Most cost forecasts assume that LLM API prices only fall. The archive says otherwise. We read Internet Archive captures of 11 providers’ own pricing pages, one per month from 2023 to November 2026, and compared each model’s price capture by capture. We found 52 list-price rises on the same model. This post lists all of them with their archive links, then turns them into a budgeting rule. BroadRouter is BroadNet’s website; BroadNet is not in this data because the archive has no captures of its pricing page.
The dataset behind it is our LLM API price history. For the cuts, see LLM API price cuts 2023 to 2026.
How often do LLM API prices go up?
Less often than they fall, but often enough to plan for. Across 161 list-price changes on the same model:
- 100 were cuts, 52 rises and 9 mixed (one price up, one down).
- The median rise was 40% on a blended price (3 input tokens per output token): 45% on input and 40% on output.
- 16 of the 52 rises at least doubled the blended price. The largest, 400%, were DeepInfra’s MythoMax-L2-13b and Llama-3.2-11B-Vision-Instruct.
- 39 rises raised both input and output, 11 raised input only and 2 raised output only.
Rises are becoming more common. By the year of the later capture: 2 rises in 2024, 19 in 2025 and 31 in 2026 to early November, when cuts numbered 28.
Who raised prices?
| Provider | Rises | Kind of provider |
|---|---|---|
| DeepInfra | 40 | Host |
| Together AI | 7 | Host |
| Mistral AI | 2 | Maker |
| Anthropic | 1 | Maker |
| DeepSeek | 1 | Maker |
| Fireworks AI | 1 | Host |
OpenAI, Google, xAI, Groq and Cohere had no list-price rise on an existing model in our captures. 4 of the 52 rises came from model makers; the other 48 from hosts serving open-weight models.
Rises did not wait for a model to age. Measured from the first capture of each model, the median rise came about 4 months in (134 days), and only 5 of the 52 came a year or more later. 12 models were both cut and raised at some point, for example Together AI’s Qwen3.8 Flash, cut 40% between the September and October 2026 captures and raised 67% back to its earlier price by November.
Which rises should you know about?
- Claude 3.5 Haiku, Anthropic. The 2 November 2024 capture has a Claude 3.5 Haiku card at USD 0.25 input and USD 1.25 output per 1M tokens. The 3 December 2024 capture shows USD 1 and USD 5, a 300% rise. The 1 January 2025 capture shows USD 0.80 and USD 4. Both prices come from the pricing cards embedded in the page. We report what the captures show and do not claim when Anthropic decided either price.
- DeepSeek V4 Pro, DeepSeek. The 1 August 2026 capture lists USD 0.435 input and USD 0.87 output, with a footnote that “The DeepSeek API service will soon adopt a peak/off-peak pricing policy” at twice the regular price in peak hours. The 3 October 2026 capture lists USD 1.32 and USD 3.96 at peak, a 264% rise blended. Even the off-peak input price, USD 0.66, is 52% above the August price.
- Codestral, Mistral. codestral-latest rose 50%, from USD 0.2 / USD 0.6 to USD 0.3 / USD 0.9, between the January and February 2025 captures, a few months after an 80% cut on the earlier Codestral.
- DeepInfra in batches. Ten DeepInfra models rose between the 11 August 2025 and 26 September 2025 captures, among them Llama-3.3-70B-Instruct-Turbo (+233%) and Mixtral-8x7B-Instruct-v0.1 (+233%). DeepInfra’s page also listed claude-sonnet-5 at USD 2.00 / USD 10.00 in August 2026 and USD 3.00 / USD 15.00 in September 2026.
- Together AI in 2026. All 7 of Together’s rises came in 2026, including Gemma 4 31B (+94%) and Llama 3.3 70B (+18%).
