Every AI pricing change in January 2025, tracked

January 2025 was a busy month for AI pricing. Seven providers made changes. I updated my master comparison table.

Quick summary: everything got cheaper. Again.

January 2025 price changes

| Provider | Model | Direction | Old price (input/M) | New price (input/M) | Change | |----------|-------|-----------|---------------------|---------------------|--------| | Anthropic | Claude 3.5 Haiku | Down | $0.80 | $0.80 | Held (output cut to $4) | | Google | Gemini 2.0 Flash | New | N/A | $0.10 | New model | | OpenAI | GPT-4o mini | Held | $0.15 | $0.15 | No change | | DeepSeek | DeepSeek V3 | New | N/A | $0.27 | New model | | DeepSeek | DeepSeek R1 | New | N/A | $0.55 | New model | | Mistral AI | Mistral Small | Down | $0.20 | $0.10 | -50% | | Groq | Various | Down | Varies | Varies | ~20-30% cuts |

Sources: Provider pricing pages and changelogs, January 2025.

The big news: Gemini 2.0 Flash at $0.10 per million input tokens. That's a frontier-capable model for essentially nothing. DeepSeek V3 at $0.27/M is almost as cheap and performs at a higher tier.

Mistral quietly cut Mistral Small by 50%. Groq reduced inference pricing across most of their hosted models by 20-30%.

The master pricing table (February 2025)

| Model | Provider | Input/M | Output/M | Quality tier | |-------|----------|---------|----------|-------------| | Gemini 2.0 Flash | Google | $0.10 | $0.40 | Mid-high | | Mistral Small | Mistral AI | $0.10 | $0.30 | Mid | | GPT-4o mini | OpenAI | $0.15 | $0.60 | Mid | | DeepSeek V3 | DeepSeek | $0.27 | $1.10 | High | | DeepSeek R1 | DeepSeek | $0.55 | $2.19 | High (reasoning) | | Claude 3.5 Haiku | Anthropic | $0.80 | $4.00 | Mid-high | | GPT-4o | OpenAI | $2.50 | $10.00 | High | | Claude 3.5 Sonnet | Anthropic | $3.00 | $15.00 | Highest | | Grok 3 | xAI | $3.00 | $15.00 | High | | Gemini 2.0 Pro | Google | $3.50 | $10.50 | High | | Claude 3 Opus | Anthropic | $15.00 | $75.00 | Legacy flagship | | GPT-4 Turbo | OpenAI | $10.00 | $30.00 | Legacy |

Sources: Anthropic, OpenAI, Google, DeepSeek, Mistral AI, Groq, xAI.

The spread between cheapest and most expensive: 150x on input tokens ($0.10 vs $15.00). A year ago, the spread was about 80x.

What $0.10 per million tokens actually means

Let me put Gemini 2.0 Flash's pricing in context:

| Use case | Tokens | Cost at $0.10/M | |----------|--------|-----------------| | Summarize a 10-page document | ~5,000 | $0.0005 | | Classify 1,000 emails | ~200,000 | $0.02 | | Generate 100 product descriptions | ~50,000 | $0.005 | | Process a full book | ~200,000 | $0.02 |

Two cents to process an entire book. Five hundredths of a cent to summarize a document.

At these prices, the cost of the API call is essentially a rounding error. Infrastructure, networking, and application logic cost more than the AI inference itself.

The deflation trend

| Date | Cheapest frontier-capable model | Input price/M | |------|-------------------------------|---------------| | Mar 2023 | GPT-3.5 Turbo | $2.00 | | Nov 2023 | GPT-3.5 Turbo (updated) | $1.00 | | Jul 2024 | GPT-4o mini | $0.15 | | Dec 2024 | Gemini 2.0 Flash | $0.10 | | Jan 2025 | Gemini 2.0 Flash / Mistral Small | $0.10 |

Sources: Historical pricing pages, provider announcements.

In 22 months, the cheapest frontier-capable model went from $2.00/M to $0.10/M. A 20x reduction.

And "frontier-capable" isn't what it used to be. Gemini 2.0 Flash at $0.10/M is comparable in quality to what GPT-4 was at $30/M two years ago.

The price floor keeps dropping. I keep updating the spreadsheet. My "cheapest useful model" cell has been overwritten 11 times in the last 18 months.

---

If you found this interesting, you might also like:

• The cost of running an AI startup in 2022: a data breakdown
• The real cost of training Llama 2: Meta's numbers vs my estimates
• Every LLM API price drop in the last 12 months, in one chart
• Wait, GPT-3 costs HOW much per token?
• Codex and the cost of code generation: my first pricing analysis

-- dataku