Gemini CLI pricing: does a cache read really cost $0.0181?

01 / Why cache reads deserve a second look

Most developers compare input and output prices first. That is reasonable, but it hides the line that quietly grows when an agent keeps re-reading the same files, tool definitions, and project context. A cache read is the repeated portion of the prompt. If your workflow sends that material again and again, the cache-read rate is the number that often shrinks your bill the most.

CheapRelay publishes a plain price list with separate input, output, and cache-read prices per 1,000,000 tokens. There is no monthly fee. You top up a prepaid balance, pay per real token, and the balance does not expire. Every new account also receives $0.09 in signup credit, usable on any listed model.

The examples below use 10,000,000 cache-read tokens on one model. The arithmetic is deliberately plain: 10 multiplied by the price per 1,000,000 tokens.

02 / Gemini CLI pricing: Gemini 2.5 Flash versus Gemini 3.8 Flash

Gemini 2.5 Flash is priced at $0.0181 per 1,000,000 cache-read tokens on CheapRelay. At 10,000,000 cache-read tokens, the example cost is:

10 × $0.0181 = $0.181

Gemini 3.8 Flash is priced at $0.0944 per 1,000,000 cache-read tokens. For the same 10,000,000-token example:

10 × $0.0944 = $0.944

Now subtract the two totals: $0.944 − $0.181 = $0.763. In this example, Gemini 2.5 Flash costs $0.763 less for the same cached volume. That is the practical starting point: read the cache-read row before you pick a model for a repeat-heavy agent session.

Gemini 3.1 Pro sits higher on that row at $0.1125 per 1,000,000 tokens. At 10,000,000 cache-read tokens, the example is 10 × $0.1125 = $1.125. Against Gemini 2.5 Flash, that is $1.125 − $0.181 = $0.944 more in this example. The numbers do the talking.

03 / Claude and GPT cache reads against the official list price

For Claude and GPT models, the price list includes official list prices, so the gap can be shown directly. The official figures use the same per-1,000,000-token unit.

• ■ Claude Haiku 4.5 costs $0.025 per 1,000,000 cache-read tokens on CheapRelay, against an official list cache read of $0.10. The example gap is $0.10 − $0.025 = $0.075 per 1,000,000 tokens, or 10 × $0.075 = $0.75 across 10,000,000 tokens.
• ■ Claude Sonnet 5.5 costs $0.0778, against an official list cache read of $0.10. The example gap is $0.0222 per 1,000,000 tokens, or $0.222 across 10,000,000 tokens.
• ■ Claude Opus 5 costs $0.1222, against an official list cache read of $0.50. The example gap is $0.3778 per 1,000,000 tokens, or $3.778 across 10,000,000 tokens.
• ■ GPT-6.1 Sol costs $0.0347, against an official list cache read of $0.10. The example gap is $0.0653 per 1,000,000 tokens, or $0.653 across 10,000,000 tokens.
• ■ GPT-5.6 Sol costs $0.0264, against an official list cache read of $0.40. The example gap is $0.3736 per 1,000,000 tokens, or $3.736 across 10,000,000 tokens.
• ■ GPT-6 Astra costs $0.0569, against an official list cache read of $1.00. The example gap is $0.9431 per 1,000,000 tokens, or $9.431 across 10,000,000 tokens.

These are examples on one token volume, not a forecast of your usage. They simply show how the arithmetic works when the same amount of cached material is read repeatedly.

04 / Where the cache-read line gets especially small

Several models sit low on the cache-read row. Here are the price-list rates, followed by the cost of a 10,000,000-token example.

