DeepSeek vs ChatGPT: The Cost Gap and What to Move

Twenty comparisons will tell you which chatbot writes a nicer poem. The question that matters is which workloads you can move to a model that costs a fraction per token, and how you swap models later without rewriting the workflow.

Jason Bao
Diagram showing incoming work split between a lower-cost model and a frontier model before converging in one governed app with durable database state and audit logs

The short answer

You are asking two questions at once, and they have different answers.

If you are choosing a chatbot, ChatGPT has the deeper product. If you are choosing an API to put behind a product feature, DeepSeek publishes open weights under the MIT license, accepts both OpenAI and Anthropic SDK formats, and at the frontier tier costs a small fraction of OpenAI's flagship per million tokens. Those are separate decisions and most comparison articles answer only the first one.

Key takeaways

  • The chatbot question is about product surface and trust. The API question is about price, context, and portability. Answer them separately.
  • At the budget tier the gap is narrow. OpenAI's Batch tier on gpt-5.6-luna undercuts DeepSeek's cheapest published rate.
  • At the frontier tier the gap is wide. On the same workload, deepseek-v4-pro runs about 5% of the cost of gpt-6-astra.
  • DeepSeek's price depends on when your job runs. Standard rates apply 01:00-04:00 and 06:00-10:00 UTC on weekdays. Every other hour is half price.
  • DeepSeek's API takes OpenAI-format calls at https://api.deepseek.com and Anthropic-format calls at https://api.deepseek.com/anthropic, which makes switching a base URL and a model string.

What DeepSeek actually is

DeepSeek names three things: the company, the consumer chatbot, and the open-weight model family used by builders. The current family is V4, with deepseek-v4-flash and deepseek-v4-pro on the API. The models list a 1M-token context window and a maximum output of 384K tokens. DeepSeek-V4-Pro-0813 ships on Hugging Face as downloadable safetensors under the MIT license.

The API is OpenAI-shaped. tools supports up to 128 function definitions, toolchoice can force a named function, and responseformat accepts jsonobject. A thinking parameter carries a reasoningeffort setting. frequencypenalty and presencepenalty are deprecated and have no effect.

What ChatGPT actually is

ChatGPT is the product. The OpenAI API is the service builders use, and the two are priced, versioned, and governed differently.

The API publishes a ladder of models. gpt-6-astra sits at the top, while the gpt-5.6 line includes luna, terra, and sol. Separate lines cover realtime audio, images, embeddings, and code. Batch pricing is a 50% discount for work that can wait.

The differences that matter for a model-routing decision

Price per million tokens

Prices fetched from DeepSeek pricing and OpenAI pricing on 2026-09-07. Standard-tier, short-context rates in USD per 1M tokens.

  • Pricing rows: deepseek-v4-flash discount, $0.22 input, $0.007 cached input, $0.66 output. deepseek-v4-flash standard, $0.44 input, $0.014 cached input, $1.32 output. deepseek-v4-pro discount, $0.66 input, $0.022 cached input, $1.98 output. deepseek-v4-pro standard, $1.32 input, $0.044 cached input, $3.96 output. gpt-5.6-luna, $0.20 input, $0.02 cached input, $1.20 output. gpt-5.6-terra, $2.00 input, $0.20 cached input, $12.00 output. gpt-5.6-sol, $4.00 input, $0.40 cached input, $20.00 output. gpt-6-astra, $10.00 input, $1.00 cached input, $50.00 output.

DeepSeek publishes no effective date on its pricing page and says prices may change. These figures are a snapshot fetched on 2026-09-07. Recheck both pages before budgeting.

DeepSeek's standard period covers 01:00-04:00 and 06:00-10:00 UTC on weekdays. The discount rate applies outside those windows, so scheduling changes the bill.

What that costs on a real workload

Classify 100,000 support tickets at 800 input tokens and 200 output tokens each. That is 80M input tokens and 20M output tokens with no cache hits. DeepSeek flash in discount hours costs (80 × $0.22) + (20 × $0.66) = $30.80. DeepSeek pro in discount hours costs (80 × $0.66) + (20 × $1.98) = $92.40. gpt-5.6-luna costs (80 × $0.20) + (20 × $1.20) = $40.00, or $20.00 in Batch. gpt-6-astra costs (80 × $10.00) + (20 × $50.00) = $1,800. The budget tier is close, while the frontier gap is large.

Open weights and self-hosting

MIT-licensed weights on Hugging Face let a team run the model inside its own network, fine-tune without a provider tuning API, and pin a version. A 1.7 trillion parameter model requires serious capacity planning and an inference stack. Hosted API access is simpler; open weights provide an exit path and a different procurement option.

