On this page
Which versions are we comparing?
The main performance comparison is Mistral Large 4 public preview versus DeepSeek-V4-Pro-0813, served under deepseek-v4-pro. We also show DeepSeek-V4.1-Flash, served under deepseek-flash, as a separate cost option.
The DeepSeek source is its direct API documentation, not a reseller. The model names, rates and limits come from the provider documentation. Flash prices are not evidence about Pro’s performance, and a Pro score does not describe Flash.
Comparison rule: keep the provider, endpoint, date and pricing conditions attached to the result. If one changes, check the comparison again.
Side-by-side specifications
| Detail | Le Chonk preview | DeepSeek-V4-Pro-0813 | DeepSeek-V4.1-Flash |
|---|---|---|---|
| Provider | Mistral AI | DeepSeek direct API | DeepSeek direct API |
| Endpoint | mistral-large-4 | deepseek-v4-pro | deepseek-flash |
| Input / output price | $0.68 / $2.09 | $0.66 / $1.98 off-peak | $0.15 / $0.60 off-peak |
| Context | 1M tokens | 1M tokens | 1M tokens |
| Modalities | Multimodal input; text output | Text input and output; no vision | Text and image input; text output |
| Weight release | Pending release | Exact snapshot not verified | Exact snapshot not verified |
| Weight license | Not verified for this exact version | Not verified for this exact version | Not verified for this exact version |
| Same-condition test by this site | Not run | Not run | Not run |
DeepSeek Pro’s documented endpoint does not support vision. Flash does. Peak rates are separate rows in the calculator below; do not use off-peak pricing for an all-day workload.
Same workload, different bill
For the default workload, Le Chonk costs $34.50 a month. DeepSeek Pro 0813 costs $33.00 off-peak or $66.00 at peak rates. Flash costs $9.00 off-peak or $18.00 at peak rates.
Pro’s off-peak advantage in this example is just $1.50 a month. It can disappear if the task uses more output tokens or needs another attempt. At peak rates, Pro’s estimated bill is $31.50 higher than Le Chonk’s.
DeepSeek applies different rates during peak and off-peak periods; check its pricing page for the current schedule. For mixed schedules, calculate the requests in each period separately and add the results. Do not average prices without knowing where your requests fall.
10,000 requests per month
| Model / provider | Input / output¹ | Per request | Monthly | vs Le Chonk² |
|---|---|---|---|---|
| Le ChonkMistral AI ↗ | $0.68 / $2.09 | $0.00345 | $34.50 | Baseline |
| Qwen3.8-MaxQwenCloud ↗ | $2.00 / $6.00 | $0.01000 | $100.00 | +$65.50 |
| DeepSeek Pro · off-peakDeepSeek direct API ↗ | $0.66 / $1.98 | $0.00330 | $33.00 | −$1.50 |
| DeepSeek Pro · peakDeepSeek direct API ↗ | $1.32 / $3.96 | $0.00660 | $66.00 | +$31.50 |
| DeepSeek Flash · off-peakDeepSeek direct API ↗ | $0.15 / $0.60 | $0.00090 | $9.00 | −$25.50 |
| DeepSeek Flash · peakDeepSeek direct API ↗ | $0.30 / $1.20 | $0.00180 | $18.00 | −$16.50 |
¹ USD per million uncached tokens. ² Monthly difference; a minus sign means cheaper. No caching, batch discounts, tool fees, taxes or tokenization differences. DeepSeek peak and off-peak rates are shown separately. Mistral also lists higher $1.36 / $4.18 rates; no discount end date confirmed.
The table compares a fixed number of tokens, not measured task cost. Tokenizers, reasoning output, tool use and retries can differ. It excludes caching, batch pricing, tax and tool charges.
What can the benchmarks tell us?
Mistral’s launch post reports a 49.8% Coding Agent Index result and places its preview ahead of Qwen3.8 Max and DeepSeek V4 Pro 0813. That is a vendor’s report of an external evaluation. We have not checked the evaluator’s complete run records.
We do not have directly verified, same-condition scores for both sides of this page. There is no site-run head-to-head test, latency measurement or quality rating. That leaves us without enough evidence to name an overall performance winner.
A coding agent test also includes more than the underlying model: tools, time limits and retry budgets matter. A score may help pick models to try, but it does not measure how well your application handles its own data.
Inspect the benchmark evidence and missing conditions →How to choose for your own work
For coding
Take a few real issues from a repository you can test. Give each model the same files, tools and time limit. Run the tests, review the patch, and count completed issues. Record the total token spend, including unsuccessful attempts.
For document work
Use documents with known answers and include questions that the documents cannot answer. Check quotations, page references and missing facts. For images or scans, verify the exact endpoint supports the input before sending a batch.
For a tight budget
Flash has the lowest listed token bill here, but it is not the Pro model. Test whether it meets your minimum quality standard. If Pro is a better fit, check whether you can schedule non-urgent work outside peak periods.
Keep a small decision sheet
- Exact model version and provider.
- Accepted tasks out of attempted tasks.
- Wall-clock time, tool errors and manual fixes.
- Input, output, cached tokens and total cost.
- Which failures matter enough to reject a model.
Set the acceptance rule before looking at the model name. That makes it easier to choose based on the work you need done.
What about self-hosting?
Le Chonk’s weights are pending a verified release. This page also does not establish a downloadable artifact and license for the exact DeepSeek snapshots listed above. A model family having open weights does not prove that a particular hosted version can be reproduced locally.
For a self-hosted comparison, record the actual repository revision, quantization, inference framework and hardware. Include idle GPU time, maintenance and throughput in the cost. API rates alone cannot tell you whether renting hardware will save money.
See why active parameters are not the memory budget →Sources & next steps
- Mistral Large 4 release announcement
API preview, planned weight release, 49B active parameters and reported benchmark results.
- Mistral Large model documentation
1.05T total, 52B active, 1M context. Displays both $0.68 / $2.09 and higher $1.36 / $4.18 prices. No discount end date confirmed.
- DeepSeek models and pricing
Direct API rates. Separate Pro 0813 and V4.1 Flash. Peak hours: weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays.