How AI API prices have changed over time — tracking price cuts, generational improvements, and the overall downward trajectory across all major providers.
| Model | Provider | Date | Old Input $/1M | New Input $/1M | Input Change | Note |
|---|---|---|---|---|---|---|
| GPT-4o (Omni) | openai | Aug 6, 2024 | $5 | $2.5 | -50% | Prices cut by 50% on input and 33% on output. |
The GPT-5.6 family (Sol, Terra, Luna), GPT-5.5/5.4 generations, and unified reasoning systems: 1M+ context on frontier tiers, 90% prompt caching discounts, and 50% batch discounts.
Claude 5 generation (Opus 5, Sonnet 5, Fable 5, Haiku 4.5) alongside Claude 3.7 Sonnet hybrid reasoning with industry-leading context windows and 90% prompt caching savings.
Gemini 3.x, 2.5, 2.0, and 1.5 series: 2M token context, ultra-fast Flash tiers, context caching discounts up to 90%, and multimodal native reasoning.
DeepSeek V4 Pro, V4 Flash, R1 Reasoning, and V3 Chat: 1M token context, disruptive low pricing, peak/off-peak tier discounts, and 90% cache-hit savings.
Llama 4 Maverick (400B MoE), Llama 4 Scout (109B MoE with 10M context), Llama 3.3 70B, Llama 3.1 405B/70B/8B, and Llama 3.2 Vision: open-weight foundation models.
European open-weight leader: Mistral Large 3, Medium 3.5, Small 4, Codestral 2501 code engine, Pixtral multimodal, and Ministral edge models.
Grok 4.6 Flagship, Grok 4.5, Grok 4.3, and Grok Build 0.1 models served from Colossus clusters with up to 1M-token context and cheap output token rates.
Qwen 3.8-Max, Qwen 3 32B, QwQ-32B reasoning models, Qwen 2.5 Max/Plus/Turbo, and specialized Qwen 2.5 Coder 32B models offering top-tier price-performance.
Enterprise RAG and agentic powerhouses: Command A+ (218B MoE unified), Command R+, Command R, Embed v3.0, and high-performance Re-ranking models.
Amazon Nova Premier, Nova 2 Pro, Nova 2 Lite, Nova Pro, Nova Lite, and Nova Micro: cost-efficient foundation models natively integrated into AWS Bedrock.
Sonar Reasoning Pro, Sonar Deep Research, Sonar Pro, and Sonar online models with real-time web-grounded search synthesis.
Ultra-low latency inference on custom Language Processing Units (LPUs): 300–800+ tokens/sec on Llama 4, Llama 3.3, QwQ, and Mistral models.
Extreme wafer-scale engine inference: up to 3,000 tokens/second on GPT-OSS-120B, Gemma 4, Qwen 3, and GLM 4.7 with generous daily free developer tiers.
High-performance GPU cloud delivering serverless and dedicated endpoints for Llama 4, DeepSeek V4/R1, Qwen 2.5, and open-weight models with fine-tuning.
Speed-optimized FireOptimizer serverless inference platform serving Llama 4 Maverick/Scout, DeepSeek V4/R1, and custom fine-tuned LoRA models.
Yi-Lightning and Yi-Large models: highly optimized bilingual Chinese/English models with fast throughput.
Kimi K3 and Kimi K2.6: 1M-token-context pioneer from Beijing with fine-grained cache discounts and deep document reasoning.
GLM 5.2 and GLM 4.7: top foundation models with bilingual English/Chinese balance, multimodal capabilities, and coding mastery.
Jamba 1.5 Large and Mini: hybrid SSM-Transformer (Mamba) architecture providing ultra-efficient 256K long-context throughput.
MiniMax-01 and Text-01: high-performance 4M-token context window models designed for massive document synthesis and multilingual dialog.
Cost-effective serverless inference provider delivering open models (Llama 3.3, DeepSeek R1, Qwen 2.5) with pure per-token pay-as-you-go pricing.
Universal AI gateway routing queries across 200+ models with smart fallback, auto-best routing, and unified billing across providers.
Phi-4 and Phi-3.5 family of highly capable small language models (SLMs) delivering state-of-the-art synthetic reasoning at compact footprints.
DBRX Instruct and Base: enterprise-grade 132B open Mixture-of-Experts (MoE) foundation models built for Lakehouse AI pipelines.
Nemotron-4 340B and Llama-3.1-Nemotron-70B: synthetic data generation, reward modeling, and enterprise alignment engines optimized with TensorRT-LLM.
Since 2023, the price per million tokens has dropped 90%+ for comparable quality models. DeepSeek R1, Gemini Flash, and budget OpenAI tiers have all driven incumbent price cuts. The trend shows no sign of stopping.
The model you chose a year ago may now have a faster, cheaper, better-quality alternative. Set a quarterly reminder to benchmark your top 3 use cases against new models at current prices.
AI API pricing has fallen dramatically over the past two years — roughly 40–70% per year on average for frontier models. New model generations consistently offer better capability-per-dollar, and competitive pressure between OpenAI, Anthropic, Google, and new entrants like DeepSeek accelerates price cuts.
Yes — we recommend quarterly reviews. A model that was the right cost-quality tradeoff 6 months ago may now have cheaper, better alternatives. Price drops don't usually require you to change your code — just the model ID in your API calls.
Competition between providers is the primary driver. When one provider cuts prices, others often follow. Additionally, providers improve their inference infrastructure over time — lower serving costs get passed through as lower API prices. New quantization techniques and hardware efficiency gains also contribute.
Not historically. Price drops in AI APIs have generally come alongside quality improvements, not tradeoffs. New model generations at the same or lower price points consistently outperform predecessors on standard benchmarks.