Research note 002 / API prices
What a million tokens cost: AI API prices, 2022–2026
Four years of list prices from six AI labs. OpenAI’s low-cost line fell 300-fold, its flagship about ninefold, and from 2025 new models often cost more per token than the ones they replaced.
In January 2022, a million tokens from OpenAI’s GPT-3 davinci cost $60. At our October 12, 2026 cutoff, a million tokens from GPT-6 Luna cost $0.20, blended three parts input to one part output.
That 300-fold fall is the headline, not the whole story. OpenAI’s flagship got about nine times cheaper, the lowest price in our data was set in August 2024, and through much of 2025 and 2026 new models launched at higher prices than the ones they replaced.
Four years in one chart
Each line follows one tier at one lab: OpenAI’s flagship and low-cost models, Claude Opus and Sonnet, Gemini Pro and Flash, DeepSeek’s main model, Grok, and Llama 70B as Together AI hosted it. A selected successor continues its tier’s line from the date recorded in the table, so a line shows the selected model’s price on that date.
How far did prices fall?
The steepest drops came first. OpenAI cut davinci from $60.00 to $20.00 on September 1, 2022, and GPT-3.5 Turbo arrived on March 1, 2023 at $2.00. GPT-4o mini took OpenAI’s low-cost tier to $0.26 in July 2024.
On August 12, 2024, Google cut Gemini 1.5 Flash to $0.13. No line in the chart has been cheaper since: the cheapest at our cutoff, GPT-6 Luna, is $0.20.
OpenAI’s flagship fell about ninefold
GPT-4 launched on March 14, 2023 at $30 per million input tokens and $60 per million output tokens: $37.50 blended. GPT-6.1 Sol costs $4.00, about a ninth as much.
Claude Opus fell less. Claude 3 Opus cost $30.00 in March 2024, and the price held through Opus 4.1. Opus 4.5 cut it by two-thirds, to $10.00, in November 2025; Opus 5.5 costs $8.00.
Above the flagships, a premium tier appeared. These models are not drawn in Figure 1.
| Model | Launched | Input | Output | Blended |
|---|---|---|---|---|
| GPT-4.5 (preview)OpenAI | $75.00 | $150.00 | $93.75 | |
| o1-proOpenAI | $150.00 | $600.00 | $262.50 | |
| Claude Fable 5Anthropic | $10.00 | $50.00 | $20.00 | |
| GPT-6 AstraOpenAI | $10.00 | $50.00 | $20.00 |
Claude Fable 5 and GPT-6 Astra both cost $20.00 blended: two and a half times Opus 5.5 and five times GPT-6.1 Sol.
The price war
DeepSeek V2 arrived on May 6, 2024 at $0.14 per million input tokens, $0.18 blended, and set off a price war among Chinese labs. [1] DeepSeek R1 followed on January 20, 2025 at $0.55 in and $2.19 out. OpenAI’s o1 cost $15 and $60, about 27 times as much. [2] On January 27, Nvidia lost $589 billion in market value in a day. [3]
When newer stopped meaning cheaper
From 2025, a new model often cost more per token than the one it replaced. GPT-5 mini launched at $0.69 in August 2025, more than double GPT-4o mini. GPT-5.5 listed at $11.25 in April 2026, twice GPT-5.4. Gemini 3.5 Flash arrived at $3.38 in May 2026, three times Gemini 3 Flash and 26 times Flash’s 2024 low.
The median price across the eight represented lines rose from $1.53 on January 1, 2025 to $4.66 a year later.
Llama’s line ended and Grok’s began in April 2025.
Then, on September 22, 2026, OpenAI and Anthropic cut prices on the same day. GPT-6 Sol launched at $4.00, half GPT-5.6 Sol’s promotional price; GPT-6 Luna at $0.20; Opus 5.5 at $8.00. The median at our cutoff is $3.50. [4] [5]
A cheaper token is not a cheaper task
These are prices per token. On the lines represented at our cutoff, output costs three to six times as much as input, and a reasoning model can write many more output tokens for the same answer. A lower price per token does not guarantee a cheaper task.
