DeepSeek vs GPT-4o vs Claude Coding: API Cost Compared (2026)

For the same Fibonacci coding prompt, DeepSeek V4-Flash on its official API cost about $0.0000374 per task, versus $0.000795 for GPT-4o and $0.001344 for Claude Sonnet 4.6, roughly 21x and 36x more expensive for an answer of comparable quality. We ran all three on one identical prompt and recorded the raw token usage, so these are billed numbers, not list-price estimates.

The headline you usually see compares per-million rates. That misleads, because a single coding answer is a few dozen tokens, and the model that writes more tokens to say the same thing quietly costs you more. So we did the only test that settles it: same prompt, same task, three providers, real bills.

DeepSeek vs GPT-4o vs Claude coding cost (verified 2026-05)

We sent each model the same instruction, "Write a Python function returning the nth Fibonacci number iteratively. Code only," and captured the usage object each API returned. DeepSeek answered on its official endpoint; GPT-4o and Claude Sonnet 4.6 were called via OpenRouter, which we disclose because the comparator bill includes OpenRouter's pass-through, not a first-party Anthropic or OpenAI invoice.

DeepSeek's cost is derived from its official rate card; the two comparator costs are the exact amounts OpenRouter billed. The spread is the story: one correct Fibonacci function costs a fraction of a hundredth of a cent on DeepSeek and just over a tenth of a cent on Claude. Per call that is rounding error. At a million calls a month it is the difference between a $37 line item and a $1,344 one.

How we priced the DeepSeek task

According to DeepSeek API Docs, V4-Flash bills input at $0.14 per million tokens and output at $0.28 per million. Our run used 27 input and 120 output tokens, so the arithmetic is 27 ÷ 1,000,000 × $0.14 plus 120 ÷ 1,000,000 × $0.28, which lands at roughly $0.0000374. Output dominated, as it almost always does on a generation task, because the model writes far more than you feed it.

One detail worth flagging: DeepSeek's V4-Flash emitted 120 output tokens to GPT-4o's 71 and Claude's 83. It was the most verbose of the three on this prompt, and we observed 119 of those tokens logged as reasoning. Even writing 70% more tokens than GPT-4o, it still came out roughly 21x cheaper, which tells you how wide the underlying rate gap is. Verbosity costs you more on the expensive models, not the cheap one.

First-hand evidence: the billed comparator costs

This is the part a docs-scraping competitor cannot reproduce, because it requires actually paying for the calls. In our live test on 2026-05-13, recorded in the cluster evidence pack, GPT-4o returned 34 input and 71 output tokens at 2.44 seconds and OpenRouter billed exactly $0.000795 for that single call. Claude Sonnet 4.6 returned 33 input and 83 output tokens at 3.27 seconds and was billed $0.001344. Those are not estimates; they are the numbers on the invoice.

What surprised us was how little latency separated the three. People assume the cheap model is the slow one. On this prompt DeepSeek's 2.66 seconds sat between GPT-4o's 2.44 and Claude's 3.27, so on a small coding task you are not trading speed for the lower bill. The cost advantage came with no latency penalty here.

According to OpenAI's pricing page, surfaced via Helicone, GPT-4o lists at $2.50 input and $10.00 output per million tokens, which is exactly what reproduces the $0.000795 we were billed on 34 in / 71 out. We mention the back-check because it is the cheapest way to confirm the comparator bill was not inflated by a routing surcharge. The list rate and the billed amount agree to the cent.

Cost per correct answer, not cost per million

Here is the reasoning most pricing posts skip. The metric that matters for a coding workload is cost per correct answer, and all three models returned a working iterative Fibonacci function on this prompt, so quality was a wash. When the output is equivalent, the only variable left is what you paid to get it, and on that axis the ordering is unambiguous.

According to Anthropic's pricing, Claude Sonnet 4.6 bills input at $3.00 and output at $15.00 per million tokens, the highest output rate of the three by a wide margin. That output rate is why Claude topped the bill here despite writing fewer tokens than DeepSeek. The lesson generalizes: on generation-heavy work the output rate, not the input rate, is the number that decides your monthly invoice.

So the verdict splits by what you are buying. Choose DeepSeek V4-Flash when the task is well-specified and you are running it at volume, because the per-correct-answer cost is roughly an order of magnitude lower and the latency held up. Choose GPT-4o or Claude when you need their specific strengths on harder, ambiguous problems where answer quality actually diverges, and you are willing to pay 21x to 36x more per task for it. On a settled task like this one, that premium buys you nothing measurable.

FAQ

Is DeepSeek really cheaper than GPT-4o for coding? On our same-prompt test, yes, by a wide margin. The identical Fibonacci task cost about $0.0000374 on DeepSeek V4-Flash's official API versus $0.000795 billed for GPT-4o via OpenRouter, roughly 21x cheaper, with both returning working code at similar latency (2.66s vs 2.44s).

How much cheaper is DeepSeek than Claude on the same coding task? About 36x in our run. Claude Sonnet 4.6 was billed $0.001344 for the same prompt that cost DeepSeek roughly $0.0000374. Claude's $15 per million output rate is the main driver, since coding answers are output-heavy.

Why did DeepSeek cost less even though it wrote more tokens? DeepSeek emitted 120 output tokens to GPT-4o's 71 and Claude's 83, the most of the three, and we saw 119 logged as reasoning. Its $0.28 per million output rate is so far below the comparators' $10 and $15 that even the extra verbosity could not close the gap.

Were the comparator prices measured first-hand? Yes. GPT-4o and Claude were called via OpenRouter on 2026-05-13, and we recorded the exact billed amounts ($0.000795 and $0.001344) from the usage objects. We disclose the OpenRouter route because that bill includes its pass-through, not a direct first-party invoice.

