← Voltar aos guias

Cost per 1M Tokens: A Practical 2026 Provider Guide

Compare the workload, not the headline price

Comparing cost per 1M tokens in 2026 starts with two separate numbers: input and output. A low input rate does not automatically mean a low bill. For example, Google's published Gemini 2.5 Flash rates distinguish standard input, output, cached input and cache storage. Treating these as one interchangeable price produces misleading estimates. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing))

Use this guide to compare both model providers and access routes. First compare different models against your quality requirements; then compare the same model through a direct API and a relay. Those are different purchasing decisions.

2026 pricing snapshot: direct APIs versus 59API

These published rates were checked on October 25, 2026. Amounts below are per 1M tokens, using standard uncached text input and output rather than batch or promotional credits.

• Anthropic direct — Claude Haiku 4.5: $1.00 input and $5.00 output. ([platform.claude.com](https://platform.claude.com/docs/en/models/haiku-4-5/overview))
• Google direct — Gemini 2.5 Flash: $0.30 input and $2.50 output, including thinking tokens in the output charge. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing))
• 59API — Claude Haiku 4.5: approximately $0.3472 input and $1.67 output. ([59api.com](https://59api.com/pricing))
• 59API — Gemini 2.5 Flash: approximately $0.2083 input and $1.67 output. ([59api.com](https://59api.com/pricing))

59API displays approximate USD equivalents but bills in CNY, so account for currency conversion when reconciling spending. These examples demonstrate lower published relay rates, not a universal cheapest-provider ranking. ([59api.com](https://59api.com/pricing))

Calculate your actual blended token cost

For an uncached workload, multiply input tokens in millions by the input rate, multiply output tokens in millions by the output rate, and add the results. Keep the two categories separate rather than averaging their prices. Both direct providers publish separate input and output charges. ([platform.claude.com](https://platform.claude.com/docs/en/models/haiku-4-5/overview))

Consider a hypothetical monthly workload with 8M input tokens and 2M output tokens. At the direct rates above, it costs $18.00 with Claude Haiku 4.5 or $7.40 with Gemini 2.5 Flash. These are calculated examples, not predictions of identical token consumption or quality. ([platform.claude.com](https://platform.claude.com/docs/en/models/haiku-4-5/overview))

Using 59API's displayed rates, the same token counts calculate to approximately $6.12 and $5.01, respectively, before additional charges. Divide each total by ten to obtain its blended cost per 1M total tokens. ([59api.com](https://59api.com/pricing))

Check caching, batch processing and extra charges

• Separate cache reads from writes. Anthropic lists distinct cache-write and cache-read rates. Estimate how often a reusable prompt will actually be read before assuming caching saves money. ([platform.claude.com](https://platform.claude.com/docs/en/models/haiku-4-5/overview))
• Include cache storage. Google's pricing lists storage charges alongside discounted cached-input rates. A cache-hit price alone is not the complete caching bill. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing))
• Evaluate batch pricing separately. Anthropic offers a 50% batch discount on input and output; Google's listed batch rates for the example model are also half its standard rates. Check whether your workflow can tolerate asynchronous processing. ([platform.claude.com](https://platform.claude.com/docs/en/models/haiku-4-5/overview))
• Budget for tools. Google's pricing separates search-grounding charges from token charges. Add any applicable tool expenses before comparing final totals. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing))

Benchmark before moving production traffic

Build a test set of 50–100 representative tasks. Keep prompts, output limits and acceptance criteria consistent. Record billed spending, completion time, retries and whether each answer passes your checks. Compare total spending divided by accepted results—not merely spending divided by requests.

For relay testing, verify the requested model, supported protocol, streaming and tool behavior. Start with a small prepaid balance and reconcile several requests against usage logs before increasing traffic.

Where 59API fits

59API is a compelling low-cost shortlist option: it combines pay-as-you-go access with Claude Code, Codex and OpenAI-compatible client support. It describes its service as preserving official-model capability; validate that promise against your own acceptance tests rather than assuming price proves quality. ([59api.com](https://59api.com/))

Its referral program also advertises a 15% rebate on invitees' top-ups. Treat referral earnings separately from baseline API savings. If the published rates fit your workload, sign up for 59API and run a small comparison before committing your production budget. ([59api.com](https://59api.com/pricing))

Preços · Docs