Subscriptions sell a usage allowance for a fixed monthly fee. Our plan comparison shows what Codex and Claude Code plans hold each month.
How we priced
- Blended price is (3 × input + output) ÷ 4: list prices in US dollars per million tokens, weighted three parts input to one part output.
- Prices are standard API list prices for uncached requests at the lowest prompt-length tier. Introductory and promotional prices count and are marked. Batch, cache and off-peak discounts, free tiers, credits and negotiated rates do not.
- Dates come from providers’ pricing pages, archives, release notes and announcements. They include model-entry dates and inferred dates, not only verified price-effective dates. January 1, 2022 is davinci’s opening state; Gemini 1.0 Pro’s price was listed before billing began. Notes identify capture intervals.
- Llama’s line is Together AI’s price for its 70B models, since Meta did not sell API access to them itself.
- Prices are rounded to the cent after blending. Ratios use unrounded prices. This is a fixed snapshot as of October 12, 2026.
Every price record
All 88 price records behind Figure 1, by line, with sources. Labels and notes preserve source and pricing caveats.
| Date | Model | Input | Output | Blended | Sources |
|---|---|---|---|---|---|
| OpenAI flagship / OpenAI | |||||
| GPT-48K contextThe 32K-context model cost $60 in and $120 out. API access opened with a waitlist. | $30.00 | $60.00 | $37.50 | ||
| GPT-4 TurboPreview | $10.00 | $30.00 | $15.00 | ||
| GPT-4o | $5.00 | $15.00 | $7.50 | ||
| GPT-4oPrice for the gpt-4o-2024-08-06 snapshot. The alias moved between September 19 and December 21, 2024; the exact day is unverified. GPT-4.1 is outside this selected line. | $2.50 | $10.00 | $4.38 | ||
| GPT-5 | $1.25 | $10.00 | $3.44 | ||
| GPT-5.1 | $1.25 | $10.00 | $3.44 | ||
| GPT-5.2 | $1.75 | $14.00 | $4.81 | ||
| GPT-5.4 | $2.50 | $15.00 | $5.63 | ||
| GPT-5.5 | $5.00 | $30.00 | $11.25 | ||
| GPT-5.6 SolSecondary sourceLaunch price from secondary reports. | $5.00 | $30.00 | $11.25 | ||
| GPT-5.6 SolPromotional pricePromotional price, stated to run at least through November 21, 2026. | $4.00 | $20.00 | $8.00 | ||
| GPT-6 Sol | $2.00 | $10.00 | $4.00 | ||
| GPT-6.1 Sol | $2.00 | $10.00 | $4.00 | ||
| OpenAI low-cost / OpenAI | |||||
| davinciOpening state on January 1, 2022, not a model launch date; davinci was already available. | $60.00 | $60.00 | $60.00 | ||
| davinci | $20.00 | $20.00 | $20.00 | ||
| text-davinci-003 | $20.00 | $20.00 | $20.00 | ||
| GPT-3.5 Turbo | $2.00 | $2.00 | $2.00 | ||
| GPT-3.5 TurboDate inferredPrice seen on a July 2023 capture; the date comes from OpenAI’s deprecations page. | $1.50 | $2.00 | $1.63 | ||
| GPT-3.5 Turbo | $1.00 | $2.00 | $1.25 | ||
| GPT-3.5 Turbo | $0.50 | $1.50 | $0.75 | ||
| GPT-4o mini | $0.15 | $0.60 | $0.26 | ||