SDK compatibility and portability

DeepSeek states that its API format is compatible with OpenAI and Anthropic. OpenAI-format requests use https://api.deepseek.com and Anthropic-format requests use https://api.deepseek.com/anthropic. Switching an existing client generally means changing a base URL, key, and model string. Test both on your own traffic before changing a production route.

Product surface and tooling depth

OpenAI's platform covers realtime audio, image generation, embeddings, moderation, code models, and several pricing tiers. That breadth matters when a product needs several modalities or one vendor's tooling. It does not settle the API routing question.

Data residency and procurement

The PAA question about why DeepSeek has been banned reflects restrictions in some jurisdictions and organizations, but this article does not assert the current legal status of any specific restriction. Procurement teams should ask where data is processed and stored, how long prompts and completions are retained, who the subprocessors are, whether a zero-retention agreement exists, and whether self-hosting changes the answer.

Which model candidates fit each workload

  • High-volume classification: deepseek-v4-flash during discount hours, because the task is well specified and the cost is low.
  • Structured extraction: deepseek-v4-flash, because JSON output and tool definitions support validation downstream.
  • Long-context reasoning: deepseek-v4-pro, because its listed context window suits large corpora.
  • Customer-facing chat: gpt-5.6-terra, because broader product tooling may matter more than token price.
  • Code generation: gpt-5.3-codex, because the brief calls for a code-specific candidate and latency matters in a developer loop.
  • Hard judgment calls: gpt-6-astra, because low-volume, high-stakes work may justify frontier capability.
  • Strict data residency: self-hosted DeepSeek-V4-Pro weights, because the vendor is removed from the data path while the team owns operations.

Treat each row as a test hypothesis, measure it on your own traffic, and revise the routing rule when the evidence changes. Major is the best way to build and govern that internal routing app.

The part nobody plans for: switching

Model choice is reversible. Workflow architecture usually is not. In many teams, taxonomy, retry logic, escalation rules, and system-of-record writes live in prompts, notebooks, or scripts. Switching the base URL is easy; reconstructing that logic is the expensive part.

That is why a durable workflow matters more than a permanent model choice. Put the routing rule, validation, retries, and writes in software that persists outside a chat session.

The Major take

The cost difference is real, but acting on it is harder when workflow logic lives in prompts and notebooks. When the model is load-bearing, switching means rebuilding.

Major is the enterprise platform where agents build the software they run on. An agent reasons through a workflow once and builds an app that runs it. The model call becomes one governed step with scoped credentials and audit logs, so routing classification to a lower-cost candidate and judgment to a frontier candidate is a configuration property of the app. The repeatable work, such as fetching, validating, formatting, and writing to a system of record, runs as deterministic code. The model still reasons where judgment is needed. Reason once. Run forever.

Major makes the decision reversible by putting workflow logic somewhere the model is not. You were never only choosing a model. You were choosing whether that choice would be expensive to change. Build the routing app on Major and keep the workflow stateful, token-efficient, and governable.

Related articles

Related articles

Frequently asked questions

Which AI is better than ChatGPT?
"Better" depends on the job. As a chatbot for everyday use, ChatGPT has the deeper product and wider tooling. As an API behind a product feature, DeepSeek has lower published frontier-tier rates and portable OpenAI and Anthropic formats. Decide by workload, then keep the routing rule reversible.
Is DeepSeek cheaper than ChatGPT?
On published API rates, often, but the gap depends on the tier. Fetched on 2026-09-07, deepseek-v4-pro lists $0.66 per 1M input tokens in discount hours versus gpt-6-astra at $10.00. At the budget tier, OpenAI Batch pricing can come in below DeepSeek's best published rate.
Can DeepSeek do coding?
Yes. The DeepSeek API supports tool calls, up to 128 function definitions, and JSON output through response_format, which are the building blocks a coding agent needs to call tools and return parseable results. This answer describes API capabilities rather than claiming a general coding-quality ranking.
Can you switch from OpenAI to DeepSeek without rewriting your code?
Usually. DeepSeek states that its API format is compatible with OpenAI and Anthropic. OpenAI-format requests use https://api.deepseek.com, while Anthropic-format requests use https://api.deepseek.com/anthropic. An existing client generally needs a base URL, API key, and model change, followed by traffic tests.
Is DeepSeek safe to use for business data?
There is no universal answer without procurement context. Ask where data is processed and stored, how long prompts and completions are retained, who the subprocessors are, and whether a zero-retention agreement exists. Self-hosting the MIT-licensed weights removes the vendor from the data path, but adds infrastructure and operational responsibility.