DeepSeek V4 Pro Is on Writingmate: Testing Agentic Coding and Tool Use

DeepSeek V4 Pro has shown up as a comparison baseline in three of our last several model posts — the Aion-3.0 roleplay test, the pricing table in our Kimi K3 writeup — and it has never once gotten a review of its own. That's a strange gap for a 1.6-trillion-parameter open-weight model that undercuts Claude Sonnet 5 on output price by roughly 17x and keeps showing up as the thing other models get measured against on cost.

My name is Artem, and I run the Writingmate blog. I've been running the same three-task agentic coding suite — a broken function, a messy module, and a multi-step job that needs real tool calls — on every model we've reviewed since Kimi K3 and Grok 4.5. DeepSeek V4 Pro kept beating those models on price in the tables without ever sitting in the hot seat itself, so this time I gave it the exact same three tasks and a stopwatch.

Here's a hands-on look at DeepSeek V4 Pro running through an OpenCode-style agentic workflow, which is close to what I put it through myself:

What Changed From DeepSeek V3.2 to V4 Pro

DeepSeek shipped V4 Pro and its smaller sibling V4 Flash on April 24, 2026, and the framing in the official release notes is specifically about agentic work, not just raw benchmark scores. The docs call out "Enhanced Agentic Capabilities" and open-source leadership on agentic coding benchmarks, plus a detail that matters more than it sounds like: V4 Pro now preserves its reasoning history across tool-result rounds and across new user messages. V3.2 discarded the reasoning trace the moment a new user message arrived, which meant every follow-up in a tool-calling loop started the model's thinking from zero.

That's the kind of change that doesn't show up on a benchmark leaderboard but shows up immediately the second you put a model in an actual agent loop with five or six sequential tool calls. The other headline number is efficiency: DeepSeek says that in the 1M-token context setting, V4 Pro needs only 27% of the single-token inference FLOPs and 10% of the KV cache that V3.2 required for the same job — which is a big part of why the API pricing dropped as far as it did.

"DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length. DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models." — DeepSeek on X

V4 Pro also ships with dual operating modes — Thinking and Non-Thinking — and the reasoning effort is optional rather than mandatory, with "high" and "xhigh" as the two selectable levels. That puts it in different territory than Kimi K3, whose reasoning is locked on at max by default. You can turn V4 Pro's reasoning off entirely for quick, low-stakes calls and turn it up for anything that needs to survive a long tool-calling loop.

DeepSeek V4 Pro Specs at a Glance

Field	DeepSeek V4 Pro	Reader takeaway
Provider	DeepSeek	Released April 24, 2026, replacing deepseek-chat and deepseek-reasoner, which retired July 24, 2026
Architecture	1.6T total parameters, 49B active per token (MoE)	About 3% of parameters active per token — cheap to run relative to its total size
Context window	1,048,576 tokens (1M), up to 384,000 output tokens	Full-codebase analysis and long tool-calling sessions fit in one context
Input / output modalities	Text in, text out	No vision or audio — a text-and-code specialist, not a multimodal generalist
Reasoning	Optional, "high" or "xhigh" effort, default "high"	You can dial reasoning off for cheap calls or up for long agent loops
Tool use	Native function calling; reasoning trace persists across tool-result rounds	Fixes V3.2's biggest agentic weak spot — no more restarting the thought process each turn
Pricing (OpenRouter)	$0.435 input / $0.87 output per 1M tokens; $0.0036 cached input	Roughly a 120x discount on cached input tokens

How I Tested Agentic Coding and Tool Use

I ran DeepSeek V4 Pro through the same three tasks we've used on every coding-focused model release this quarter, comparing it against Claude Sonnet 5 — our current daily-driver pick from the coding assistant roundup — as the baseline. Same prompts, same fresh conversations, same default reasoning setting, so the results would actually be comparable.

• Task 1 — broken function fix: a Flask session handler throwing an intermittent KeyError: 'user_id', with the full traceback and the three files it touched. The real cause is a race condition, not a typo.
• Task 2 — messy module refactor: a 140-line Express route handler doing authentication, validation, a database call, and a confirmation email inline, three levels of nested callbacks deep. The ask: split it into small, testable functions without changing behavior.
• Task 3 — multi-step agentic task: add a rate limiter to an API, write unit tests for it, run the suite, read the failures, and fix them — a real loop with sequential tool calls, not a single reply.

