Pricing Trends of LLMs by Major Providers (2023-2025)

Source:https://chatgpt.com/share/67a4cf07-efec-8000-ad83-486163512568

Great! I will gather data on the pricing trends of LLMs released between 2023 and 2025 by OpenAI, Google, DeepSeek, and Anthropic. The research will include the price per million tokens for both input and output tokens, covering models such as GPT-4o, GPT-4o-mini, O1-Preview, O3, DeepSeek-R1, Claude 3.5 Sonnet, Gemini 1.5, Gemini 2.0, Gemini 2.0 Flash-Lite, and Gemini 2.0 Pro. Additionally, I will compile a table detailing each model’s context window, model version, and size (if available). I will generate two separate charts for input and output token pricing to visualize the trends clearly. I'll update you once I have the results.

Pricing Trends of 2023–2025 LLMs by OpenAI, Google, DeepSeek, and Anthropic

Between 2023 and 2025, major AI providers released new large language models (LLMs) with varying capabilities and significantly different pricing structures. Below we outline the price per million tokens for input (prompt) and output (generated) tokens for notable models from OpenAI, Google, DeepSeek, and Anthropic. We also provide each model’s context window and model size (where available), followed by comparative breakdowns of input and output token pricing by provider. All information is sourced from up-to-date, publicly available references.

Model Details and Context Windows

The table below summarizes each model’s version, context window (maximum tokens in one request), and size (number of parameters, if disclosed):

Model

Context Window

Model Size

OpenAI GPT-4o

n128,000 tokenswww.deeplearning.ai

Not publicly disclosed (multimodal flagship model)

OpenAI GPT-4o Mini

n128,000 tokenswww.deeplearning.ai

~8 billion parameters (approx.)www.deeplearning.ai

OpenAI O1 (Preview)

n200,000 tokens (input)docsbot.aidocsbot.ai
n100,000 tokens (output)docsbot.aidocsbot.ai

Not disclosed (advanced reasoning model)

OpenAI O3

n200,000 tokens (input)www.vals.ai
n100,000 tokens (output)www.vals.ai

Not disclosed (small “reasoning” model)

DeepSeek R1

n128,000 tokenswww.plainconcepts.com

n671B parameters (Mixture-of-Experts; ~37B active per token)blog.promptlayer.com

Anthropic Claude 3.5 Sonnet

n200,000 tokenswww.anthropic.com

Not disclosed (Claude 3.5 family model)

Google Gemini 1.5

n2,000,000 tokensai.google.dev(long-context Pro model)

Not disclosed (Gemini 1.5 Pro; “Flash” variant is 8B)

Google Gemini 2.0 Flash-Lite

Large (≥1M) – in preview (optimized for long text output)

Not disclosed (cost-optimized smaller model)

Google Gemini 2.0 Pro

Large (context details TBD)

Not disclosed (production model with multimodal support)

Table 1: Key context window and size information for each model. Models with extremely large contexts (e.g. OpenAI’s O1 and Google’s Gemini Pro series) use special architectures (like chain-of-thought or flash attention) to handle long inputs. Model sizes are only approximate where noted, as many providers do not publicly reveal parameter counts.

Price per Million Input Tokens (Prompt) – Comparison by Provider

The cost per 1 million input tokens varies widely across these models. The list below groups models by provider, showing their input pricing (in USD per million tokens):

Insights (Input Pricing): OpenAI’s advanced reasoning model O1 has the highest input cost at $15/M, reflecting its powerful but expensive chain-of-thought reasoning

docsbot.ai. In contrast, OpenAI’s smaller GPT-4o Mini is extremely cheap (0.15/M)[deeplearning.ai](https://www.deeplearning.ai/the-batch/openais-gpt-4o-mini-offers-big-performance-at-a-small-price/#:~:text=%2A%20API%20access%20to%20GPT,3.5), enabling cost-effective use. Google’s Gemini models have become very price-competitive by late 2024–2025: **Gemini 2.0 Flash-Lite’s** input cost is only n0.075 per million tokensai.google.dev– the lowest among all models listed. Even Google’s high-performance Gemini 2.0 (Pro) model is just 0.10/M input[ai.google.dev](https://ai.google.dev/pricing#:~:text=Input%20Pricing), undercutting OpenAI’s flagship GPT-4o by a factor of 50×. DeepSeek’s R1 is also extremely affordable at n0.14/M, leveraging an open-source approach to drive down costswww.plainconcepts.com. Anthropic’s Claude 3.5 Sonnet sits in the mid-range at $3.00/M inputwww.anthropic.com, cheaper than OpenAI’s GPT-4o but higher than Google’s latest offerings.

