Pricing Comparison
| Model | Provider | Input $/M | Output $/M | Context | Quality |
|---|---|---|---|---|---|
| Mistral Small | Mistral | $0.10 | $0.30 | 128K | 68 |
| Llama 4 Scout | Meta | $0.10 | $0.30 | 512K | 79 |
| GPT-4o mini | OpenAI | $0.15 | $0.60 | 128K | 78 |
| Gemini 2.5 Flash | $0.15 | $0.60 | 1M | 82 | |
| Llama 4 Maverick | Meta | $0.20 | $0.60 | 1M | 86 |
| Mistral Medium | Mistral | $0.40 | $1.20 | 128K | 76 |
| Claude Haiku 4.5 | Anthropic | $1.00 | $5.00 | 200K | 80 |
| o3-mini | OpenAI | $1.10 | $4.40 | 200K | 88 |
| Gemini 2.5 Pro | $1.25 | $5.00 | 1M | 94 | |
| GPT-5 | OpenAI | $2.00 | $8.00 | 256K | 92 |
| Mistral Large | Mistral | $2.00 | $6.00 | 128K | 84 |
| GPT-4o | OpenAI | $2.50 | $10.00 | 128K | 87 |
| Claude Sonnet 5 | Anthropic | $3.00 | $15.00 | 1M | 93 |
| Grok 3 | xAI | $3.00 | $15.00 | 128K | 90 |
| Claude Opus 5 | Anthropic | $5.00 | $25.00 | 1M | 95 |
| o3 | OpenAI | $10.00 | $40.00 | 200K | 96 |
Prices are per million tokens via each provider's API. Quality scores are composite benchmark estimates (MMLU, HumanEval, GPQA, etc.). Last updated August 2026.
Input Cost per Million Tokens
Quality vs. Average Cost
Top-left = best value (high quality, low cost)
Cost Calculator
Estimate your monthly AI spend across models
| Model | Monthly Cost |
|---|---|
| Mistral SmallMistral | $0.20 |
| Llama 4 ScoutMeta | $0.20 |
| GPT-4o miniOpenAI | $0.38 |
| Gemini 2.5 FlashGoogle | $0.38 |
| Llama 4 MaverickMeta | $0.40 |
| Mistral MediumMistral | $0.80 |
| o3-miniOpenAI | $2.75 |
| Claude Haiku 4.5Anthropic | $3.00 |
| Gemini 2.5 ProGoogle | $3.13 |
| Mistral LargeMistral | $4.00 |
| GPT-5OpenAI | $5.00 |
| GPT-4oOpenAI | $6.25 |
| Claude Sonnet 5Anthropic | $9.00 |
| Grok 3xAI | $9.00 |
| Claude Opus 5Anthropic | $15.00 |
| o3OpenAI | $25.00 |
*Assumes ~4K tokens per request average
Provider Deep Dive
OpenAI
Five models spanning reasoning (o3) to budget (4o mini). GPT-5 launched May 2026 at $2/$8 — 60% cheaper than GPT-4 while outperforming it. o3 remains the most expensive at $10/$40 but leads on hard reasoning benchmarks.
Anthropic
Three tiers: Opus 5 ($5/$25) for maximum capability, Sonnet 5 ($3/$15, intro $2/$10) for the best quality-per-dollar at frontier level, and Haiku 4.5 ($1/$5) for speed-sensitive workloads. All share a 1M context window except Haiku (200K).
Two models, both with 1M context windows. Gemini 2.5 Pro ($1.25/$5) competes head-to-head with Claude Sonnet 5. Gemini 2.5 Flash ($0.15/$0.60) is the cheapest frontier-class model available — the default for cost-sensitive production use.
Mistral
European AI lab offering three tiers. Mistral Small at $0.10/$0.30 is the cheapest proprietary model on the market. Large ($2/$6) targets enterprise workloads. All models have 128K context windows.
Meta
Open-weight Llama 4 family — Maverick (400B MoE, $0.20/$0.60) and Scout (109B MoE, $0.10/$0.30) via inference providers. Pricing varies by provider (Together AI, Fireworks, etc.). Self-hosting eliminates per-token costs entirely.
xAI
Grok 3 ($3/$15) is xAI's flagship with 128K context. Competitive with Claude Sonnet 5 on benchmarks at the same price point. Integrated into X (Twitter) with real-time data access via the xAI API.
Pricing Trends
Key shifts in AI model pricing since 2024
90% Cost Reduction in 2 Years
GPT-4's launch pricing in 2023 was $30/$60 per million tokens. GPT-5 in 2026 costs $2/$8 — a 93% reduction at the input tier while being significantly more capable. Every provider has followed the same trajectory.
The Flash/Mini Tier Standardized
Every major provider now offers a sub-$1/M input model: GPT-4o mini ($0.15), Gemini 2.5 Flash ($0.15), Mistral Small ($0.10), Llama 4 Scout ($0.10), Claude Haiku 4.5 ($1). This tier handles 80%+ of production workloads.
Context Windows Exploded
1M token context is now standard at the frontier (Claude, Gemini). This was science fiction in 2024 when 128K was cutting-edge. Longer context windows reduce the need for RAG in many applications.
Reasoning Premium Emerged
OpenAI's o3 at $10/$40 established a new pricing tier for models that 'think longer.' This reasoning premium reflects the real compute cost of multi-step inference and creates a clear price/capability ladder.
Common Questions
Which AI model is cheapest per token in 2026?
Google's Gemini 2.5 Flash and Meta's Llama 4 Scout tie at $0.15 and $0.10 per million input tokens respectively. Mistral Small is even cheaper at $0.10/$0.30. For the cheapest option with strong quality, Claude Haiku 4.5 at $1/$5 is the best value-to-quality ratio among proprietary models.
How do I choose between models?
Match cost to task complexity. Use Flash/Mini/Small tier ($0.10-$1/M) for classification, extraction, and simple generation. Use mid-tier ($2-$5/M) for coding, analysis, and complex writing. Reserve premium models ($5-$40/M) for tasks where quality directly impacts business outcomes — legal review, code architecture, research synthesis.
Are open-source models really free?
Open-weight models like Llama 4 have no license fee, but you pay for compute. Self-hosting on GPUs costs $1-3/hour per GPU. For most teams, inference providers (Together AI, Fireworks, Groq) offer Llama 4 at $0.10-$0.60/M tokens — cheaper than hosting unless you have sustained high volume (>$5K/month in API costs).
How often do AI model prices change?
Prices have dropped roughly 50-70% per year since 2023. New model releases typically come with lower per-token prices than their predecessors. The trend is clear: within 12-18 months of a model's release, a cheaper alternative with comparable quality appears.