AI API pricing has collapsed. Models that cost $60 per million tokens two years ago now cost $0.40. The same capability that required enterprise contracts in 2024 is available on a free tier in 2026. But the pricing landscape has also become more complex—batch APIs, prompt caching, context window tiers, and reasoning model premiums make direct comparisons surprisingly difficult.
Here’s the actual pricing from OpenAI, Anthropic, and Google as of January 2026, plus the open-source alternatives that are increasingly competitive.
Quick reference: Cost per million tokens
These are the models you’ll actually use for production workloads:
Budget tier ($0.10-0.60): GPT-4o-mini ($0.15/$0.60), Gemini 2.0 Flash ($0.10/$0.40), DeepSeek V3.2 ($0.28/$0.42), Haiku 3 ($0.25/$1.25)
Mid tier ($1-5): GPT-5 ($1.25/$10), Gemini 2.5 Pro ($1.25/$10), Sonnet 4.5 ($3/$15), Haiku 4.5 ($1/$5)
Premium tier ($5-25): Opus 4.5 ($5/$25), o3 reasoning ($2/$8), Gemini 3 Pro ($2/$12)
The pattern: Output tokens cost 3-5x more than input tokens. This matters because your actual costs depend heavily on your use case—chatbots generate lots of output; RAG systems use lots of input.
OpenAI pricing breakdown
OpenAI’s flagship GPT-5 runs $1.25 input / $10 output per million tokens, with a 400K context window. That’s actually cheaper than GPT-4o ($2.50/$10) despite being significantly more capable. GPT-5 nano offers a budget option at $0.05/$0.40.
The best value play is still GPT-4o-mini at $0.15/$0.60—it’s 93% cheaper than GPT-4 was and handles most production use cases competently.
The o-series reasoning models saw an 80% price drop. o3 now costs $2/$8, down from o1’s $15/$60. That makes reasoning models actually viable for production workloads instead of expensive experiments.
Batch API: 50% off for asynchronous processing. If you can tolerate 24-hour turnaround, GPT-5 drops to $0.625/$5.
Anthropic (Claude) pricing breakdown
Anthropic’s latest Claude 4.5 family prices:
Opus 4.5: $5/$25 per million tokens. This is 67% cheaper than Opus 4/4.1 ($15/$75) while being the most capable model in the family. Extended thinking—Claude’s “reasoning mode”—bills thinking tokens as output tokens, not at a premium rate.
Sonnet 4.5: $3/$15. The workhorse model for most production use cases. Long context (over 200K tokens) costs $6/$22.50.
Haiku 4.5: $1/$5. Fast and cheap, but noticeably less capable than the budget offerings from OpenAI or Google.
Prompt caching: Cache reads cost 90% less than regular input. If you’re running similar prompts repeatedly, this adds up fast. Cache writes cost 1.25-2x base price depending on TTL.
Google (Gemini) pricing breakdown
Google’s pricing is the most aggressive in the market, especially at the budget end:
Gemini 2.0 Flash: $0.10/$0.40. This is the cheapest viable model from a major provider. Ultra-budget option Gemini 2.0 Flash-Lite runs even cheaper at ~$0.075/$0.30.
Gemini 2.5 Pro: $1.25/$10 (doubles for context over 200K tokens). Production-stable and competitive with GPT-5 on most benchmarks.
Gemini 3 Pro Preview: $2/$12. Google’s latest flagship, with long context pricing at $4/$18 over 200K tokens.
Free tier: Google AI Studio offers up to 1,000 daily requests with no credit card required. That’s enough for serious prototyping.
The open-source alternative
DeepSeek deserves special attention. The V3.2-Exp model costs $0.28/$0.42 per million tokens—27x cheaper than OpenAI o1 and 95% cheaper than GPT-5. Cache hits drop input costs to $0.028. Off-peak discounts (50-75% during certain hours) push it even lower.
DeepSeek R1, their reasoning model, runs $0.55/$2.19—still dramatically cheaper than o3’s $2/$8.
The quality gap between DeepSeek and the major providers has narrowed significantly. For many use cases, it’s now “good enough” at a fraction of the cost.
Other options: Together AI hosts 200+ models including Llama 3.3 70B at ~$0.88 per million. Groq offers ultra-fast inference with Llama 70B at ~$0.60 per million. Fireworks AI runs fine-tuning at $0.50 per million tokens for models up to 16B parameters.
Cost scenarios: What you’ll actually pay
Let’s run real numbers. Assuming 40% input / 60% output token distribution:
Startup (100M tokens/month):
Gemini 2.0 Flash: $28. DeepSeek V3.2: $36. GPT-4o-mini: $42. Anthropic Haiku 4.5: $340. Anthropic Sonnet 4.5: $1,020.
The difference between choosing Gemini Flash and Claude Sonnet is 36x in cost. Whether that matters depends entirely on whether the quality difference matters for your use case.
Scale-up (1B tokens/month):
Gemini 2.0 Flash: $280. DeepSeek V3.2: $364. GPT-4o-mini: $420. With batch processing (50% discount), these drop further—OpenAI GPT-4o-mini batch runs $210/month at this volume.
Enterprise (10B tokens/month):
At this scale, self-hosting becomes viable. Running Llama 70B on 8x H100s costs roughly $15,000-20,000/month. GPT-4o-mini API would cost $4,200, Gemini 2.0 Flash $2,800. The API wins unless you need specific customization or compliance requirements.
Optimization strategies
Prompt caching: Anthropic offers 90% savings on cache reads. Google offers up to 75%. If you’re running similar system prompts repeatedly (chatbots, RAG systems), this can cut costs dramatically.
Batch processing: All major providers offer 40-50% discounts for asynchronous batch jobs. If you don’t need real-time responses—data processing, analysis, content generation pipelines—batch is free money.
Model tiering: Use cheap models for simple tasks, expensive models for hard tasks. A common pattern: GPT-4o-mini for classification and routing, Sonnet or o3 for complex reasoning that actually needs it.
Output optimization: Output tokens cost 3-5x more than input. Asking for shorter responses, structured outputs, or JSON instead of prose can meaningfully reduce costs.
Bottom line
The 2024 assumption that “AI API costs are prohibitive” is no longer true. GPT-4o-mini at $0.15/$0.60 handles most use cases. Gemini 2.0 Flash at $0.10/$0.40 handles even more at lower cost. DeepSeek V3.2 at $0.28/$0.42 offers near-frontier quality at budget prices.
The decision framework is simpler than the pricing tables suggest:
Cost-sensitive: Gemini 2.0 Flash or DeepSeek V3.2.
Balanced: GPT-4o-mini for OpenAI ecosystem, Gemini 2.5 Pro for Google Cloud.
Quality-first: Claude Sonnet 4.5 for writing, o3 for reasoning.
Premium: Claude Opus 4.5 when you need the best available model.
Prices are still dropping. What costs $100/month today will probably cost $30 next year. Plan accordingly.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
