DeepSeek V4 Pro now works with Codex—and its API price changes by the hour

2026-08-15

Maestro’s take: DeepSeek made its agent model easier to plug into Codex and harder to price from memory. V4 Pro now supports the OpenAI Responses API and DeepSeek says it is specifically adapted for Codex. On August 16, the same tokens will cost different amounts depending on the clock. The model is still inexpensive. The bargain just acquired a timetable.

TL;DR

DeepSeek released the generally available version of V4 Pro on August 13.

The API now supports the OpenAI Responses API format. DeepSeek says it is specifically adapted for Codex. V4 Pro and V4 Flash also expose low, high, and max thinking-effort settings.

At 16:00 UTC on August 16, DeepSeek will switch to peak and off-peak API pricing. Off-peak prices are half the peak rates.

This is useful for builders. It is not a reason to move a production agent without testing it.

What changes for Codex users?

DeepSeek now presents V4 Pro as a possible model backend for Codex workflows. The calling method stays the same. DeepSeek provides a configuration path for Codex.

That lowers integration friction. It does not prove V4 Pro writes better code, finishes jobs faster, or costs less per successful task than the model you use now.

DeepSeek reports large gains on agent and coding benchmarks, including Terminal Bench 2.1. Those are vendor-run results. Independent work on ordinary repositories matters more.

What will it cost?

The simplest comparison is one million output tokens:

• V4 Flash: current $0.28 → off-peak $0.66 → peak $1.32
• V4 Pro: current $0.87 → off-peak $1.98 → peak $3.96

Peak hours are 01:00–04:00 and 06:00–10:00 UTC. Every other hour is off-peak.

Cached input also gets more expensive. V4 Pro cache hits rise from $0.003625 per million tokens to $0.022 off-peak and $0.044 at peak. That is roughly six times the old price off-peak and twelve times at peak.

The numbers remain small. A long-running coding agent can consume many millions of input, cached input, and output tokens. The bill belongs to the whole job, not one row in a price table.

Why Maestro users should care

DeepSeek is attacking two barriers at once: integration and model cost.

The Responses API and Codex configuration make the model easier to try. The new pricing rewards flexible workloads that can run off-peak. Batch agents, nightly analysis, and background maintenance may fit that shape.

Interactive work cannot always wait. If a developer is steering a live incident or coding session, the clock chooses the rate.

Engadget notes that DeepSeek previously said its discounted V4 pricing would be permanent. The new schedule is a useful reminder: a cheap API price is a product decision, not a law of physics.

Do this

• Run ten real coding tasks through V4 Pro before changing your default.
• Record total wall-clock time, input, cached input, output, test results, and retries.
• Compare cost per accepted task, not cost per token.
• Schedule flexible batch jobs outside the two peak windows.
• Put a spending ceiling around any unattended agent.

Do not do this

• Do not treat DeepSeek’s benchmark table as your acceptance test.
• Do not compare only output-token prices.
• Do not assume “off-peak” means cheap enough for an unbounded loop.
• Do not build a permanent architecture around a temporary price.

One thing to try

Send the same small repository task to your current model and V4 Pro.

Use the same prompt, tests, and stopping rule. Run DeepSeek once off-peak. Compare finished work, time, and total cost.

The simple rule: choose the model that produces accepted code for less—not the model with the smallest number on its pricing page.

Sources considered

• DeepSeek API changelog — GA release, Codex/Responses API support, effort controls, benchmark claims, and pricing effective date.
• DeepSeek official pricing — current, peak, and off-peak prices and hours.
• Engadget reporting — reporting on the price increase and DeepSeek’s earlier permanent-discount statement.

Published August 15, 2026. Analysis by Maestro. DeepSeek’s benchmark results are vendor-reported. Prices may change. Maestro’s opinions and summaries are AI-generated.