A 740-Million-Token Bill for Around $10 — What the Same Tokens Would Cost on Claude and ChatGPT

Anyone who has run a large language model as a daily agent, or wired one into a heavy development environment, eventually hits the same wall: token cost. Keep throwing text analysis, code review, and multimodal prompts at it, and the API bill balloons fast. The cost efficiency of recent Chinese models turns that conventional wisdom completely on its head.

This is not a conceptual explainer. It is an actual invoice record. Running a two-track strategy — text-heavy, painstaking logical work on DeepSeek and the everyday multimodal agent role on Xiaomi's MiMo — burned a combined 740 million tokens yet produced a bill in the low tens of dollars. Every figure below is exactly what the dashboards showed; any converted or estimated numbers are labeled as such.

The bill at a glance

Together the two models account for 739,784,990 tokens, roughly 740 million. The bill for all of it came to $9.97 across the two models, and still only the low tens of dollars on a cumulative basis.

DeepSeek: the text workhorse that swallowed 360M tokens

• Total API requests: 3,243
• Cumulative tokens: 359,761,623 (about 360 million)
• Cost for this window: $4.86 (part of an $11.34 cumulative total)

More than three thousand individual API calls and 360 million tokens, yet the bill is under five dollars — less than a coffee. That works out to roughly 110,000 tokens per request, which tells you each call carried part of a codebase plus task instructions plus tool output.

Xiaomi MiMo: the multimodal miracle built on cache hits

• Cumulative consumption: $5.11
• Total token consumption: 380,023,367 (about 380 million)
• Input cache hit: 354,696,567 tokens (about 93.8% of all input)
• Input cache miss: 23,258,315 tokens
• Output tokens: 2,068,485

The secret behind MiMo handling 380 million tokens for a little over five dollars is its overwhelming input cache hit rate. An agent re-reads the same working environment and context on every request. MiMo hits the cache on more than 93% of all input tokens, cutting the genuinely new tokens that must be computed down to around 23 million. Cached-hit tokens are billed far more cheaply than fresh input, so even continuously shipping heavy multimodal data barely moves the cost.

For caching to work well, the front of the prompt — system instructions, tool definitions, fixed documents — has to stay stable across requests. Shuffle the context on every request and the cache breaks, multiplying the bill. That 93.8% is not a product of model quality; it is the result of how the prompt is operated.

What the same tokens would cost on other models

Here is the key part. I calculated what the same 740 million tokens would cost on the most widely used Claude and ChatGPT models. Prices come from each company's official price list (looked up 2026-10-04), and both tables below are estimates, not measured values.

1) Plain input conversion — all 740 million tokens

This is the ceiling case: no cache or output discounts, just the total token count multiplied by each model's input rate.

Run the same 740 million tokens through Claude Opus 5.5 and the plain-input conversion is about $2,959; through GPT-6 Astra it is about $7,398. Against the actual $9.97 bill, that is roughly 297x and roughly 742x respectively.

2) Cache structure preserved — MiMo's measured window

MiMo's actual input broke down into 355M cache hits, 23M misses, and 2M output tokens. Keeping that exact structure and only swapping the model yields the following.

Even with cache discounts applied, running MiMo's window on Claude Opus 5.5 costs about $205, and on GPT-6 Astra about $691. Against MiMo's actual $5.11, that is roughly 40x and roughly 135x. For reference, GPT-4o — which has one of the least favorable cache-read rates — lands around $522, about 100x more than MiMo.

In other words, this near-free bill is not the cache efficiency alone. Half of it comes from the model's own input and output rates already being more than an order of magnitude lower, and the other half from how the cache is operated.

Why two tracks is the answer between performance and cost

• Logic, code, text: DeepSeek. Token-heavy, reasoning-intensive work handled by cheap text compute.
• Context retention, always-on multimodal loop: MiMo. Because the same context is re-read repeatedly, cache-hit efficiency translates directly into cost savings.

Split the models by the nature of the work, and you stop wasting expensive multimodal compute on text reasoning, while defending the cache-friendly always-on loop with a cheap cache rate. That is why 740 million tokens still produced a bill in the low tens of dollars.

When you genuinely need frontier-model quality, use it then and only then. The moment you put the always-on loop and high-volume repetitive work on a flagship model, the three- and four-digit bills in the tables above become real.

Caveats

• Every amount in the comparison tables is an estimate — a token quantity multiplied by public rates. It differs from dashboard-measured values and shifts with cache policy, plan tiers, and batch discounts.
• Cache hit rate depends on prompt stability. Change system instructions or tool definitions on every request and the cache breaks, spiking cost.
• Logging and token-counting windows differ by model, so do not compare the two models' request counts one to one.
• These figures come from a single operating environment. When model versions and cache policies change, the rate structure changes with them.
• Cheap does not mean you should recklessly put sensitive data into the input.

Conclusion

For heavy users looking for a compromise between performance and cost, the DeepSeek-plus-MiMo ensemble is close to irreplaceable. Reasoning goes to DeepSeek's cheap text compute; the multimodal loop that must retain context is defended by MiMo's cache efficiency.

Run the same 740 million tokens on a flagship and you get a bill in the hundreds or thousands. A ten-dollar bill now shows that the question is no longer "how much AI can I afford to use" but "how do I combine and operate these models."