The full list of 52 rises
Largest first. Prices are USD per 1M tokens, input / output. A gap note means the two captures are more than 60 days apart.
| Provider and model | Captures | Before, in / out | After, in / out | Blended change | Archive |
|---|---|---|---|---|---|
| DeepInfra MythoMax-L2-13b | 3 Feb 2026 → 4 Mar 2026 | 0.08 / 0.08 | 0.40 / 0.40 | +400% | before, after |
| DeepInfra Llama-3.2-11B-Vision-Instruct | 7 Apr 2026 → 2 May 2026 | 0.049 / 0.049 | 0.245 / 0.245 | +400% | before, after |
| Anthropic Claude 3.5 Haiku | 2 Nov 2024 → 3 Dec 2024 | 0.25 / 1.25 | 1.00 / 5.00 | +300% | before, after |
| DeepInfra Llama-3.2-3B-Instruct | 9 Jul 2025 → 11 Aug 2025 | 0.003 / 0.006 | 0.012 / 0.024 | +300% | before, after |
| DeepSeek deepseek-v4-pro | 1 Aug 2026 → 3 Oct 2026 (63 days apart) | 0.435 / 0.87 | 1.32 / 3.96 | +264% | before, after |
| DeepInfra Llama-3.3-70B-Instruct-Turbo | 11 Aug 2025 → 26 Sep 2025 | 0.038 / 0.12 | 0.13 / 0.39 | +233% | before, after |
| DeepInfra Mixtral-8x7B-Instruct-v0.1 | 11 Aug 2025 → 26 Sep 2025 | 0.08 / 0.24 | 0.40 / 0.40 | +233% | before, after |
| Together AI Gemma 3n E4B Instruct | 7 Apr 2026 → 1 May 2026 | 0.02 / 0.04 | 0.06 / 0.12 | +200% | before, after |
| DeepInfra DeepSeek-R1-Distill-Qwen-32B | 11 Aug 2025 → 26 Sep 2025 | 0.075 / 0.15 | 0.27 / 0.27 | +188% | before, after |
| DeepInfra Meta-Llama-3.1-70B-Instruct-Turbo | 26 Sep 2025 → 2 Oct 2025 | 0.10 / 0.28 | 0.40 / 0.40 | +176% | before, after |
| DeepInfra DeepSeek-R1-Distill-Llama-70B | 2 Oct 2025 → 12 Nov 2025 | 0.20 / 0.60 | 0.60 / 1.20 | +150% | before, after |
| DeepInfra Llama-3.3-Nemotron-Super-49B-v1.5 | 1 Jun 2026 → 3 Jul 2026 | 0.10 / 0.40 | 0.40 / 0.40 | +129% | before, after |
| DeepInfra QwQ-32B | 11 Aug 2025 → 26 Sep 2025 | 0.075 / 0.15 | 0.15 / 0.40 | +127% | before, after |
| DeepInfra Qwen3-235B-A22B-Instruct-2507 | 3 Jul 2026 → 2 Aug 2026 | 0.09 / 0.10 | 0.09 / 0.55 | +122% | before, after |
| DeepInfra gemma-3-4b-it | 11 Aug 2025 → 26 Sep 2025 | 0.02 / 0.04 | 0.04 / 0.08 | +100% | before, after |
| DeepInfra Llama-3.1-Nemotron-70B-Instruct | 12 Nov 2025 → 3 Dec 2025 | 0.60 / 0.60 | 1.20 / 1.20 | +100% | before, after |
| DeepInfra Qwen2.5-72B-Instruct | 7 Apr 2026 → 2 May 2026 | 0.12 / 0.39 | 0.36 / 0.40 | +97% | before, after |
| Together AI Gemma 4 31B | 1 May 2026 → 1 Jun 2026 | 0.20 / 0.50 | 0.39 / 0.97 | +94% | before, after |
| DeepInfra DeepSeek-R1-Distill-Llama-70B | 11 Aug 2025 → 26 Sep 2025 | 0.10 / 0.40 | 0.20 / 0.60 | +71% | before, after |
| Together AI Qwen3.5 9B | 1 Jun 2026 → 1 Jul 2026 | 0.10 / 0.15 | 0.17 / 0.25 | +69% | before, after |
| Together AI Qwen3.8 Flash | 3 Oct 2026 → 2 Nov 2026 | 0.09 / 0.28 | 0.15 / 0.47 | +67% | before, after |
| Mistral AI Codestral | 2 Jan 2025 → 3 Feb 2025 | 0.20 / 0.60 | 0.30 / 0.90 | +50% | before, after |
| DeepInfra claude-sonnet-5 | 2 Aug 2026 → 1 Sep 2026 | 2.00 / 10.00 | 3.00 / 15.00 | +50% | before, after |
| DeepInfra Meta-Llama-3.1-70B-Instruct | 11 Aug 2025 → 26 Sep 2025 | 0.23 / 0.40 | 0.40 / 0.40 | +47% | before, after |
| Fireworks AI OpenAI gpt OSS 20b | 9 Aug 2025 → 3 Sep 2025 | 0.05 / 0.20 | 0.07 / 0.30 | +46% | before, after |
