Gemini 4 Argon vs Claude Opus 5: The Receipt

01 / The price tags before the reasoning starts

A difficult debugging prompt does not arrive with a fixed bill. Context, reasoning, tool loops, and retries determine what you actually spend. For Gemini 4 Argon versus Claude Opus 5, start with a receipt—not a claim that either model always reasons better.

As of October 25, 2026, Google lists introductory Argon pricing of $2 per million input tokens and $10 per million output tokens through November 30, 2026. Those rates become $4 and $20 on December 1, 2026. These are Google's published API prices, not a CheapRelay quote. ([blog.google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/?utm_source=openai))

Anthropic lists standard Claude Opus 5 pricing at $5 per million input tokens and $25 per million output tokens. The comparison below uses standard processing, without batch pricing or premium-speed options. ([anthropic.com](https://www.anthropic.com/news/claude-opus-5))

02 / Receipt one: 20,000 input and 12,000 output tokens

Assume a hypothetical reasoning task: inspect a specification, diagnose a bug, and propose a tested correction. Budget 20,000 uncached input tokens and 12,000 billable output tokens per attempt. These are planning quantities, not measured consumption or benchmark results.

• ■ Argon input: 20,000 ÷ 1,000,000 × $2 = $0.04.
• ■ Argon output: 12,000 ÷ 1,000,000 × $10 = $0.12.
• ■ Argon total: $0.16 per attempt; $16 for 100 identical attempts.
• ■ Opus input: 20,000 ÷ 1,000,000 × $5 = $0.10.
• ■ Opus output: 12,000 ÷ 1,000,000 × $25 = $0.30.
• ■ Opus total: $0.40 per attempt; $40 for 100 identical attempts.

At identical token counts, the introductory Argon receipt is $0.24 lower per attempt. After its announced price change, that same receipt becomes $0.32: $0.08 input plus $0.24 output. A budget built around today's introductory rate therefore needs a dated assumption. ([cloud.google.com](https://cloud.google.com/blog/products/ai-machine-learning/gemini-4-argon-available-on-google-cloud?hl=en&utm_source=openai))

03 / The 12,000-token line includes more than the answer

Do not estimate this bill by counting the final paragraph. Google's documentation explicitly includes thinking tokens in Argon's output pricing. A short visible answer can accompany a larger billable output count. Use reported usage rather than the length of text displayed in your application. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/latest-model?frmapp=yes&utm_source=openai))

For a fair experiment, keep the task, available tools, and acceptance criteria constant. Record each model's actual consumption independently; identical prompts do not establish identical token counts. Start with a bounded output allowance and increase it only when your evaluation shows that the extra work helps.

04 / Receipt two: a confirmed cache hit

Argon's documented introductory cached-input rate is $0.10 per million tokens. Its caching is implicit, so treat a hit as something to verify—not a saving to assume. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/latest-model?frmapp=yes&utm_source=openai))

Suppose usage confirms that 16,000 of the 20,000 input tokens were cached. Keeping the hypothetical 12,000 output tokens unchanged gives this receipt:

• ■ Cached input: 16,000 ÷ 1,000,000 × $0.10 = $0.0016.
• ■ Fresh input: 4,000 ÷ 1,000,000 × $2 = $0.0080.
• ■ Output: 12,000 ÷ 1,000,000 × $10 = $0.1200.
• ■ Total: $0.1296, saving $0.0304 against the uncached example.

This is a conditional calculation, not an observed cache result or a cache-adjusted Opus comparison. Notice that output still dominates the receipt.

05 / Three cheap attempts can lose to one accepted answer

Using the uncached examples, two Argon attempts cost $0.32; three cost $0.48. One Opus attempt costs $0.40. If three attempts are needed to match one accepted result, the lower token price no longer wins. That is arithmetic, not a claim about either model's success rate.

Evaluate cost per accepted result: total spending divided by outputs that pass your checks. Track human review separately. At an assumed labor cost of $60 per hour, five minutes of review adds $5—far more than the $0.24 first-attempt difference.

06 / Keep the relay bill separate from provider list prices

CheapRelay offers Claude Opus 5 with pay-as-you-go billing per real token, including cache reads and writes, without a monthly fee or subscription. Check its current model-specific rates before substituting them into these receipts; provider list prices are not relay prices. Argon's comparison here does not imply availability through CheapRelay.

Sign up for CheapRelay to run a small, capped Opus trial against your own acceptance tests. Keep the winning answer and its complete receipt together: tokens, retries, cache usage, and review time.