LLM API Price Explorer

Data as of 2026-11-03. Data version 5. Prices are as shown on the vendor's pricing page on the date given for each point: 2026-10-07 for the live pages, the capture date for Wayback Machine snapshots. Three vendor announcement posts add dated figures. The OpenAI table rows of 2026-10-07 were read on 2026-10-09 from the capture of that day, and the DeepSeek change log (S38b) was read on 2026-10-09. On 2026-10-20 all six vendors' live pages were read again. A model has a point on 2026-10-20 only where a vendor page showed something new that day. Those are Claude Haiku 5.5, the Claude Sonnet 5.5 row (its cache price changed), Gemini 4 Argon, and the two OpenAI life sciences models (one new, one moved). Every other price on the pages was unchanged, so its latest point stays on 2026-10-07. On 2026-11-03 the live pages were read again. A model has a point on 2026-11-03 only where a vendor page showed something new that day. Those are OpenAI gpt-6.1-terra, gpt-6-rosalind and the promotional gpt-6-astra-law, Mistral Large 4 (a new price) and the Gemini 3.5 Flash-Lite Batch caching price. Two third-party models that Mistral hosts have host rows that are not published. Every other price was unchanged, so its latest point stays on its earlier date.

What the chart shows

• Vendor list prices in USD per 1M tokens at the standard pay-as-you-go rate and short context. Input and output are separate views.
• One line per model family. A family is one model generation: preview and GA releases, dated snapshots and the alias that points to them. Example: family gpt-4o holds gpt-4o, gpt-4o-2024-05-13, gpt-4o-2024-08-06 and gpt-4o-2024-11-20. Point details show the model name exactly as the vendor wrote it.
• The line follows the cheapest model in the family on each observed date. A hollow point is another model in the same family listed at a higher price on that date.
• A point is the date of a snapshot or announcement, not the day a price changed. The line steps at the later snapshot, because the change took effect at some time between the two captures.
• 2023 prices quoted per 1K tokens were multiplied by 1,000. No price is rounded. No exchange rate is applied, so Mistral AI's EUR and DeepSeek's CNY prices are not drawn.
• One rule: the series value is the list price in effect on the observation date. Promotions and scheduled changes appear as notes with their dates in the dataset and are not chart values. Examples: DeepSeek's off-peak rate and its 2025 deepseek-chat discount; Mistral Large 4's sale price; OpenAI gpt-5.6-sol, which the page shows only at a promotional price, so it has no chart value. Google listed Gemini 3.6, 3.7 and 3.8 Flash at $0.75 input and $3.75 output through 2026-12-31 and at $1.50 and $7.50 from 2027-01-01. The chart shows $0.75 and $3.75 on 2026-10-07; the 2027 price is a scheduled change and is not drawn. Mistral Large 4 showed a sale price until 2026-10-20 and shows one price on 2026-11-03. Gemini 4 Argon is listed at $2.00 input and $10.00 output through 2026-11-30, then $4.00 and $20.00 from 2026-12-01. The chart shows $2.00 and $10.00 on 2026-10-20. The December price is a scheduled change and is not drawn.
• Closed to new customers: where the vendor's own deprecations page says only existing users can use a model, the point has a dashed outline and the table says so. OpenAI did this for gpt-4-32k, gpt-4-vision-preview and three gpt-3.5-turbo snapshots in 2024. Google did it for Gemini 2.5 Pro, 2.5 Flash and 2.5 Flash-Lite, seen on the page on 2026-10-07 and again on 2026-10-20 and 2026-11-03. The quarter-end tier series never use such a model. Models with no such statement on a page that records them are open; for other dates and vendors the status is unknown.
• Percent change = (latest price - launch price) / launch price x 100, rounded to one decimal place. A family seen only once shows 0.0%.
• Release dates come from model ID date suffixes, dates stated by the vendor, version numbers that encode the month, or announcement posts. Retirement dates come from vendor deprecation pages. A blank means the source ledger holds no date. A retirement date after 2026-11-03 is a scheduled shutdown. The Released column names the basis of each start date. The bases are a launch post and a model ID date suffix (a build date that can come before public availability). The others are the release date column of the vendor's deprecations page and a version number that encodes a month. Gemini 3.5 Flash and 3.7 Flash are no longer on Google's pricing page (2026-10-20 and 2026-11-03); their earlier points stay and no retirement date is recorded. Mistral Large 3 is no longer on Mistral's pricing page (2026-11-03); Mistral's model table lists its retirement for 2027-04-30. OpenAI's deprecations page records the shutdown of gpt-4, gpt-4-turbo, gpt-3.5-turbo, gpt-4.1-nano, o1, o3-mini and o4-mini on 2026-10-23.

Not included

• Host and reseller prices: OpenRouter, Together AI, Amazon Bedrock and Azure OpenAI. Llama has no first-party API price, so it is not in the chart.
• Secondary analyses (Epoch AI, arXiv) and the derived quarter-end series.
• Batch, flex, fast, cached input, cache write, priority, long context, off-peak and promotional prices, scheduled future prices, fine-tuning, search and storage fees.
• Negotiated, committed-use and enterprise rates, free tiers and regional uplifts.
• Per-token prices are not per-text prices. Tokenizers differ by vendor and by model generation.

Tier definitions

Frontier and small follow each vendor's own naming. "All models" adds models neither tier covers, such as OpenAI o-series reasoning models, pro and specialised variants, and Mistral AI's open and edge models.

Sources

Every point carries one of these keys. Keys whose last column says "dates only" supply release or retirement dates. The companion page lists the same sources with their retrieval dates.

Ledger keys not used here:

• S37, Mistral AI via Wayback: snapshot without an API price table.
• S44, xAI via Wayback: empty snapshot.
• S46, Together AI (host): host price (Together AI).
• S47, Fireworks AI (host): host, no token price on the page.
• S49, Amazon Web Services (host): host price (Amazon Bedrock).
• S50, Microsoft (host): host price (Azure OpenAI).
• S51, Epoch AI (secondary): secondary analysis (Epoch AI).
• S52, arXiv, Mingdeng Du (secondary): secondary analysis (arXiv).
• S53, OpenRouter (secondary, host aggregator): host aggregator (OpenRouter).
• S54, OpenRouter (secondary, host aggregator): host aggregator (OpenRouter).

Read the key findings, charts and full methodology on the companion page: LLM API Price History Tracker.