Tools: LLM Pricing in February 2026: What Every Model Actually Costs

The full pricing table ## What stands out ## Beyond the price tag ## Who should use what ## What's next Originally published on Kael Research TL;DR: Cheapest option is OpenAI's open-source GPT-OSS-20B at $0.05/M input. Best value is GPT-5 mini at $0.25/M. Most expensive is Grok-4 at $30/M — 600x more than GPT-OSS-20B. Claude Opus 4.6 dropped to $5/$25 (down from $15/$75 on Opus 4). Full table with 18 models below. If you're building on top of LLMs right now, you're probably spending more than you need to. Pricing has changed so fast over the past year that most teams are running on outdated assumptions. Here's what every major model actually costs as of February 2026, with the context that matters for choosing between them. All prices are per million tokens. Sources: OpenAI pricing, Anthropic models, Google AI pricing, xAI pricing, DeepSeek pricing, Together.ai, Groq for open-source model hosting. All checked February 19, 2026. The gap between cheapest and most expensive is staggering. GPT-OSS-20B at $0.05/M input vs Grok-4 at $30/M input. That's 600x. Even comparing production-grade models, GPT-5 mini at $0.25/M vs Claude Opus 4.6 at $5/M is a 20x spread. For most workloads, the cheaper models handle 80%+ of tasks just fine. xAI is pricing itself out. Grok-4 at $30/$150 per million tokens is the most expensive API on the market. That's 6x Claude Opus 4.6 and 17x GPT-5.2 on input. Unless you need something Grok does better (hard to name what that is), the pricing makes no sense for production use. Google is quietly the cheapest. Gemini 2.0 Flash at $0.10/$0.40 matches GPT-4.1 nano and undercuts almost everything else. If your use case tolerates the quality tradeoff, it's the best deal available. Open-weight models changed the math. Llama 4 Maverick at $0.27/$0.85 through hosted APIs is cheap, but the real story is self-hosting. Running Llama on your own GPUs drops the effective cost below $0.10/M tokens for input. The breakeven vs API depends on volume, but for companies doing 10B+ tokens/month, self-hosting wins. The table is just the start. What actually matters: Output tokens cost 3-8x more than input. This is consistent across every provider. If your app generates long responses (code, reports, content), output cost dominates your bill. Trim your outputs. Caching changes everything. OpenAI and Anthropic both offer prompt caching that cuts repeat-context costs by 50-90%. If you're sending the same system prompt or few-shot examples on every call, caching alone might cut your bill in half. Quality gaps are shrinking. A year ago, there was a clear hierarchy: GPT-4 > Claude 3 > everything else. Now GPT-5 mini, Claude Sonnet 4, and Gemini 2.5 Flash are all competitive for most tasks. The premium models (GPT-5.2, Opus 4) still win on complex reasoning and long-form analysis, but the gap keeps closing. Latency matters more than price. The cheapest model that takes 8 seconds to respond might cost you more in user drop-off than a 2x pricier model that responds in 1.5 seconds. Benchmark latency alongside cost. High-volume production (chatbots, classification, extraction): GPT-5 mini or Gemini 2.0 Flash. Both under $0.50/M input with solid quality. Code generation: Claude Sonnet 4 or GPT-5.2. Sonnet is generally better at following complex coding instructions. GPT-5.2 has an edge on multi-file refactoring. Research and analysis: Claude Opus 4.6 if budget allows ($5/$25 is much more reasonable than the old Opus 4 pricing). GPT-5.2 if not. Cost-sensitive startups: Llama 4 Maverick self-hosted, or GPT-4.1 nano for API. Get to market first, pick the right model later. Pricing has dropped roughly 10x per year for equivalent quality over the past three years. There's no reason to think that stops. By Q4 2026, expect GPT-5 mini-equivalent quality at $0.05/M input or less. The real shift is happening at the infrastructure layer. Custom silicon (Google TPUs, Amazon Trainium, Microsoft Maia) is starting to undercut Nvidia GPU economics. As that scales, hosted API pricing will drop faster than self-hosting costs — potentially flipping the build-vs-buy calculation for mid-size companies. We'll update this comparison monthly. Subscribe to get updates when pricing changes. This analysis is part of Kael Research's ongoing coverage of AI market economics. We track pricing, adoption, and competition across the AI industry. See our full research briefs for deeper analysis on specific markets. Templates let you quickly answer FAQs or store snippets for re-use. Are you sure you want to hide this comment? It will become hidden in your post, but will still be visible via the comment's permalink. Hide child comments as well For further actions, you may consider blocking this person and/or reporting abuse