| GPT-5 mini | $0.25 | $2.00 | $0.69 | ||
| GPT-5.4 mini | $0.75 | $4.50 | $1.69 | ||
| GPT-5.6 LunaSecondary sourceLaunch price from secondary reports, consistent with the 80% cut OpenAI’s changelog records. | $1.00 | $6.00 | $2.25 | ||
| GPT-5.6 Luna | $0.20 | $1.20 | $0.45 | ||
| GPT-6 Luna | $0.10 | $0.50 | $0.20 | ||
| Claude Opus / Anthropic | |||||
| Claude 1Converted to tokensLaunched per character; Anthropic’s April 2023 price sheet restates it per token at 3.8 characters per token. | $11.02 | $32.68 | $16.44 | ||
| Claude 2 | $11.02 | $32.68 | $16.44 | ||
| Claude 2.1 | $8.00 | $24.00 | $12.00 | ||
| Claude 3 Opus | $15.00 | $75.00 | $30.00 | ||
| Claude Opus 4 | $15.00 | $75.00 | $30.00 | ||
| Claude Opus 4.1 | $15.00 | $75.00 | $30.00 | ||
| Claude Opus 4.5 | $5.00 | $25.00 | $10.00 | ||
| Claude Opus 4.6Prompts up to 200K tokens | $5.00 | $25.00 | $10.00 | ||
| Claude Opus 4.7Prompts up to 200K tokensThe release notes report about 30% more tokens for the same text with the new tokenizer. | $5.00 | $25.00 | $10.00 | ||
| Claude Opus 4.8 | $5.00 | $25.00 | $10.00 | ||
| Claude Opus 5 | $5.00 | $25.00 | $10.00 | ||
| Claude Opus 5.5 | $4.00 | $20.00 | $8.00 | ||
| Claude Sonnet / Anthropic | |||||
| Claude 3 Sonnet | $3.00 | $15.00 | $6.00 | ||
| Claude 3.5 Sonnet | $3.00 | $15.00 | $6.00 | ||
| Claude 3.5 Sonnet (Oct 2024) | $3.00 | $15.00 | $6.00 | ||
| Claude 3.7 Sonnet | $3.00 | $15.00 | $6.00 | ||
| Claude Sonnet 4 | $3.00 | $15.00 | $6.00 | ||
| Claude Sonnet 4.5 | $3.00 | $15.00 | $6.00 | ||
| Claude Sonnet 4.6Prompts up to 200K tokens | $3.00 | $15.00 | $6.00 | ||
| Claude Sonnet 5Promotional priceIntroductory price; made standard on August 10, 2026. | $2.00 | $10.00 | $4.00 | ||
| Claude Sonnet 5.5 | $2.00 | $10.00 | $4.00 | ||
| Gemini Pro / Google | |||||
| Gemini 1.0 ProConverted to tokensVertex AI priced per character; converted at 4 characters per token. Charged from January 15, 2024. | $1.00 | $2.00 | $1.25 | ||
| Gemini 1.0 ProConverted to tokensDate inferredFirst seen on a March 10, 2024 capture, after December 20, 2023. Converted at 4 characters per token. | $0.50 | $1.50 | $0.75 | ||
| Gemini 1.5 ProPrompts up to 128K tokensDate inferredPreviewPrice seen on a June 3, 2024 capture; billing began around Google I/O. | $3.50 | $10.50 | $5.25 | ||
| Gemini 1.5 ProPrompts up to 128K tokens | $1.25 | $5.00 | $2.19 | ||
| Gemini 2.5 ProPrompts up to 200K tokensSecondary sourcePreviewPaid preview price reported by Simon Willison. | $1.25 | $10.00 | $3.44 | ||
| Gemini 3 ProPrompts up to 200K tokensPreview | $2.00 | $12.00 | $4.50 | ||
| Gemini 3.1 ProPrompts up to 200K tokensSecondary sourcePreviewStill in preview at the cutoff; price reported by Simon Willison. | $2.00 | $12.00 | $4.50 | ||
| Gemini Flash / Google | |||||