Price per Million Output Tokens (Completion) – Comparison by Provider

The cost per 1 million output tokens (model-generated tokens) is generally higher than input across all providers, but with similar relative trends:

Insights (Output Pricing): OpenAI’s O1 carries the steepest output token price at $60/M

docsbot.ai, reflecting the intensive computation it performs for each generated token. OpenAI’s GPT-4o output cost (15/M)isequaltoAnthropic’sClaude3.5Sonnet(15/M) is equal to Anthropic’s Claude 3.5 Sonnet (n15/M)www.anthropic.com, indicating these top-tier models charge a premium for generation. However, OpenAI’s smaller models and new reasoning model show much lower rates: GPT-4o Mini outputs cost only 0.60/M[deeplearning.ai](https://www.deeplearning.ai/the-batch/openais-gpt-4o-mini-offers-big-performance-at-a-small-price/#:~:text=%2A%20API%20access%20to%20GPT,3.5), and **O3** (OpenAI’s small reasoning model) is n4.40/Mwww.vals.ai– a dramatic 93% reduction compared to O1’s rate, aligning with OpenAI’s claim that O3-mini is 93% cheaper than O1community.openai.comcommunity.openai.com.

Google’s pricing again is highly competitive on output tokens. Gemini 2.0 Flash-Lite is just $0.30 per million output tokens

ai.google.dev– the lowest output cost among these models – while Gemini 2.0 (Pro) is 0.40/M[ai.google.dev](https://ai.google.dev/pricing#:~:text=Input%20Pricing). These are **orders of magnitude cheaper** than OpenAI and Anthropic’s comparable models. Even Google’s earlier **Gemini 1.5 Pro** had a moderate n5.00/M output costai.google.dev(which was further reduced in late 2024 as part of pricing cuts). DeepSeek R1 again offers rock-bottom pricing at $0.28/M for outputswww.plainconcepts.com, making it an attractive low-cost alternative if its performance meets the needs.

Key Takeaways

  • OpenAI’s trend: The introduction of GPT-4o in 2024 significantly lowered costs relative to GPT-4 (which was about 30/Minput,30/M input, n60/M output) – GPT-4o is priced at 5/5/n15

    www.deeplearning.aiper million tokens, a major reduction. Additionally, GPT-4o Mini made usage even cheaper for high-token scenarios (only 0.15/0.15/n0.60 per 1M)www.deeplearning.ai. However, OpenAI’s reasoning-optimized models (O1 and O3 series) came at a premium for O1 (the preview was very costly at 15/15/n60 per 1Mdocsbot.aidocsbot.ai) and then a huge drop for O3-mini as a cost-efficient alternative (1.10/1.10/n4.40www.vals.ai). This reflects a trend of offering specialized high-performance models at high cost, alongside smaller versions to make certain capabilities accessible.

  • Google’s trend: Google aggressively improved Gemini model pricing and context capabilities over time. Gemini 1.5 Pro launched with a 2M token context window

    ai.google.devand was priced at 1.25/1.25/n5.00 per 1M tokensai.google.devfor typical contexts, with further reductions announced in late 2024developers.googleblog.com. By early 2025, Google’s Gemini 2.0 Flash-Lite (a lightweight model) and Gemini 2.0 (production “Flash” model) have very low pricing (0.075–0.075–n0.10 per 1M input; 0.30–0.30–n0.40 per 1M output)ai.google.devai.google.dev, undercutting most competitors. Google’s strategy shows a clear downward trend in token prices, likely to drive adoption, while providing industry-leading context lengths (1M–2M tokens) for enterprise use cases.

  • DeepSeek’s trend: DeepSeek R1 emerged as a disruptively low-cost model in 2025, with input/output pricing in the sub-$0.30 range

    www.plainconcepts.com. Its novel Mixture-of-Experts architecture (671B parameters with selective activation) enables strong performance in math/coding while keeping costs about 100–200× cheaper than OpenAI’s O1www.plainconcepts.com. This highlights a trend of new entrants using innovative techniques to compete on price and openness.

  • Anthropic’s trend: Anthropic’s Claude 3.5 Sonnet (released mid-2024) maintained moderate pricing (3/Minput,3/M input, n15/M output)

    www.anthropic.comsimilar to its earlier Claude 2 models, but with a large 200K context window. Anthropic positioned Claude 3.5 as a faster, cost-effective mid-tier model that still beats many competitors on qualitywww.anthropic.comwww.anthropic.com. Compared to OpenAI and Google, Anthropic’s pricing has been relatively stable, not dropping as steeply – suggesting they compete more on model capability and the generous context length rather than on being the cheapest per token.

Overall, token prices have trended downward from 2023 to 2025, especially for input tokens, as providers optimize models and scale up infrastructure. OpenAI and Anthropic introduced expensive high-end models but also cheaper variants, while Google and DeepSeek significantly pushed prices lower across the board. These pricing trends, alongside context window expansions (up to millions of tokens), are making large-scale AI much more accessible for developers and businesses

developers.googleblog.comwww.plainconcepts.com.