glimmer 30B (meta): blended price $0.93 to $1.155 per 1M tokens; input +16.7%, output +25.0%.
- +7.0% R1 0528 (DeepSeek): blended price $1.655 to $1.7715 per 1M tokens; input +14.2%, output +6.3%.
- +2.4% Qwen3.8 27B (Alibaba): blended price $1.3575 to $1.39 per 1M tokens; input -60.0%, output +4.5%. Mixed direction. Input and output moved in opposite directions, so the effect on your bill depends on your token mix.
Price decreases (17)
- -60.7% GLM 5.3 (Zhipu): blended price $2.5452 to $1.001 per 1M tokens; input -16.8%, output -61.1%. Outlier listing. At one end of this window, the price we quoted was an outlier listing: input almost free and output at or above the usual rate, far from what most hosts charge. The percentage reflects that listing.
- -60.5% GLM 5.2 (Zhipu): blended price $4.7621 to $1.88 per 1M tokens; input +8757.1%, output -64.4%. Outlier listing. At one end of this window, the price we quoted was an outlier listing: input almost free and output at or above the usual rate, far from what most hosts charge. The percentage reflects that listing. Mixed direction. Input and output moved in opposite directions, so the effect on your bill depends on your token mix.
- -53.3% Qwen3 Coder 30B A3B Instruct (Alibaba): blended price $0.465 to $0.217 per 1M tokens.
- -51.0% DeepSeek V4 Flash Vision Exp (DeepSeek): blended price $1.056 to $0.5174 per 1M tokens.
- -40.0% gpt-oss-120b (OpenAI): blended price $0.465 to $0.279 per 1M tokens.
- -39.4% Qwen3.5-27B (Alibaba): blended price $1.898 to $1.1505 per 1M tokens; input -25.0%, output -40.0%.
- -36.3% Granite 4.2 8B (ibm-granite): blended price $0.193 to $0.123 per 1M tokens; input 0.0%, output -40.0%.
- -35.4% gpt-oss-20b (OpenAI): blended price $0.10 to $0.0646 per 1M tokens; input -43.3%, output -34.6%.
- -32.3% Mistral Small 3.2 24B (Mistral): blended price $0.24 to $0.1625 per 1M tokens; input -25.0%, output -33.3%.
- -23.6% GLM 4.5 Air (Zhipu): blended price $0.83 to $0.634 per 1M tokens; input -35.0%, output -22.7%.
- -21.8% GLM 4.7 (Zhipu): blended price $1.72 to $1.345 per 1M tokens; input -33.3%, output -20.5%.
- -17.9% Kimi K3.1 Pro (Moonshot): blended price $7.60 to $6.2359 per 1M tokens; input -87.4%, output -12.0%. Outlier listing. At one end of this window, the price we quoted was an outlier listing: input almost free and output at or above the usual rate, far from what most hosts charge. The percentage reflects that listing.
- -17.2% Qwen3.5 397B A17B (Alibaba): blended price $2.70 to $2.235 per 1M tokens; input -25.0%, output -16.7%.
- -13.3% Kimi K2.7 Code (Moonshot): blended price $2.63 to $2.28 per 1M tokens; input 0.0%, output -14.3%.
- -12.4% DeepSeek V3 0324 (DeepSeek): blended price $0.885 to $0.775 per 1M tokens; input -13.8%, output -12.3%.
- -9.1% Nemotron 3 Ultra (Nvidia): blended price $1.86 to $1.69 per 1M tokens; input -16.7%, output -8.3%.
- -1.9% Beam 1.0 501B A23B (reflection): blended price $0.79 to $0.775 per 1M tokens; input -16.7%, output 0.0%.
Since tracking began
Average quoted price
Median price per 1M tokens
Models added and removed
Every price change, day by day
The price we quote changed at least once for 88 of 386 models between 30 Aug and 21 Nov: 676 changes, most-changed model first. 390 are established as a change of quoted host (outlined).
How these figures are calculated
- Prices are blended. 30% input and 70% output, the same blended price the Efficiency score uses. Your own workload has its own token mix, which is why the pricing table's cost-per-request column is weighted by the use case you pick.
- List prices only. Batch and cache discounts depend on you implementing them, so they stay out of the comparison.
- Corrections are left out. When a quoted rate was wrong at the source and later fixed, the difference is a data correction, not a price move. The model stays off this page while the wrong figure sits inside the window.
- A change of quoted host is flagged. Most models are sold by several hosts, often at rates several times apart, and we quote one of them. A move is flagged when the new rate matches a different host from the one we quoted at the start of the window, and no host itself moved from the old rate to the new one. The row is listed apart and kept off the home page summary.
- An outlier listing is flagged too. A quoted rate with input at most 35% of the median host's, and an output-to-input ratio at least 4 times the median, is far from the model's typical host rate. The percentage then reflects that listing, so the row is listed apart as well.
- A missing baseline gives no entry. New listings, renames, delistings and free-to-paid transitions get no percentage.
- The average quoted price is a trimmed index. It follows the models listed since tracking began. Each model's quoted price is taken relative to the first day. The 5% of models that rose most and the 5% that fell most are set aside, and the index is the geometric mean of the rest, so every model counts equally and a single outlier listing cannot move it.
- Added and removed are first and last appearances. A model gone for a few days and back counts as neither. A rename counts as one of each, because the archive identifies a model by its name.
The full cost-per-request formulas are written out in the cost per request methodology guide.
What to do with a price move
- Re-test your model mix on your own use case. A blended percentage is not your percentage. A model that dropped 20% on the blend can be flat for a workload that is mostly input tokens.
- Re-check batch and cache eligibility after a price change. Hosts reprice tiers independently, and a discount that was not worth implementing at the old rate sometimes is at the new one.
- Revisit routing thresholds rather than the default model. Where a cheaper model crosses the quality bar for part of the traffic, the saving comes from the split, not from switching everything at once.
LLM pricing · Models by maker · Cloud compute pricing · Compare models · Price barometer · FinOps guides · API and MCP server