| Gemini 1.5 FlashPrompts up to 128K tokensDate inferredPreviewPrice seen on a June 3, 2024 capture; generally available May 23, 2024. | $0.35 | $1.05 | $0.53 | ||
| Gemini 1.5 FlashPrompts up to 128K tokens | $0.075 | $0.30 | $0.13 | ||
| Gemini 2.0 Flash | $0.10 | $0.40 | $0.18 | ||
| Gemini 2.5 FlashSecondary sourcePreviewOutput without thinking; thinking output cost $3.50. Price reported by Simon Willison. | $0.15 | $0.60 | $0.26 | ||
| Gemini 2.5 Flash | $0.30 | $2.50 | $0.85 | ||
| Gemini 3 Flash | $0.50 | $3.00 | $1.13 | ||
| Gemini 3.5 Flash | $1.50 | $9.00 | $3.38 | ||
| Gemini 3.6 Flash | $1.50 | $7.50 | $3.00 | ||
| Gemini 3.7 FlashPromotional priceIntroductory price through December 31, 2026; $1.50 in and $7.50 out from 2027. | $0.75 | $3.75 | $1.50 | ||
| Gemini 3.8 FlashPromotional priceSame introductory price as Gemini 3.7 Flash: through December 31, 2026; $1.50 in and $7.50 out from 2027. | $0.75 | $3.75 | $1.50 | ||
| DeepSeek / DeepSeek | |||||
| DeepSeek-V2 | $0.14 | $0.28 | $0.18 | ||
| DeepSeek-V2.5 | $0.14 | $0.28 | $0.18 | ||
| DeepSeek-V3Promotional pricePromotional V2 price until February 8, 2025; list price $0.27 in and $1.10 out. | $0.14 | $0.28 | $0.18 | ||
| DeepSeek-R1 | $0.55 | $2.19 | $0.96 | ||
| DeepSeek-V3.1Secondary sourceReleased August 21, 2025; one price for both modes from September 5, per secondary reports. | $0.56 | $1.68 | $0.84 | ||
| DeepSeek-V3.2-Exp | $0.28 | $0.42 | $0.32 | ||
| DeepSeek-V3.2 | $0.28 | $0.42 | $0.32 | ||
| DeepSeek-V4-Flash | $0.14 | $0.28 | $0.18 | ||
| DeepSeek-V4-FlashPeak price, 01:00–04:00 and 06:00–10:00 UTC; off-peak discounts are excluded. | $0.44 | $1.32 | $0.66 | ||
| DeepSeek-V4.1-FlashPeak price, Monday–Friday; off-peak discounts are excluded. | $0.30 | $1.20 | $0.53 | ||
| Llama 70B at Together AI / Meta | |||||
| Llama 2 70B | $3.00 | $3.00 | $3.00 | ||
| Llama 2 70BDate inferredChanged between captures of August 12 and October 17, 2023. | $1.00 | $1.00 | $1.00 | ||
| Llama 2 70B | $0.90 | $0.90 | $0.90 | ||
| Llama 3 70B | $0.90 | $0.90 | $0.90 | ||
| Llama 3 70BTogether AI’s default endpoint changed from Reference (FP16) at $0.90 to Turbo (FP8) at $0.88; Reference stayed $0.90. | $0.88 | $0.88 | $0.88 | ||
| Llama 3.1 70B | $0.88 | $0.88 | $0.88 | ||
| Llama 3.3 70BPrice inferredTogether AI’s 70B price; Llama 3.3 is first named beside it on a January 31, 2025 capture. | $0.88 | $0.88 | $0.88 | ||
| Grok / xAI | |||||
| Grok 3Secondary sourceAPI launch date from TechCrunch; price confirmed in xAI’s price table. | $3.00 | $15.00 | $6.00 | ||
| Grok 4Prompts up to 128K tokens | $3.00 | $15.00 | $6.00 | ||
| Grok 4.20 | $2.00 | $6.00 | $3.00 | ||
| Grok 4.3Date inferredChanged between captures of April 28 and May 6, 2026. | $1.25 | $2.50 | $1.56 | ||
| Grok 4.5 | $2.00 | $6.00 | $3.00 | ||
| Grok 4.6 | $2.00 | $6.00 | $3.00 | ||
| Grok 4.7 | $2.00 | $6.00 | $3.00 | ||