Choosing an LLM isn't just about capability — cost per token can vary by 100x or more across providers. Below we break down the input and output pricing for the top 10 large language models in 2026, ranked from cheapest to most expensive, so you can budget accurately. See our top 10 LLM models ranked by capability with use cases for each.
The Cost Rankings — Cheapest First
| # | Model | Provider | Input / 1M tok | Output / 1M tok | Context |
|---|---|---|---|---|---|
| 1 | Gemma 4 | FREE | FREE | 256K | |
| 2 | DeepSeek V4 Pro | DeepSeek | $0.27 | $1.10 | 1.6T MoE |
| 3 | GLM 5.1 | Z.ai | $0.07 | $0.28 | 200K |
| 4 | Kimi K2.6 | Moonshot AI | $0.08 | $0.32 | 256K |
| 5 | Gemini 3.5 Flash | Google DeepMind | $0.07 | $0.30 | 1M |
| 6 | Gemini 3.1 Pro | Google DeepMind | $1.25 | $5.00 | 1M |
| 7 | GPT-5.4 | OpenAI | $2.50 | $15.00 | 1M |
| 8 | Claude Sonnet 4.6 | Anthropic | $3.00 | $15.00 | 1M |
| 9 | GPT-5.5 | OpenAI | $5.00 | $30.00 | 1M |
| 10 | Claude Opus 4.8 | Anthropic | $15.00 | $75.00 | 200K |
Prices are approximate per 1M tokens in USD as of June 2026. Actual costs vary by usage tier and commitment level. Some models offer batch or reserved-throughput discounts not reflected here. Gemma 4 is free (open model, self-hosted).
Per-Model Cost Breakdown
Gemma 4 — Google
Completely freeGemma 4 is the only model on this list that costs $0 in API fees — because it's an open model you run yourself. At just 14B parameters in a 14GB package, it runs at 85 tokens/sec on consumer hardware (laptops, phones). For on-device AI, edge computing, and privacy-first local inference, the total cost of ownership is unbeatable.
DeepSeek V4 Pro — DeepSeek
Best cost-to-performanceA 1.6T-parameter Mixture-of-Experts model (49B active) that delivers frontier reasoning, coding, and agentic performance at a fraction of Western API costs. At just $0.27/1M input and $1.10/1M output, it's the leading choice for organizations that need maximum capability on a tight budget.
GLM 5.1 — Z.ai
The leading open-weights model on the Intelligence Index, with a 744B MoE architecture (40B active). At $0.07/1M input and $0.28/1M output, it's the top open-source alternative to proprietary frontier models for self-hosted deployments, academic research, and domain-specific fine-tuning.
Kimi K2.6 — Moonshot AI
A ~1T-parameter MoE model (32B active) purpose-built for agent-oriented coding and long-context understanding. At $0.08/1M input and $0.32/1M output, it delivers exceptional value for analyzing large codebases, lengthy documents, and multimedia content in a single pass — with native video support.
Gemini 3.5 Flash — Google DeepMind
Outperforms Gemini 3.1 Pro on coding and agentic benchmarks while running at lower cost. At $0.07/1M input and $0.30/1M output with 1M-token context and class-leading speed, it's the top choice for real-time coding assistants, low-latency chatbots, and high-throughput production pipelines.
Gemini 3.1 Pro — Google DeepMind
Leads 13 of 16 major benchmarks and scores 94.3% on GPQA Diamond. At $1.25/1M input and $5.00/1M output, it's the most cost-effective frontier model — delivering top-tier reasoning, 1M-token context, and Google Search grounding at roughly half the price of comparable proprietary models.
GPT-5.4 — OpenAI
OpenAI's cost-quality sweet spot at $2.50/1M input and $15.00/1M output — half the price of GPT-5.5 with most of the capability. With a 1M-token context window, it's ideal for high-volume content production, cost-sensitive enterprise apps, and prototyping at scale.
Claude Sonnet 4.6 — Anthropic
Anthropic's mid-tier model that now delivers Opus-level intelligence at Sonnet pricing. At $3.00/1M input and $15.00/1M output with a 1M-token context window, it's the best value among proprietary frontier models for production workloads — strong on coding, analysis, and creative tasks.
GPT-5.5 — OpenAI
The most capable general-purpose model in June 2026. At $5.00/1M input and $30.00/1M output with a 1M-token context window and thinking mode, it's the premium choice for complex reasoning, full-stack software development, agentic workflows, and creative tasks where output quality directly impacts revenue.
Claude Opus 4.8 — Anthropic
Most expensiveAnthropic's most reliable model — 4x more dependable than 4.7 on code verification tasks. At $15.00/1M input and $75.00/1M output, it's the most expensive model on this list — best reserved for production code generation, safety-critical enterprise apps, and high-value tasks where output quality is paramount. A "Fast" variant is available at $10/$50 per 1M tokens.
Cost Spread at a Glance
The price gap between the cheapest and most expensive LLM is staggering — Claude Opus 4.8 costs 68× more per output token than DeepSeek V4 Pro, and 250× more than self-hosted Gemma 4. Here's how the full field compares visually (output cost per 1M tokens):
How to Think About LLM Costs
Cost per token alone can be misleading. Here's what actually matters when budgeting:
- Output tokens dominate bills. Output pricing is typically 3–5× higher than input pricing, and most real-world workloads generate more output tokens than input — especially chat and summarization.
- Self-hosting changes the equation. Models like Gemma 4, GLM 5.1, and DeepSeek V4 Pro can run on your own GPUs, eliminating per-token API fees entirely. For high-volume use, the breakeven on hardware often hits within months.
- Cheapest ≠ best value. DeepSeek V4 Pro is 68× cheaper than Claude Opus 4.8 per output token — but if Claude's reasoning quality saves your team hours of debugging or generates higher-converting copy, the cost premium pays for itself.
- Model routing cuts bills by 40–60%. The most cost-efficient AI deployments route simple queries (FAQ, classification) to cheap models like Gemini 3.5 Flash or Kimi K2.6, and complex queries (analysis, coding, writing) to premium models like GPT-5.5 or Claude Opus 4.8.