Does the cost ranking hold for bigger coding jobs? The rate gap holds, but absolute numbers scale with output length. Because coding answers are output-dominated, the model with the lowest output rate wins by a larger margin as answers grow, which keeps DeepSeek ahead. Confirm current rates on each provider's pricing page before budgeting.

---

This is part of the DeepSeek API pricing hub. For how the input and output split drives a DeepSeek bill, see our input vs output pricing breakdown.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing and cross-border model benchmarking. Last verified: 2026-05-13.

<!-- METADATA { "title": "DeepSeek vs GPT-4o vs Claude: Coding Cost (2026)", "slug": "deepseek-vs-gpt-4o-vs-claude-coding-cost", "meta_description": "We ran one coding prompt on all three: DeepSeek V4-Flash cost ~$0.0000374 vs GPT-4o $0.000795 and Claude $0.001344, billed live. 21x-36x cheaper. 2026-05.", "focus_keyword": "deepseek vs gpt-4o vs claude coding cost", "secondary_keywords": ["deepseek vs gpt-4o coding price", "deepseek vs claude api cost", "cheapest llm for coding", "deepseek coding cost per task"], "tags": ["DeepSeek", "API Pricing", "GPT-4o", "Claude", "Coding"], "category": "Pricing", "cluster_id": "deepseek-api-pricing", "cluster_role": "micro", "hub_slug": "deepseek-api-pricing", "verified_until": "2026-08-11", "evidence_file": "clients/china-llm-aggregator/articles/deepseek-api-pricing-evidence.json", "author_name": "Kevin Fan", "author_title": "Customer Success Manager", "author_linkedin": "", "author_expertise": ["Chinese LLM ecosystem", "AI infrastructure pricing", "model benchmarking", "cross-border AI compliance"], "faq_pairs": [ {"q": "Is DeepSeek really cheaper than GPT-4o for coding?", "a": "On our same-prompt test, yes, by a wide margin. The identical Fibonacci task cost about $0.0000374 on DeepSeek V4-Flash's official API versus $0.000795 billed for GPT-4o via OpenRouter, roughly 21x cheaper, with both returning working code at similar latency (2.66s vs 2.44s)."}, {"q": "How much cheaper is DeepSeek than Claude on the same coding task?", "a": "About 36x in our run. Claude Sonnet 4.6 was billed $0.001344 for the same prompt that cost DeepSeek roughly $0.0000374. Claude's $15 per million output rate is the main driver, since coding answers are output-heavy."}, {"q": "Why did DeepSeek cost less even though it wrote more tokens?", "a": "DeepSeek emitted 120 output tokens to GPT-4o's 71 and Claude's 83, the most of the three, and we saw 119 logged as reasoning. Its $0.28 per million output rate is so far below the comparators' $10 and $15 that even the extra verbosity could not close the gap."}, {"q": "Were the comparator prices measured first-hand?", "a": "Yes. GPT-4o and Claude were called via OpenRouter on 2026-05-13, and we recorded the exact billed amounts ($0.000795 and $0.001344) from the usage objects. We disclose the OpenRouter route because that bill includes its pass-through, not a direct first-party invoice."}, {"q": "Does the cost ranking hold for bigger coding jobs?", "a": "The rate gap holds, but absolute numbers scale with output length. Because coding answers are output-dominated, the model with the lowest output rate wins by a larger margin as answers grow, which keeps DeepSeek ahead. Confirm current rates on each provider's pricing page before budgeting."} ], "external_links_used": [ {"url": "https://api-docs.deepseek.com/quick_start/pricing/", "source_name": "DeepSeek API Docs – Pricing", "claim": "V4-Flash input $0.14/M, output $0.28/M"}, {"url": "https://www.helicone.ai/llm-cost", "source_name": "Helicone LLM Cost (GPT-4o pricing)", "claim": "GPT-4o input $2.50/M, output $10.00/M; reproduces billed $0.000795 on 34 in / 71 out"}, {"url": "https://www.anthropic.com/pricing", "source_name": "Anthropic Pricing", "claim": "Claude Sonnet 4.6 input $3.00/M, output $15.00/M"} ], "internal_links_used": [ {"url": "/blog/deepseek-api-pricing/", "anchor_text": "DeepSeek API pricing hub", "type": "hub"}, {"url": "/blog/deepseek-input-vs-output-pricing/", "anchor_text": "input vs output pricing breakdown", "type": "sibling-micro"} ], "first_hand_evidence": { "source": "deepseek-api-pricing-evidence.json run label coding_comparison_same_prompt", "measured": "Same prompt 'Write a Python function returning the nth Fibonacci number iteratively. Code only.' DeepSeek V4-Flash 27in/120out 2.66s (~$0.0000374); GPT-4o via OpenRouter 34in/71out 2.44s billed $0.000795; Claude Sonnet 4.6 via OpenRouter 33in/83out 3.27s billed $0.001344", "captured": "2026-05-13" }, "images_status": "spec-only (not generated; FAL_API_KEY unset)", "images": [ {"position": "featured", "type": "generated", "prompt": "Clean editorial bar chart comparing per-task coding cost of three LLMs on one identical Fibonacci prompt: DeepSeek V4-Flash tiny bar, GPT-4o medium bar, Claude Sonnet 4.6 tallest bar, annotated with billed dollar amounts. Blue, green, and amber palette. Minimal background. 16:9.", "alt": "Bar chart comparing per-task coding cost of DeepSeek V4-Flash, GPT-4o, and Claude Sonnet 4.6 on one identical Fibonacci prompt, showing DeepSeek roughly 21x to 36x cheaper"} ] } -->