Skip to main content

DeepSeek V4 Flash 0731: Post-Training Results, Price and Agent Integration

MODELS & COSTS

DeepSeek V4 Flash 0731: Post-Training Results, Price and Agent Integration

Open Source Factory · Aug 02, 2026 · 3 min read

October 3, 2026 update: The 0731 build is historical. Subsequent Flash releases changed model aliases, available weights and pricing. Read the current API history before following this article’s old configuration or price advice.

A benchmark gain is a reason to test an agent endpoint, not the end of the integration work. DeepSeek's 0731 table reports a post-training improvement. Tool schemas, failed calls and the complete successful-task bill still need checking in the harness that will use the model.

In this article
  1. The 0731 release
  2. DeepSeek's reported benchmark results
  3. July launch rates
  4. Agent endpoint compatibility
  5. Evaluation and deployment caveats
  6. Testing the complete workflow
  7. Linked source documents

The architecture remains the 284B model, while the training recipe changed. The announcement attributes the agent gains to that post-training update. Reproduction needs the same task versions and harness, particularly when comparing Flash with a different release of Pro.

Preview vs 0731 comparison chart
April Preview vs July official. DeepSWE went 7.3 → 54.4 (+645%)... yes, same model.

The 0731 release

The model is deepseek-v4-flash (0731 checkpoint). Specs:

· Parameters: 284B total / 13B active (MoE)
· Context: 1M tokens / up to 384K output
· Concurrent requests: 2,500 (Pro is 500)
· Thinking mode ON by default (can be disabled)
· Native Responses API → plugs into Codex directly
· No weights on HuggingFace yet — API only

Side note: the old deepseek-chat / deepseek-reasoner aliases died on July 24. If you're still on those, you're probably already getting errors.

DeepSeek's reported benchmark results

Here are the numbers. But all of this was measured by DeepSeek on their own harness — zero third-party reproductions so far. Treat it as a "direction," not a "confirmed ranking."

DeepSeek official benchmark table
The original table DeepSeek published. Beats V4 Pro Preview on every single row.
Agent benchmark comparison chart
My comparison chart. Blue is Flash 0731, orange is Opus 4.8.

Two numbers caught my eye personally:

Terminal Bench 2.1: 82.7 — Opus 4.8 scores 85.0, so the gap is 2.3 points. The price gap is 89×.
DeepSWE: 7.3 → 54.4 — +645% over the Preview. If training alone does that, it's a pretty meaningful result for the post-training research crowd.

July launch rates

The launch rate card below is a separate comparison from the benchmark table. Rates are per million tokens and describe the July release; later billing changes need the subsequent pricing notice.

V4 Flash: input $0.14 / cache hit $0.0028 / output $0.28
V4 Pro: input $0.435 / output $0.87 (3.1× Flash)
GPT-5.6 Luna: input $0.20 / output $1.20
Gemini 3.1 Pro: input $2.00 / output $12.00
Claude Opus 5: input $5.00 / output $25.00 ← 89× Flash
Price comparison chart by model
Log scale, so the gap feels smaller than it is — Opus output is 89× Flash.

What does a cache hit of $0.0028 mean? Repeated prefixes (like your system prompt) get a 98% discount. If your workload keeps reusing the same context — agent loops, for example — your effective input cost approaches zero.

※ Peak-time 2× surcharge (Beijing 09–12 and 14–18) was announced but the start date is TBD. Prices above are current.

Agent endpoint compatibility

This one surprised me: native Responses API support. Not through an OpenAI-compatible shim — directly. Codex CLI, ChatGPT desktop, and the VS Code extension all hook up with a single config line.

Right now, Flash is the only DeepSeek model with official Codex support. Pro was supposed to follow in early August, but no news yet.

Evaluation and deployment caveats

Benchmarks = self-measured. There's a semi-independent number — Artificial Analysis Intelligence Index 50 (median 17) — but that measures general intelligence, not agent performance.
The output is verbose. Per AA measurements, it burns 3.4× the median tokens. Cheap per token doesn't necessarily mean cheap per task. Judge by "cost per completed task," not "token price."
Thinking mode is ON by default. Reasoning tokens are billed at output rates — money leaves even when you don't see them. Turn it off for simple tasks.
No weights. The 0731 build is API-only. If you want local, the only option is the April Preview's MIT weights.

Testing the complete workflow

For an existing agent harness, test endpoint compatibility and tool calls before changing the default model. A short configuration change can connect an endpoint, but it does not establish successful execution of the complete workflow.

A useful comparison would measure completed tasks, retries and cost against the current model. DeepSeek's launch table supplies candidates for that test, while the later rate card matters for the budget. Neither alone establishes an Opus replacement.

Related on this blog: DeepSeek adding vision to Flash · the V4 Flash vision build

Sources: DeepSeek API Changelog (2026.07.31) · Artificial Analysis · BenchLM.ai · Digital Applied
Open Source Factory · opensourcefactory1.blogspot.com

Linked source documents

Comments

Post a Comment