Every AI pricing change in Q4 2025, tracked

Q4 2025 was the most active pricing quarter of the year. 14 changes from 9 providers.

The headline: Google dropped Gemini 2.5 Flash to $0.05 per million input tokens. That's five cents for a million tokens. The race to the bottom continues.

Q4 2025 pricing changes

| Provider | Model | Change | Old input/M | New input/M | |----------|-------|--------|-------------|-------------| | Google | Gemini 2.5 Flash | Cut | $0.15 | $0.05 | | Google | Gemini 2.5 Pro | Cut | $1.25 | $0.80 | | OpenAI | GPT-4o | Cut | $2.50 | $2.00 | | OpenAI | GPT-4o mini | Cut | $0.15 | $0.10 | | Anthropic | Claude 4 Sonnet | Cut | $3.00 | $2.50 | | DeepSeek | DeepSeek R2 | New | N/A | $0.20 | | DeepSeek | DeepSeek V4 | New | N/A | $0.15 | | Mistral AI | Mistral Large 3 | Cut | $2.00 | $1.50 | | Mistral AI | Mistral Small 2 | Cut | $0.10 | $0.06 | | Groq | Various | Cut | Various | ~15% reduction | | Together AI | Llama hosting | Cut | Various | ~20% reduction | | Fireworks AI | Open source hosting | Cut | Various | ~15% reduction | | xAI | Grok 3 | Cut | $3.00 | $2.50 | | Cerebras | Inference | Cut | Various | ~25% reduction |

Sources: Provider pricing pages and announcements, Q4 2025.

Every single change was a price cut. Not a single increase. The entire market moved down.

Updated master comparison table (January 2026)

| Model | Provider | Input/M | Output/M | Quality tier | |-------|----------|---------|----------|-------------| | Gemini 2.5 Flash | Google | $0.05 | $0.20 | Mid-high | | Mistral Small 2 | Mistral AI | $0.06 | $0.18 | Mid | | GPT-4o mini | OpenAI | $0.10 | $0.40 | Mid | | DeepSeek V4 | DeepSeek | $0.15 | $0.60 | High | | DeepSeek R2 | DeepSeek | $0.20 | $0.80 | High (reasoning) | | Gemini 2.5 Pro | Google | $0.80 | $6.00 | High | | Mistral Large 3 | Mistral AI | $1.50 | $4.50 | High | | GPT-4o | OpenAI | $2.00 | $8.00 | High | | Claude 4 Sonnet | Anthropic | $2.50 | $12.50 | High | | Grok 3 | xAI | $2.50 | $12.50 | High | | Claude Opus 4 | Anthropic | $15.00 | $75.00 | Flagship |

Sources: OpenAI, Anthropic, Google, DeepSeek, Mistral AI, xAI, Artificial Analysis.

The cheapest input price is now $0.05/M (Gemini 2.5 Flash). The most expensive is $15.00/M (Claude Opus 4). A 300x spread.

Quarterly price deflation rate

| Quarter | Avg frontier price change | Avg economy price change | |---------|--------------------------|------------------------| | Q1 2025 | -12% | -8% | | Q2 2025 | -18% | -15% | | Q3 2025 | -10% | -5% | | Q4 2025 | -17% | -33% |

The economy tier saw the sharpest cuts in Q4 (-33%), driven by Google's aggressive Flash pricing. The frontier tier cut 17%, led by OpenAI and Anthropic lowering their main workhorse models.

What $0.05/M input tokens means

At $0.05 per million input tokens:

| Use case | Tokens | Cost | |----------|--------|------| | Process 1,000 emails (classification) | 200K | $0.01 | | Read a 300-page book | 200K | $0.01 | | Summarize 100 articles | 500K | $0.025 |

One cent to classify a thousand emails. One cent to read a book. We're approaching "essentially free" for input tokens.

Output tokens are still 4x more expensive ($0.20/M for Flash), but even there, generating a 500-word response costs about $0.00014.

My pricing floor prediction update

| Tier | Current price | Predicted floor | When | |------|-------------|----------------|------| | Economy input | $0.05/M | $0.02/M | Mid-2026 | | Economy output | $0.20/M | $0.08/M | Late 2026 | | Frontier input | $2.00/M | $0.50/M | 2027 | | Frontier output | $8.00/M | $2.00/M | 2027 |

Sources: My analysis of compute costs, hardware trends.

I'm updating my floor estimates downward again. The pricing compression shows no signs of stopping.

My master pricing spreadsheet has 89 rows now. Every few months I think "surely this is the floor." Every few months I'm wrong.

---

If you found this interesting, you might also like:

• The cost of running an AI startup in 2022: a data breakdown
• Every AI pricing change in January 2025, tracked
• The cost of AI dropped 97% in two years. One chart.
• 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

-- dataku