DeepSeek V4 Flash Vision vs Opus 4.8: Cost & Benchmarks

Full benchmark comparison

DeepSeek’s self-reported results show a model that is competitive on agentic and chart tasks, but not consistently ahead of Claude Opus 4.8. Its strongest relative results appear on Agents' Last Exam and ZeroBench, while Chartography and Terminal Bench 2.1 remain close. 

The largest measured loss in the transcribed results is repository-scale code reasoning: on NL2Repo, DeepSeek V4 Flash Vision trails Opus 4.8 by 12.0 points. That pattern supports using DeepSeek where price matters and testing Opus where stronger code reasoning can justify the higher cost.

Note: DeepSeek’s published table covers 11 benchmarks, and 4 are not yet transcribed.

How DeepSeek ran these benchmarks

DeepSeek evaluated its own models and published these results using its own Harness Minimal Mode. The reported configuration sets max tokens at the model ceiling, with top_p 0.95 and temperature 1.0. No independent verification of these DeepSeek V4 Flash Vision benchmarks exists yet.

That matters when interpreting the comparison. A model vendor controls the evaluation harness, configuration, prompting choices, and other test conditions when running its own benchmarks. Those choices can naturally suit the model the vendor knows best. This does not imply manipulation or dishonesty. Self-run evaluations are normal when a lab announces a new model, especially immediately after release.

Treat the scores as directional evidence, not neutral proof of relative performance. They are useful for deciding which models deserve testing, but not for choosing a production model by themselves. Before routing a production vision workload, run DeepSeek V4 Flash Vision and Claude Opus 4.8 against your own documents, charts, prompts, and expected outputs.

Where DeepSeek V4 Flash Vision beats Opus 4.8

DeepSeek V4 Flash Vision posts two verified wins over Claude Opus 4.8 in the transcribed DeepSeek self-reported benchmarks. On Agents' Last Exam, it scores 27.3 vs 25.7, a 1.6-point lead. On ZeroBench (Pass@5), it scores 35.0 vs 34.0, a 1.0-point lead.

Both benchmarks focus on agentic or visual reasoning rather than raw knowledge recall. That suggests DeepSeek is genuinely competitive for vision workloads where the model must interpret visual information and reason through a task, rather than simply identify or retrieve information.

The margins are narrow. A 1.6-point or 1.0-point advantage is within the range where evaluation harness differences can matter, especially because DeepSeek ran these tests using its own methodology. These results justify testing DeepSeek against Opus 4.8 on agentic vision workloads, but they do not establish a broad performance advantage.

Where Claude Opus 4.8 still wins

Claude Opus 4.8 leads on the largest reasoning-heavy gaps in DeepSeek’s self-reported benchmarks. The biggest difference is NL2Repo, where DeepSeek V4 Flash Vision scores 57.7 vs 69.7, a 12.0-point deficit. On DSBench-Hard, Opus 4.8 leads by approximately 8 points. The gap falls to 2.9 points on ApexBench, with 36.5 vs 39.4, and to 1.1 points on Terminal Bench 2.1, with 83.9 vs 85.0.

The pattern is clear in these results. The performance gap widens when tasks require multi-step reasoning across large codebases or datasets. It narrows on more direct visual tasks, where DeepSeek comes closer to Opus 4.8.

Route repository-scale code and complex data reasoning to Claude Opus 4.8, and route lower-cost visual workloads to DeepSeek V4 Flash Vision.

What vision added over DeepSeek V4 Flash

The useful comparison is not only against Claude Opus 4.8. It is also against the text-only DeepSeek V4 Flash 0731 base model. On ApexBench, DeepSeek V4 Flash Vision scores 36.5, up from 26.2 for V4 Flash 0731, a 10.3-point gain. On Terminal Bench 2.1, the score moves from 82.7 to 83.9, a much smaller 1.2-point gain, at the same price tier.

That difference helps isolate what image input contributes. Adding vision produces a large improvement on the benchmark that requires visual context, while changing performance very little on the benchmark that does not. In DeepSeek’s self-reported results, vision is therefore doing real work on tasks that can use visual information rather than simply raising benchmark scores across the board.

Pricing: cost per image, not cost per token

Per-million-token pricing can mislead when you compare vision models. Frontier models consume roughly 1,000 to 4,000 tokens per image depending on resolution, while DeepSeek V4 Flash Vision caps each image at 384 input tokens. That cap is the mechanism behind the cost difference. The $0.22 per 1M input tokens rate matters, but limiting how many tokens an image can consume is what keeps image processing costs predictable.

