Models & Industry

DeepSeek V4 Flash for Coding: Price, Benchmarks, and Tool Integration

Review DeepSeek V4 Flash model IDs, pricing, benchmark modes, Claude Code and OpenCode setup, compatibility issues, and a fair test method.

BetterToken

Editorial Team

Aug 5, 20268 min read

DeepSeek V4 Flash is worth testing as a low-cost model for frequent coding calls, subagents, and tasks with a clear definition of done. A low price per token does not remove the cost of long reasoning, client errors, and reruns. For heavy architectural work, compare Flash with Pro on the same repository instead of treating it as a “cheap copy” of the flagship.

Want to test Flash on a small budget? In the BetterToken catalog on August 13, 2026, deepseek-v4-flash-0731 cost about $0.095 per million input tokens and$ 0.191 per million output tokens. Create your own API key, run one safe test, and check the Model ID and usage in Workspace. Recheck the catalog before the next run: Model IDs and prices depend on the selected provider and may change.

Which DeepSeek V4 Flash is being discussed?

The official April 24 announcement introduced V4 Preview. On August 13, the DeepSeek model page showed two stable API Model IDs: deepseek-v4-flash and deepseek-v4-pro. Their MODEL VERSION entries were DeepSeek-V4-Flash-0731 and DeepSeek-V4-Pro-0813.

The version and Model ID serve different purposes. The version tells you which build DeepSeek currently serves, while the stable Model ID remains the string used by the official API. Another provider may use different catalog names: on August 13, BetterToken listed deepseek-v4-flash-0731 and deepseek-v4-pro, but not deepseek-v4-pro-0813. Open the selected endpoint's catalog and copy the name from there; do not add 0813 to the configuration automatically.

Where you call the model	Flash	Pro	What to copy into configuration
Official DeepSeek API	version `DeepSeek-V4-Flash-0731`	version `DeepSeek-V4-Pro-0813`	`deepseek-v4-flash` or `deepseek-v4-pro`
BetterToken, checked August 13	`deepseek-v4-flash-0731`	`deepseek-v4-pro`	Model ID from the current BetterToken catalog

The 0731 and 0813 numbers indicate the age of the served build, but they do not define a universal alias across endpoints.

According to the documentation and official model card, Flash has 284 billion total and 13 billion active parameters, a one-million-token context, and a maximum output of 384K. Thinking and non-thinking modes are available, with High and Max reasoning levels for comparison.

A million-token context is a technical ceiling, not a promise of equal quality for every code volume. Test whether the model can retrieve information from the middle of the context and preserve project rules after a long tool history.

What one agent run costs

Official DeepSeek API pricing:

Input, cache hit: $0.0028 per 1 million tokens.
Input, cache miss: $0.14 per 1 million tokens.
Output: $0.28 per 1 million tokens.

A run with 200,000 input and 20,000 output tokens without a cache:

0.2 × $0.14 + 0.02 × $0.28 = $0.0336.

If the entire input hits the cache, the calculation is $0.00616. A complete hit is a theoretical edge case; a real agent changes history, tool results, and files, so parts of the prefix stop matching.

Using the BetterToken catalog rates, the same volume without separate cache accounting costs:

0.2 × $0.095 + 0.02 × $0.191 = $0.02282.

The low total should not hide reasoning usage. If the client requests a large max_tokens, the model may spend a substantial part of the budget before producing a useful answer. For a financial comparison, record not just the price per million but the tokens consumed per completed task.

What changes between Non-Think, High, and Max

The official model card compares three modes:

LiveCodeBench: 55.2 in Non-Think, 88.4 in High, and 91.6 in Max.
Terminal Bench 2.0: 49.1 in Non-Think, 56.6 in High, and 56.9 in Max.
SWE Verified: 73.7 in Non-Think, 78.6 in High, and 79.0 in Max.
SWE Pro: 49.1 in Non-Think, 52.3 in High, and 52.6 in Max.
MRCR 1M: 37.5 in Non-Think, 76.9 in High, and 78.7 in Max.

These are vendor/model-card results, not an independent production benchmark. They show two useful patterns. First, disabling thinking sharply reduces scores on some coding and long-context tests. Second, moving from High to Max adds much less than moving from Non-Think to High.

