DeepSeek V4 Pro Gains Benchmarks Then Raises API Prices Sharply

DeepSeek put the full DeepSeek-V4-Pro-0813 model live on its web app, mobile client and API on August 13, pairing a jump in agent-oriented benchmarks, an open-source coding-agent harness, and API price increases of 50 percent to more than 1,100 percent that took effect August 16.

The package improves real-world software-engineering scores and hands developers a free alternative to closed agent environments, yet the new peak and off-peak schedule ends the era of near-flat ultra-low rates that defined earlier DeepSeek releases.

What Arrived With V4 Pro 0813

DeepSeek describes V4 Pro as aimed at autonomous AI agent tasks and software engineering. The company says it approaches the level of Moonshot AI’s Kimi K3 while remaining cheaper to run at scale. The DeepSeek-V4-Pro GA release note lists flexible reasoning-effort settings (low, high, max), native OpenAI Responses API support, and one-click Codex setup.

V4 Pro is selectable in Expert Mode on the consumer app and web interface. On the API the model name stays deepseek-v4-pro. Context length reaches 1 million tokens and maximum output hits 384K tokens. Both thinking and non-thinking modes are supported.

• Full GA on app, web and API after the April preview
• Reasoning effort controls for simple versus complex agent work
• OpenAI Responses API plus Anthropic-compatible endpoint
• Companion open-source DeepSeek Harness (dsh) v0.1 under MIT
• Peak and off-peak billing live from 16:00 UTC August 16

The same day DeepSeek also published the first public cut of its agent runtime.

Self-Reported Gains Meet Independent Scores

DeepSeek reports V4 Pro scored 62.7 on the DeepSWE benchmark for real-world software engineering, up from 12.8 in earlier testing, and 87.9 on Terminal Bench 2.1. Those figures come from the company’s own tables and were measured with its Harness in some cases.

Independent firm Artificial Analysis placed the reasoning max-effort version at an Artificial Analysis Intelligence Index score of 53. That sits well above the median of comparable models (27) yet trails Moonshot’s Kimi K3 near 60 and Anthropic’s Claude Opus 5 near 63. On Terminal-Bench v2.1 the independent run landed around 79 percent, lower than the self-reported 87.9. Flash 0731 scored 52 on the same index, only one point behind Pro despite far lower active parameters.

AA also notes strong gains on agentic knowledge-work Elo (GDPval-AA) versus the April release. Speed sits above average at roughly 78 output tokens per second. Cost per Intelligence Index task lands in an attractive quadrant relative to closed peers even after the price reset.

Some developers on X and forums observed that Flash often matches or exceeds the older Pro preview on agent loops, raising questions about when the Pro premium is justified.

The New Peak and Off-Peak Rate Card

DeepSeek moved from flat rates to time-based pricing. Off-peak rates are half of peak. Peak windows are 01:00-04:00 and 06:00-10:00 UTC; all other hours are off-peak. The change took effect at 16:00 UTC on August 16, 2026.

According to the official peak and off-peak rate card, V4-Pro pricing per million tokens is now:

Flash sits lower: cache-hit input $0.007/$0.014, cache-miss $0.22/$0.44, output $0.66/$1.32. Concurrency limits are 500 for Pro and 2,500 for Flash. Cache-hit input on Pro rose by roughly 500-1,100 percent from prior levels near $0.0036, the sharpest single change. Output roughly doubled to quadrupled depending on the window.

Even at peak, Pro output at $3.96 per million remains far below many Western list prices that reach $15-$50. OpenRouter and other routers still list DeepSeek among the lower-cost high-capability options once peak logic is applied.

Who Pays More and Who Schedules Around It

Startups and teams running continuous agent fleets feel the cache-hit and peak jumps first. Long conversations and repetitive tool loops that once enjoyed deep cache discounts now cost substantially more. Batch jobs can shift into off-peak windows and cut the bill in half.

Cost snapshot after August 16

• Pro peak output: $3.96 per 1M tokens
• Pro off-peak output: $1.98 per 1M tokens
• Flash peak output: $1.32 per 1M tokens
• Cache-hit Pro increase: up to roughly 1,100 percent versus prior flat rate

Enterprise users comparing per-task cost against capability still see DeepSeek undercutting Kimi K3 (around $3 input / $15 output) and Claude Opus-class models. The practical decision becomes whether a few points on the Intelligence Index justify staying on a more expensive closed model or whether scheduling discipline plus Harness keeps total spend lower on DeepSeek.

