$1 buys 704K Gemini 2.5 Flash tokens with this relay service

01 / Gemini 2.5 Flash starts at $0.1778 per million input tokens

CheapRelay is a relay service built around a simple question: what does the work actually cost? The price list gives you a direct answer, with no monthly fee and no subscription attached to the usage.

For Gemini 2.5 Flash, the current prices per 1,000,000 tokens are $0.1778 for input, $1.42 for output, and $0.0181 for cache reads. The model id to enter is gemini-2.5-flash.

That gives us three useful numbers instead of a vague promise. Input is the material you send to the model. Output is the material it sends back. Cache reads let a repeated context cost less on later calls. CheapRelay charges per real token, so those three lines are the bill you can calculate before you make the request.

Here is the first-day arithmetic. In this example, one request contains 1,000,000 input tokens, generates 100,000 output tokens, and performs 1,000,000 cache reads. The cost is:

• ■ Input: 1,000,000 ÷ 1,000,000 × $0.1778 = $0.1778
• ■ Output: 100,000 ÷ 1,000,000 × $1.42 = $0.1420
• ■ Cache read: 1,000,000 ÷ 1,000,000 × $0.0181 = $0.0181
• ■ Total: $0.1778 + $0.1420 + $0.0181 = $0.3379

Ten of those workloads cost $3.379. There is no need to guess whether the meter is moving quickly. The price list supplies the units, and the bill follows.

02 / What a fixed dollar buys on Gemini 2.5 Flash

Suppose your budget is exactly $1. In this example, spent entirely on output, that buys 1,000,000 ÷ $1.42, or about 704,225 output tokens. Spend it entirely on input instead, and the same $1 buys about 5,624,297 input tokens. Spend it entirely on cache reads, and it buys about 55,248,619 cached tokens.

Those figures are not interchangeable. Output is the expensive part because the model has to generate it. Input is cheaper, and cache reads are cheaper again. The practical move is to compare your workload with these three lines rather than treating a token as one uniform unit.

Consider a mixed example with a $10 balance. Suppose you send 5,000,000 input tokens, receive 500,000 output tokens, and make 5,000,000 cache reads:

• ■ Input: 5 × $0.1778 = $0.8890
• ■ Output: 0.5 × $1.42 = $0.7100
• ■ Cache reads: 5 × $0.0181 = $0.0905
• ■ Total: $0.8890 + $0.7100 + $0.0905 = $1.6895

That leaves $8.3105 in the balance. CheapRelay uses a prepaid balance, so the arithmetic is visible before and after the call. There is no monthly fee, and the balance does not expire.

Every new account also receives $0.09 in sign-up credit, described on the site as 注册赠送额度. It is usable on any model. In this example, using all of it on Gemini 2.5 Flash output buys about 63,380 output tokens: $0.09 ÷ $1.42 × 1,000,000. It is a small test budget, but a test budget with a visible ceiling is more useful than an unlimited promise.

03 / Use Gemini CLI and the native protocol

The first rule of a cheap first day is to keep the client setup compatible with the model. Gemini models on CheapRelay use the native protocol under /v1beta, through Gemini CLI. They are not served through an OpenAI-compatible endpoint.

That means the setup is specific:

• ■ Use Gemini CLI.
• ■ Use the native protocol under /v1beta.
• ■ Enter the model id gemini-2.5-flash.
• ■ Do not try to route Gemini 2.5 Flash through an OpenAI-compatible client configuration.

CheapRelay separates model families deliberately. Gemini models follow the native Gemini route. GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and the other compatible models use the OpenAI-compatible route under /v1, where clients such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK and cURL fit the documented setup.

Claude is a separate case again: Claude models work only with the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. Keeping these paths straight prevents a surprising mismatch on day one.

The operational limits are also plain: 120 calls per minute per key, and 8 calls at once per account. If you receive HTTP 402, the balance is at or below zero. Top up the prepaid balance, then continue. There is no mystery fee hiding behind the status code.

04 / Prepaid money, refunds and referral arithmetic

CheapRelay bills from a prepaid balance, so you decide how much money enters the account and how much work it buys. The balance does not expire, and there is no monthly fee. If a planned experiment ends early, the unused balance remains available for later token usage.

The refund rule is specific too. Unused balance can be refunded within 12 days of the top-up, minus what was spent and minus the payment fee. If you top up $20 and spend $4.25, this example leaves $15.75 of unused balance, with the payment fee deducted from the refundable amount.

Referrals are another visible calculation. A directly invited user pays you 20% of what they pay. From the second level, you receive 6%. Those amounts go into a withdrawable rewards balance, and the smallest withdrawal is ¥20, via Alipay.

For example, if a directly invited user pays $10, your rewards balance receives $2. If someone in your second level pays $10, it receives $0.60. Across a $10 direct payment and a $10 second-level payment, this example adds $2.60. Once the withdrawable rewards balance reaches ¥20, you can withdraw it through Alipay.

That structure is easy to audit because every number comes from the same formula. You can check the invited payments, multiply by 20% or 6%, and compare the result with the rewards balance.

05 / Compare Gemini 2.5 Flash with Gemini 3.8 Flash before switching

CheapRelay also lists Gemini 3.8 Flash, with a model id of gemini-3.8-flash. Its price is $0.95 per 1,000,000 input tokens, $4.28 per 1,000,000 output tokens, and $0.0944 per 1,000,000 cache reads. Gemini 2.5 Flash is lower on all three lines: $0.1778 versus $0.95 for input, $1.42 versus $4.28 for output, and $0.0181 versus $0.0944 for cache reads.

Put both models into the same example workload: 1,000,000 input tokens, 100,000 output tokens, and 1,000,000 cache reads.

• ■ Gemini 2.5 Flash: $0.1778 + $0.1420 + $0.0181 = $0.3379
• ■ Gemini 3.8 Flash: $0.9500 + $0.4280 + $0.0944 = $1.4724
• ■ Difference: $1.4724 − $0.3379 = $1.1345

In this example, Gemini 3.8 Flash costs about 4.36 times as much as Gemini 2.5 Flash. If the workload is the same, that gap belongs in your decision before the tokens are spent.

The price list also includes GPT-6.1 Sol at $0.1486 for input and $0.7375 for output per 1,000,000 tokens. Its official list price is $2.00 for input and $10.00 for output, with the published list-price conversion at 7.2 yuan per US dollar. The output price shown by CheapRelay is 93% under that official list price. Claude Haiku 4.5 is listed at $1.40 for output against an official list price of $5.00, a 72% gap. Those published comparisons belong to their own rows on the price list.

For Gemini 2.5 Flash, the first-day case is straightforward: $0.1778 input, $1.42 output, $0.0181 cache reads, Gemini CLI, native /v1beta, and a prepaid balance with no monthly fee. CheapRelay makes the cost legible before you commit the money, which is the point: know the unit price, multiply the tokens, and keep the difference.