Only the DeepSeek row is verified against provider pricing. The competitor rows use the unverified published token rates provided for this comparison and typical image consumption of roughly 1,000 to 4,000 tokens.

DeepSeek also provides a free Files API. You can upload an image once and reference it by file_id across requests, removing repeat upload cost from multi-turn workflows that reuse the same image.

On Eden AI, deepseek/deepseek-v4-flash-vision-exp costs $0.22 per 1M input tokens and $0.66 per 1M output tokens, the same rates DeepSeek charges directly. Eden AI adds no gateway markup to this model.

Benchmark scores per dollar

Another way to read the DeepSeek self-reported benchmarks is to compare how much of Claude Opus 4.8’s score DeepSeek V4 Flash Vision reaches on tests where both have reported results.

Arithmetic: DeepSeek score ÷ Claude Opus 4.8 score × 100.

On three of these four benchmarks, DeepSeek delivers more than 90% of Opus 4.8’s score. At the same time, DeepSeek output costs $0.66 per 1M tokens, against $25.00 per 1M tokens for Opus 4.8 based on the unverified competitor pricing supplied for this comparison. The score gap is therefore much smaller than the output-price gap on Terminal Bench 2.1, ApexBench, and Chartography.

Percentage-of-score is only a rough heuristic, not a measure of equivalent quality. It becomes especially weak when a small benchmark difference translates into costly production errors.

When to route to DeepSeek vs Opus 4.8

Route to DeepSeek V4 Flash Vision

• High-volume document and receipt processing: DeepSeek caps each image at 384 input tokens, bringing image cost to approximately $0.000085 per image or $0.08 per 1,000 images.
• Chart reading: On DeepSeek’s self-reported Chartography benchmark, V4 Flash Vision scores 64.3 vs 65.0 for Claude Opus 4.8. The performance gap is small relative to the pricing gap.
• Agentic visual workflows: DeepSeek wins Agents' Last Exam, scoring 27.3 vs 25.7, and also leads ZeroBench (Pass@5) at 35.0 vs 34.0.
• Cost-constrained agents: DeepSeek costs $0.22 per 1M input tokens and $0.66 per 1M output tokens, while the unverified Opus 4.8 pricing supplied for this comparison is $5.00 and $25.00.
• Self-hosted text workloads where residency matters: the text-only DeepSeek V4 Flash 0731 base model is open weight under MIT. The vision variant is not open weight and cannot be self-hosted from released weights.

Route to Claude Opus 4.8

• Repository-scale code reasoning: On NL2Repo, DeepSeek scores 57.7 vs 69.7, a 12.0-point deficit.
• Hard data analysis: On DSBench-Hard, Claude Opus 4.8 leads by approximately 8 points in DeepSeek’s self-reported table.
• Complex tasks where mistakes cost more than tokens: ApexBench still favors Opus, 39.4 vs 36.5, while DeepSeek’s lower $0.66 output price only matters if the remaining quality gap is acceptable.
• Production workloads that exclude experimental models: DeepSeek V4 Flash Vision is explicitly EXPERIMENTAL, so route to Opus when that status alone fails your production requirements.

With Eden AI, this is a routing decision rather than an either/or choice: you can send each workload to the model whose benchmark profile and price fit the task.

How to call both models through one API

Eden AI exposes an OpenAI-compatible endpoint, so you can route a vision request without rebuilding your integration for each provider. For DeepSeek V4 Flash Vision, the model string is deepseek/deepseek-v4-flash-vision-exp. Switching models is then a model-string change inside the same request structure.

You can also configure a fallback list so the request can continue if the experimental DeepSeek model returns an error or hits a rate limit. This matters in production because DeepSeek V4 Flash Vision is still marked EXPERIMENTAL, while your application may still need to return a result when that first route fails.

import requests

url = "https://api.edenai.run/v3/chat/completions"

headers = {
    "Authorization": "Bearer <your-api-key>",
    "Content-Type": "application/json"
}

payload = {
    "model": "deepseek/deepseek-v4-flash-vision-exp",
    "messages": [
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "Analyze this image and summarize the key information."
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://example.com/image.jpg"
                    }
                }
            ]
        }
    ],
    "fallbacks": [
        "<claude-opus-4.8-model-id>"
    ],
    "stream": False
}

response = requests.post(url, headers=headers, json=payload)
data = response.json()

print(data)

‍

The sample sends an image to DeepSeek V4 Flash Vision through Eden AI and keeps a fallback route available if the primary model cannot complete the request.

FAQs - DeepSeek V4 Flash Vision vs Opus 4.8