A practical choice:

Non-Think — formatting, extraction, or an obvious local edit after verification on your examples;
High — the primary candidate for bug fixes, review, and several related files;
Max — difficult debugging, a long agent plan, or a task where an error costs more than additional tokens.

Do not make Max the default solely because it has the largest number in the table.

Workloads where Flash looks rational

Flash is especially interesting when there are many requests and each step can be checked automatically:

a subagent locates files and gathers facts for the main agent;
CI explains a specific test failure;
a batch review checks one rule;
an agent generates tests for an already defined interface;
a code migration is split into short, identical batches.

It is a less obvious choice for new architecture without clear constraints, conflicting requirements, or work where the model must make product decisions independently for hours. In those cases, retries may erase Flash's price advantage, and Pro or another model may cost less per accepted result.

Connecting to Claude Code

DeepSeek provides an Anthropic-compatible endpoint and splits model roles in its official example: Pro is the main agent, while Flash serves as Haiku and the subagent.


export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="YOUR_API_KEY"
export ANTHROPIC_MODEL="deepseek-v4-pro[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-v4-flash"
export CLAUDE_CODE_SUBAGENT_MODEL="deepseek-v4-flash"
export CLAUDE_CODE_EFFORT_LEVEL="max"

This split is useful: the more expensive model makes architectural decisions, while Flash handles narrower subtasks. If you want to test Flash as the main agent, do not present that as the officially recommended configuration. Create a separate profile and compare the result.

After launch, verify the Base URL, actual subagent model, and one tool call. A text response alone is insufficient; compatibility problems often appear only after the first tool invocation.

Connecting to OpenCode

The official guide requires OpenCode 1.14.24 or later:


opencode --version
opencode

In the interface, run /connect, select DeepSeek, and paste the key in the dialog. Then open the model list. The official guide demonstrates Pro; select Flash only if deepseek-v4-flash is actually present under the official provider. BetterToken used deepseek-v4-flash-0731 at the time of verification. Do not swap these names without checking the specific endpoint's catalog.

The test should have several steps: ask it to read a file, call a tool, then continue reasoning after the tool result. That second turn reveals reasoning_content problems.

Why the reasoning_content error occurs

In thinking mode, DeepSeek requires reasoning_content to be passed back in the history of later requests. OpenCode issue #24130 describes a client that lost this field and failed after a tool call; the related fix was discussed in PR #24146. The current official DeepSeek guide separately requires OpenCode 1.14.24 or later. Do not infer a fixed version solely from an issue or PR; check the current guide.

If the error persists:

update OpenCode;
confirm that it uses the current DeepSeek provider;
do not remove reasoning fields while normalizing messages yourself;
repeat a two-turn test with one tool;
only then inspect the Model ID and network.

Handwritten JSON from a random comment may help temporarily, but the official provider is safer because it is updated alongside the client contract.

Where the 32K limit comes from

OpenCode issue #29363 reported that the client configuration capped output at 32K even though the model documents 384K. These are different layers: the model limit and the limit the application sends in its request.

Do not request 384K automatically. A large ceiling increases potential cost and time. If a response stops with length, however, check the effective max_tokens, provider adapter, and reasoning mode before concluding that “Flash cannot produce a long response.”

What individual user reports tell you

In one practical coding discussion, users described both successful repository audits and short migrations, as well as missed requirements, plans that needed correction, and repeated errors in a small test task. Those accounts do not share a prompt, commit, harness, or independent verification. They supply scenarios for your own acceptance test, not an average performance measure.

A separate issue #1483 reports responses switching to Chinese for some V4 Flash requests. This is one user report, not a known defect rate. Add explicit checks for response and comment language, project-rule compliance, and recovery after a failed tool call.

How to test without fooling yourself

Compare Non-Think, High, and Max on ten identical tasks. Freeze the commit and do not change the prompt between modes. Measure:

the share of tasks that pass tests;
time to a completed diff;
input, output, and reasoning tokens;
number of retries;
rule and language violations;
cost per accepted task.

If High completes nine out of ten tasks and Max completes the same nine at twice the usage, Max does not pay off. If Max alone solves a difficult bug, its higher cost is justified for that class of work.

Sources

Ready to optimize your LLM workflow?

Join thousands of developers building faster, smarter, and more cost-effective AI applications with BetterToken.