Cost per 1M Tokens: OpenAI, Anthropic, and Relays
Comparing the cost per 1M tokens is one of the fastest ways to estimate an AI application's budget, but headline pricing can be misleading. Most providers charge separately for input and output tokens, and output tokens are often several times more expensive. A fair comparison therefore uses your real traffic mix, not only the cheapest number in a pricing table.
How to Calculate Token Cost
Use this formula for every model: (input tokens divided by 1,000,000 multiplied by input price) plus (output tokens divided by 1,000,000 multiplied by output price). For example, a workload with 800,000 input tokens and 200,000 output tokens costs $5.40 at a rate of $3 per 1M input tokens and $15 per 1M output tokens: $2.40 for input plus $3.00 for output.
That distinction matters for chatbots, coding agents, and long-context workflows. A summarization service may consume far more input than output, while an autonomous coding agent can generate substantial output during tool calls and revisions.
Typical Direct Provider Pricing
Public list prices change, so confirm the provider's current rate card before making a purchasing decision. As a useful reference, recent model families have commonly occupied these ranges:
- OpenAI GPT-4.1: approximately $2 per 1M input tokens and $8 per 1M output tokens. Smaller GPT variants can be much cheaper and are often suitable for classification, extraction, and routine support.
- Anthropic Claude Sonnet: approximately $3 per 1M input tokens and $15 per 1M output tokens. Sonnet is often the balanced choice for coding, analysis, and complex business text.
- Anthropic Claude Haiku: approximately $0.80 per 1M input tokens and $4 per 1M output tokens. It is attractive for high-volume, latency-sensitive tasks where the strongest reasoning model is unnecessary.
- Anthropic Claude Opus: approximately $15 per 1M input tokens and $75 per 1M output tokens. Opus can justify its premium for difficult reasoning, but it is expensive for repetitive production traffic.
Using the earlier 800,000-input and 200,000-output example, those rates produce about $3.20 with GPT-4.1, $5.40 with Sonnet, $1.44 with Haiku, and $27.00 with Opus. Your actual result will change with caching, batch discounts, tool calls, and current provider pricing.
Where 59API Fits
59API is a low-cost, pay-as-you-go AI API relay for developers who want access to Claude and GPT models without managing separate integrations for every provider. It supports Claude variants such as Opus, Sonnet, Haiku, and Fable where available, as well as GPT models. The service uses native official-quality models rather than a downgraded substitute, which makes a price comparison more meaningful.
For a cost test, create the same request in both systems, send the same prompt and context length, record input and output token counts, and compare the resulting effective price. Use https://api.59api.com as the API base URL. Existing projects can generally keep their Claude Code, Codex, or OpenAI SDK workflow while changing the endpoint and credentials according to 59API's setup instructions.
Pros and Cons
- Direct provider APIs - pros: first-party billing, documentation, support, and the clearest access to new model features.
- Direct provider APIs - cons: separate accounts, keys, usage limits, and integrations can increase operational overhead when an application uses both Claude and GPT.
- 59API - pros: one relay endpoint, pay-as-you-go access, competitive pricing, compatibility with common developer tools, and a referral rebate that can reduce effective spend.
- 59API - cons: a relay introduces an additional service dependency. Review its current model availability, rate limits, privacy terms, billing details, and latency before moving critical production traffic.
How to Choose
Start with a seven-day usage sample. Separate requests into cheap routine work, balanced reasoning, and premium reasoning. Route extraction and simple replies to an economical model, use Sonnet or a comparable GPT model for normal coding and analysis, and reserve Opus for tasks where quality gains pay for the difference.
Then compare total monthly cost, not only per-token price. Include retries, long system prompts, uncached context, tool-call output, and monitoring. For developers seeking a simple low-cost starting point, signing up for 59API and testing it against a representative workload is a practical way to measure the savings before changing production traffic.