| DeepInfra Llama-3.2-11B-Vision-Instruct | 1 Jun 2026 → 3 Jul 2026 | 0.245 / 0.245 | 0.345 / 0.345 | +41% | before, after |
| Together AI Llama 3 8B Instruct Lite | 1 Jun 2026 → 1 Jul 2026 | 0.10 / 0.10 | 0.14 / 0.14 | +40% | before, after |
| DeepInfra Meta-Llama-3.1-8B-Instruct-Turbo | 11 Aug 2025 → 26 Sep 2025 | 0.015 / 0.02 | 0.02 / 0.03 | +38% | before, after |
| DeepInfra Qwen3-30B-A3B | 2 May 2026 → 1 Jun 2026 | 0.08 / 0.28 | 0.09 / 0.45 | +38% | before, after |
| DeepInfra MythoMax-L2-13b | 11 Aug 2025 → 26 Sep 2025 | 0.06 / 0.06 | 0.08 / 0.09 | +38% | before, after |
| DeepInfra Mixtral-8x7B-Instruct-v0.1 | 2 Oct 2025 → 12 Nov 2025 | 0.40 / 0.40 | 0.54 / 0.54 | +35% | before, after |
| DeepInfra Mistral-Nemo-Instruct-2407 | 1 Oct 2026 → 2 Nov 2026 | 0.019 / 0.03 | 0.029 / 0.03 | +34% | before, after |
| DeepInfra Llama-4-Maverick-17B-128E | 3 Jul 2026 → 2 Aug 2026 | 0.15 / 0.60 | 0.20 / 0.80 | +33% | before, after |
| DeepInfra DeepSeek-R1 | 11 Aug 2025 → 26 Sep 2025 | 0.45 / 2.15 | 0.70 / 2.40 | +29% | before, after |
| DeepInfra Qwen3-14B | 3 Feb 2026 → 4 Mar 2026 | 0.08 / 0.24 | 0.12 / 0.24 | +25% | before, after |
| DeepInfra gemma-3-4b-it | 1 Jun 2026 → 3 Jul 2026 | 0.04 / 0.08 | 0.05 / 0.10 | +25% | before, after |
| DeepInfra DeepSeek-V3.1-Terminus | 2 May 2026 → 1 Jun 2026 | 0.21 / 0.79 | 0.27 / 0.95 | +24% | before, after |
| DeepInfra gemma-3-12b-it | 1 Jun 2026 → 3 Jul 2026 | 0.04 / 0.13 | 0.05 / 0.15 | +20% | before, after |
| Together AI Qwen3.7-Max | 2 Sep 2026 → 3 Oct 2026 | 1.25 / 3.75 | 1.50 / 4.50 | +20% | before, after |
| DeepInfra DeepSeek-V3.1 | 3 Jul 2026 → 2 Aug 2026 | 0.21 / 0.79 | 0.25 / 0.95 | +20% | before, after |
| DeepInfra Qwen3-30B-A3B | 1 Jun 2026 → 3 Jul 2026 | 0.09 / 0.45 | 0.12 / 0.50 | +19% | before, after |
| DeepInfra Llama-3-70B-Instruct | 4 Apr 2025 → 3 May 2025 | 0.23 / 0.40 | 0.30 / 0.40 | +19% | before, after |
| Together AI Llama 3.3 70B | 1 Jun 2026 → 1 Jul 2026 | 0.88 / 0.88 | 1.04 / 1.04 | +18% | before, after |
| DeepInfra DeepSeek-V3-0324 | 1 Jun 2026 → 3 Jul 2026 | 0.20 / 0.77 | 0.24 / 0.90 | +18% | before, after |
| DeepInfra Qwen3-235B-A22B-Instruct-2507 | 1 Jun 2026 → 3 Jul 2026 | 0.071 / 0.10 | 0.09 / 0.10 | +18% | before, after |
| DeepInfra gemma-4-31B-it | 1 Oct 2026 → 2 Nov 2026 | 0.15 / 0.40 | 0.20 / 0.40 | +18% | before, after |
| DeepInfra Qwen3-Coder-480B-A35B-Instruct-Turbo | 7 Apr 2026 → 2 May 2026 | 0.22 / 1.00 | 0.30 / 1.00 | +14% | before, after |
| DeepInfra Qwen3-14B | 2 Oct 2025 → 12 Nov 2025 | 0.06 / 0.24 | 0.08 / 0.24 | +14% | before, after |
| DeepInfra Llama-4-Scout-17B-16E | 1 Jun 2026 → 3 Jul 2026 | 0.08 / 0.30 | 0.10 / 0.30 | +11% | before, after |
| DeepInfra Meta-Llama-3.1-8B-Instruct-Turbo | 3 Jul 2026 → 2 Aug 2026 | 0.02 / 0.03 | 0.02 / 0.04 | +11% | before, after |
| DeepInfra gemma-4-31B-it | 1 Sep 2026 → 1 Oct 2026 | 0.13 / 0.38 | 0.15 / 0.40 | +10% | before, after |
| Mistral AI Mistral Medium | 5 May 2024 → 2 Aug 2024 (89 days apart) | 2.70 / 8.10 | 2.75 / 8.10 | +0.9% | before, after |
The 9 mixed changes (one price up, one down) are in the dataset, for example DeepSeek’s deepseek-chat, whose input rose from USD 0.27 to USD 0.28 while output fell from USD 1.10 to USD 0.42 between the September and October 2025 captures.
