DeepSeek V4 vs GPT-5.5 and GPT-6 for coding: what it costs

This post first ran in July 2026 as "DeepSeek V3 vs GPT-5 for coding." Both models have since been replaced. The original also carried a scorecard from a 30-task test with no published data behind it, so we removed it rather than keep numbers nobody can check.

This version sticks to what can be verified: the current models, their published prices, and how to run both through one endpoint so you can compare them on your own code.

What changed since the first version

• DeepSeek V3 is gone from DeepSeek's price list. It now has two models: deepseek-v4-pro, and deepseek-flash, which is served by V4.1 Flash. The older name deepseek-v4-flash is still accepted. Both have a 1M-token context window, up to 384K output tokens, and thinking mode on by default.
• DeepSeek no longer has one price. Peak hours cost twice the off-peak rate. Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, except Chinese public holidays.
• GPT-5 is no longer OpenAI's current model. OpenAI's price list now leads with the GPT-6 family: Astra, Sol and Luna. GPT-5.5 is still listed, and it is the newest model kRouter's Codex provider offers through a ChatGPT plan.

Prices per million tokens

Checked in October 2026 against DeepSeek's pricing page and OpenAI's pricing page. OpenAI's rates are for prompts up to 272K input tokens. Longer prompts cost more; GPT-5.5 goes to $10 in and $45 out.

Three things stand out.

At the top end, DeepSeek is still far cheaper. V4 Pro's output costs about 15x less than GPT-5.5 off-peak, and about 7.5x less at peak. Against GPT-6 Sol the gap is about 5x off-peak and 2.5x at peak.

At the bottom end, the old story has flipped. GPT-6 Luna lists below DeepSeek Flash on uncached input and on output, even at DeepSeek's off-peak rates. Only cached input is still cheaper at DeepSeek. If you picked DeepSeek only because it was the cheapest option, compare it with Luna first.

For agents, read the cached-input column. Claude Code, Codex and similar tools resend the conversation on every turn, so a long session is mostly input the provider has seen before. When the provider's prompt cache catches it, that column can matter more than the headline input price. And if you work during DeepSeek's peak windows, budget at the peak rows.

What about quality?

We are not publishing a scorecard. Each vendor reports benchmark results from its own harness and prompts, so a DeepSeek number and an OpenAI number rarely measure the same thing, and neither measures your codebase. The cheap way to find out is to run the same tasks through both.

kRouter turns that into a one-word change, because the client stays the same and only the model name moves:

1. Connect DeepSeek with an API key, plus whichever GPT route you have (see the next section).
2. Pick ten or so tasks you have already solved, ideally ones with tests: a bug fix, a refactor across several files, a migration.
3. Run each task once per model from a clean branch, changing nothing but the model name.
4. Score each run on two questions: do the tests pass, and would you merge the diff? The Usage page in the dashboard has a by-model view of the tokens each model used. Work out the cost from the current price list, since DeepSeek's rate depends on the hour.

Reaching each model through kRouter

A few notes:

• GPT-6 Luna on OpenCode Go works only through OpenCode's Responses API. kRouter converts to and from it, so Claude Code, Codex and Chat Completions clients all work, from kRouter 0.5.163 on. OpenCode Go's three protocols explains why.
• Codex is a sign-in with your ChatGPT account. kRouter shows a risk notice before you connect it: the subscription is not licensed for proxy use, and the account may be restricted.
• DeepSeek also runs an Anthropic-compatible API of its own. If DeepSeek is the only model you want in Claude Code, you can point Claude Code straight at it. A router earns its place when you want a second provider to take over, or GPT in the same session.

Using both without picking one

There are two ways, depending on whether kRouter decides or you do.

A fallback combo keeps you working when one provider fails. Create it on the Combos page:

Combo: "coding"
  1. ds/deepseek-v4-pro
  2. ocg/deepseek-v4-pro
  3. cx/gpt-5.5

Requests go to DeepSeek's API first. If that key is rate limited or failing, kRouter cools it down, tries any other DeepSeek keys you added, then moves to the next entry. A combo does not judge how hard a request is, and kRouter has no spend cap, so a combo will not save GPT for the difficult tasks on its own. The first version of this post said it would; that was wrong. Combos and fallback lists exactly when kRouter moves on.

Slot mapping lets you decide. Claude Code has three model slots, and the CLI Tools page can point each one at a different model. Written by hand, it looks like this:

export ANTHROPIC_BASE_URL=http://localhost:20128/v1
export ANTHROPIC_AUTH_TOKEN=<your-krouter-key>
export ANTHROPIC_DEFAULT_OPUS_MODEL=cx/gpt-5.5
export ANTHROPIC_DEFAULT_SONNET_MODEL=ds/deepseek-v4-pro
export ANTHROPIC_DEFAULT_HAIKU_MODEL=ds/deepseek-v4-flash

Work in the Sonnet slot runs on DeepSeek, and GPT-5.5 only runs when you switch to the Opus slot with /model opus. A slot can also hold a combo name, such as coding above. Claude Code on Kimi, GLM and DeepSeek covers which open model holds up in which slot.

Any OpenAI-compatible client works the same way:

export OPENAI_BASE_URL=http://localhost:20128/v1
export OPENAI_API_KEY=<your-krouter-key>

API keys are on the dashboard's Endpoint page. To get started:

npm install -g @sifxprime/krouter
krouter

Common questions

Is DeepSeek V4 cheaper than GPT for coding?

Against GPT-5.5 and GPT-6 Sol, yes. DeepSeek V4 Pro's output costs about 15x less than GPT-5.5 off-peak, and still about 2.5x less than GPT-6 Sol at DeepSeek's peak rate. At the low end it is not: GPT-6 Luna lists below DeepSeek Flash on uncached input and on output. These are list prices as of October 2026.

When does DeepSeek charge peak rates?

From 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. Peak costs twice the off-peak rate. Weekends are off-peak all day.

Can I use DeepSeek V4 in Claude Code or Codex?

Yes. DeepSeek's own Anthropic-compatible API works with Claude Code directly. Through kRouter you can also use it from Codex, because kRouter converts Codex's Responses requests for DeepSeek, and you can fall back to another provider when DeepSeek is unavailable.