• ■ MiMo V2.6 Pro: $0.002778 per 1,000,000 tokens, so 10 × $0.002778 = $0.02778.
• ■ DeepSeek V4 Flash: $0.005556 per 1,000,000 tokens, so 10 × $0.005556 = $0.05556.
• ■ MiniMax M3: $0.025 per 1,000,000 tokens, so 10 × $0.025 = $0.25.
• ■ DeepSeek V4 Pro: $0.0333 per 1,000,000 tokens, so 10 × $0.0333 = $0.333.
• ■ GPT-6.1 Sol: $0.0347 per 1,000,000 tokens, so 10 × $0.0347 = $0.347.
• ■ Qwen3.7 Plus: $0.0361 per 1,000,000 tokens, so 10 × $0.0361 = $0.361.
• ■ Claude Opus 5.5: $0.0597 per 1,000,000 tokens, so 10 × $0.0597 = $0.597.
• ■ Qwen3.8 Max: $0.0681 per 1,000,000 tokens, so 10 × $0.0681 = $0.681.
• ■ Claude Sonnet 5.5: $0.0778 per 1,000,000 tokens, so 10 × $0.0778 = $0.778.
• ■ Claude Sonnet 5: $0.0861 per 1,000,000 tokens, so 10 × $0.0861 = $0.861.
• ■ Kimi K3: $0.0931 per 1,000,000 tokens, so 10 × $0.0931 = $0.931.
• ■ Gemini 3.8 Flash: $0.0944 per 1,000,000 tokens, so 10 × $0.0944 = $0.944.
• ■ Gemini 3.1 Pro: $0.1125 per 1,000,000 tokens, so 10 × $0.1125 = $1.125.
• ■ GLM-5.3: $0.1153 per 1,000,000 tokens, so 10 × $0.1153 = $1.153.
• ■ Grok 4.6: $0.1167 per 1,000,000 tokens, so 10 × $0.1167 = $1.167.
• ■ Claude Opus 5: $0.1222 per 1,000,000 tokens, so 10 × $0.1222 = $1.222.
• ■ Grok 4.7: $0.1264 per 1,000,000 tokens, so 10 × $0.1264 = $1.264.

Qwen3.6 Plus Uncensored has no cache-read price on the price list, so it is not included in this comparison. In the example above, the spread between MiMo V2.6 Pro at $0.02778 and Grok 4.7 at $1.264 is $1.23622 for the identical 10,000,000 cached tokens. Pick the line that fits the job, then check the rest of the bill.

05 / Build the bill from all three token types

Cache reads are only one line. A real request also contains fresh input and generated output, so a useful example should show all three together. Take this example workload on Gemini 2.5 Flash: 1,000,000 input tokens, 1,000,000 output tokens, and 10,000,000 cache-read tokens.

• ■ Input: 1 × $0.1778 = $0.1778
• ■ Output: 1 × $1.42 = $1.42
• ■ Cache reads: 10 × $0.0181 = $0.181
• ■ Total example cost: $0.1778 + $1.42 + $0.181 = $1.7788

Now run the same example on Gemini 3.1 Pro, where the price list shows $1.12 input, $6.49 output, and $0.1125 cache read per 1,000,000 tokens:

• ■ Input: 1 × $1.12 = $1.12
• ■ Output: 1 × $6.49 = $6.49
• ■ Cache reads: 10 × $0.1125 = $1.125
• ■ Total example cost: $1.12 + $6.49 + $1.125 = $8.735

The difference in this example is $8.735 − $1.7788 = $6.9562. The cache-read line contributes $1.125 − $0.181 = $0.944 of that gap, so it matters, but input and output still decide most of the total. The sensible method is to price the whole workload, then use the cache-read rate to trim the repeat-heavy part.

06 / Keep the setup simple

CheapRelay uses a prepaid balance and charges for the real tokens you use. There is no monthly fee, and the balance does not expire. A new account starts with $0.09 in signup credit, usable on any listed model, which is enough to run a small first test and confirm the three lines against your own workload.

For Claude models, use the official Claude Code or Claude Desktop client, version 2.1.285 or newer. For Gemini models, use the native protocol under /v1beta through Gemini CLI, and enter the model id exactly as listed, for example gemini-2.5-flash. GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax, and the other OpenAI-compatible models use an OpenAI-compatible client under /v1.

The account limit is 120 calls per minute per key, with 8 calls at once per account. If a request returns HTTP 402, the balance is at or below zero, so top up and try again. If you top up within 12 days, the unused balance can be refunded, minus what was spent and the payment fee.

Referrals add a little more: directly invited users generate 20% of what they pay, the second level generates 6%, and that goes into a withdrawable rewards balance with a ¥20 minimum withdrawal through Alipay. You do not need the referral percentages to make the point. The cache-read row on the price list already does it: read it, multiply it by your real cached-token volume, subtract, and see which model leaves more money in your prepaid balance.