How should you budget for price rises?
The risk of a rise depends far more on who sells you the model than on the model. We counted rises per 100 model-years of observation (each model’s time from first to last capture, models seen in two or more months):
| Seller | Model-years observed | Rises | Rises per 100 model-years | Cuts per 100 model-years |
|---|---|---|---|---|
| Model makers (7) | 121.5 | 4 | 3.3 | 12.3 |
| Hosts other than DeepInfra | 54.4 | 8 | 14.7 | 18.4 |
| DeepInfra | 73.9 | 40 | 54.1 | 101.5 |
A practical rule from these figures:
- Price your plan at today’s rate, not a forecast cut. 80% of models never changed price in our captures. See how often LLM API prices change.
- Add headroom sized to the seller. With a maker’s own API, rises were rare. With a host, a rise of about 40% on one model within a year is a fair stress case; at the most active host there was about one rise for every two model-years.
- Keep a second provider tested. For open-weight models several hosts sell the same weights, often at very different prices. A tested fallback turns a rise into a config change. See LLM API failover and fallback and the same model at different hosts.
- Set a spend alert, not just a budget. A rise shows up first as a higher daily bill. LLM API spend control covers caps and alerts.
- Re-read the pricing page monthly for every model in production, or use a dataset that does it for you, such as the LLM API price index.
Methodology
- Source. One Internet Archive capture per provider per month of each provider’s own pricing page, 2023 to November 2026; 308 of 417 fetched captures had prices. Each price keeps its archive URL and a verbatim quote. Dataset: /data/llm-api-price-history/; every row: CSV.
- What counts as a rise. Two consecutive captures of the same model (same API ID or page name) where no price went down and at least one went up. Standard pay-as-you-go text prices only; where a page splits by context length, the short-context price; DeepSeek’s peak price where it lists peak and off-peak.
- Excluded. Moves into or out of a price the page marks as promotional or introductory, one zero-price page error and euro prices. An end of a labelled promotion is therefore never counted as a rise.
- Dates. Capture dates, not announcement dates. A rise took effect somewhere between the two captures.
- Rates. Model-years run from each model’s first to last capture. Makers are OpenAI, Anthropic, Google, DeepSeek, Mistral, xAI and Cohere; hosts are DeepInfra, Together AI, Groq and Fireworks AI. Our script recomputes every figure from the dataset JSON with the dataset’s own rules.
Where BroadNet Inference fits
BroadRouter is BroadNet’s own site. BroadNet Inference has no archived price history, so we cannot show whether its prices have moved. Its current per-token prices are in pricing.json. If a fixed cost matters more than the lowest rate, BroadNet’s Full Moon plan for DeepSeek V4.1 Flash is $60 a month per session with no token meter; one session runs one request at a time. Flat rate vs pay per token shows when that beats per-token billing.