Reuters and others framed the move as a test of whether a premium tier can support a sustainable business after years of aggressive discounting. The long-term rate path remains open.

DeepSeek Harness Opens the Agent Layer

Alongside the model, DeepSeek released DeepSeek Harness (dsh) v0.1 in developer preview under the MIT license. The MIT-licensed DeepSeek Harness repository describes an architecture in which everything is a plugin: models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and UI.

It is built on Cordis, DeepSeek’s composable plugin runtime. A single command can spin up a local web UI. The project ships architecture docs and an AGENTS.md written so other agents can read the codebase. DeepSeek positions it as model-agnostic infrastructure rather than a full drop-in product replacement for Claude Code or Codex.

The official Harness developer preview post drew heavy engagement. One widely shared take on X put it bluntly: a frontier lab gave away, for free, the exact layer companies charge a subscription for, noting Claude Code’s top plan at $200 per month. Developers immediately began writing custom plugins and swapping models, including non-DeepSeek backends.

A frontier lab just gave away, for free, the exact layer companies charge a subscription for.

That line, from a high-engagement summary post, captured the mood among agent builders. Early users report lower friction for local file-system agents and multi-step coding loops, though the preview status means breaking changes are expected.

From Discount Leader to Premium Tier Experiment

Earlier DeepSeek releases built reputation on extreme price-performance. V4 Pro keeps absolute costs low by frontier standards, yet the magnitude of the cache and peak adjustments marks a clear shift. The company is testing whether users will accept higher unit prices once capability and tooling improve.

Independent benches show Pro is no longer an automatic free win on every agent metric; Flash sits close enough that many workloads can stay on the cheaper tier. The open Harness lowers switching costs in the other direction: teams can keep the runtime and change the model underneath.

For large-scale autonomous coding and knowledge-work agents, the new math is capability per dollar after scheduling, not headline token price alone. Peak windows force operational changes. Off-peak batching and cache discipline become first-class engineering concerns.

Availability and Immediate Next Steps

DeepSeek-V4-Pro-0813 is live now through the web interface, mobile app and API. New peak and off-peak rates have been active since 16:00 UTC on August 16. Weights are listed as open under MIT on Hugging Face according to Artificial Analysis. Harness remains in developer preview on GitHub with active commits.

Teams evaluating the stack can compare a slightly lower Intelligence Index score against multi-fold lower token costs, factor the August price change into budgets, test Harness against existing Claude Code or Codex workflows, and shift non-urgent jobs into off-peak hours. The combination of stronger agent scores, a free open runtime, and deliberately higher API prices is the package DeepSeek chose to ship.

Frequently Asked Questions

What is DeepSeek Harness and how does it differ from Claude Code?

DeepSeek Harness (dsh) is an MIT-licensed open-source agent runtime released in v0.1 developer preview. It treats models, tools, sessions, sandboxes and even the agent loop as swappable plugins on the Cordis framework, and it is model-agnostic. Claude Code is a closed, subscription product tightly integrated with Anthropic models; Harness supplies the orchestration layer without locking the model or charging a monthly fee.

When did the new DeepSeek API peak pricing take effect?

The peak and off-peak schedule became active at 16:00 UTC on August 16, 2026. Peak hours are fixed at 01:00-04:00 and 06:00-10:00 UTC; every other hour uses the lower off-peak rates, which are half the peak rates.

How does V4 Pro score on independent versus company benchmarks?

Artificial Analysis gives the max-effort reasoning version an Intelligence Index of 53. DeepSeek’s own figures list 62.7 on DeepSWE and 87.9 on Terminal Bench 2.1. Independent Terminal-Bench runs have come in lower, around 79 percent, illustrating the gap that often appears between provider-reported and third-party agent evaluations.

Is DeepSeek V4 Pro still cheaper than Kimi K3 or Claude after the hikes?

Yes on list rates. Peak Pro output is $3.96 per million tokens and off-peak $1.98, while Kimi K3 output is commonly listed near $15 and Claude Opus-class models higher still. Input and cache pricing remain competitive once peak windows are avoided, though the relative advantage narrowed sharply on cache hits.

Where can developers run DeepSeek V4 Pro right now?

The model is available on DeepSeek’s official web interface and mobile app in Expert Mode, via the DeepSeek API (OpenAI- and Anthropic-compatible endpoints), and through third-party routers such as OpenRouter. Open weights are published for local or self-hosted use under an MIT license according to model cards.