Task 1: Fixing a Broken Function From a Stack Trace

DeepSeek V4 Pro found the actual race condition — the background job clearing the session dict before the request handler finished reading it — on the first pass, and correctly rejected the lazy fix of wrapping the dictionary access in a try/except with a default value. That's the same root cause both Claude Sonnet 5 and GPT-5.6 Sol Pro caught in our original roundup, and V4 Pro got there without a second prompt.

Where it differed from Claude Sonnet 5 was verbosity. With reasoning left on its "high" default, V4 Pro produced several paragraphs of visible thinking before the diff, similar to what we saw from Kimi K3's always-on reasoning. It's not wrong, just slower to scan — if you're paying by the token or reviewing a lot of these in a row, dropping reasoning effort for straightforward bug reports is worth testing before you commit to "high" as your default.

Task 2: Refactoring a Messy Module

On the Express handler, V4 Pro split the code into five functions and kept the file's existing error-handling style intact — the same bar Claude Sonnet 5 cleared in our original test, and one that Qwen3 Coder Plus and Gemini 3.1 Pro Preview both missed on their first pass. It also caught the same detail Claude did: the confirmation email step needed to be async-safe rather than blocking the response, and flagged it as a comment in the diff instead of silently changing the behavior.

It wasn't flawless. V4 Pro introduced one small helper function that duplicated logic already available elsewhere in the file I'd given it for context — a minor style miss, not a correctness bug, but the kind of thing a reviewer would flag in a pull request.

Task 3: A Multi-Step Agentic Task With Tool Use

This is where the reasoning-persistence change from V3.2 actually mattered. V4 Pro wrote the rate limiter, wrote a test that specifically exercised burst requests at the limit boundary, ran the suite, read its own failure output, and fixed an off-by-one error in the limiter — all without me re-explaining the task on the second tool-result round. That's a direct result of the model no longer discarding its reasoning trace between turns.

The single most useful thing I saw across all three tasks was that V4 Pro's second and third tool calls referenced specifics from its own first response instead of re-deriving them. In a long agent loop, that's the difference between a model that stays coherent and one that slowly drifts off the original plan.

It needed one follow-up prompt to close out the loop cleanly, putting it roughly on par with Claude Sonnet 5's two follow-ups from the original roundup and ahead of Gemini 3.1 Pro Preview and Qwen3 Coder Plus, which both needed three or more. Independent reviews report the same general pattern with a caveat worth repeating here: on longer, more autonomous agent runs than the one I tested, Claude Opus-tier models still recover from mid-loop errors more reliably than V4 Pro does. My three-task suite didn't push it that far, but it's the honest limit worth knowing before you hand V4 Pro a fully unattended, hours-long agent job.

DeepSeek V4 Pro vs. Claude Sonnet 5, GPT-5.6 Sol Pro, and Kimi K3

Independent benchmarks back up what I saw in testing: DeepSeek V4 Pro scores 80.6% on SWE-bench Verified, 93.5% on LiveCodeBench, and 67.9% on Terminal-Bench 2.0, with a Codeforces rating of 3,206. Those numbers sit close to Claude Sonnet 5 and GPT-5.6 Sol Pro on individual coding tasks, even though V4 Pro costs a fraction of either on a per-token basis.

Model	SWE-bench Verified	Terminal-Bench	Input / Output ($ per 1M)	Context
DeepSeek V4 Pro	80.6%	67.9% (2.0)	$0.435 / $0.87	1M
Claude Sonnet 5	—	—	$3.00 / $15.00	1M
GPT-5.6 Sol Pro	—	88.8% (2.1)	$5.00 / $30.00	1M in / 128K out
Kimi K3	—	88.3% (2.1)	$3.00 / $15.00	1M

Terminal-Bench versions aren't directly comparable across the table (2.0 vs. 2.1 is a harder revision), so treat that column as directional rather than an exact ranking. What is directly comparable is price: V4 Pro's output pricing is roughly 17x cheaper than Claude Sonnet 5's and about 34x cheaper than GPT-5.6 Sol Pro's, while landing functionally close on the three tasks I ran myself.

Pricing: Why DeepSeek V4 Pro Undercuts Everyone

DeepSeek made its aggressive launch pricing permanent on May 22, 2026, turning what started as a 75%-off promotion into the standing rate: $0.435 per million input tokens and $0.87 per million output tokens, with cached input priced at $0.0036 per million — close to a 120x discount over the base input rate. For a long agentic session that re-sends the same system prompt and file context on every tool-result round, that cache discount compounds fast.

"We are making our discount permanent! Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life!" — DeepSeek on X

On Writingmate, none of this matters as a per-token bill — DeepSeek V4 Pro is included in paid plans alongside 200+ other models, so the pricing table above is mainly useful for judging value relative to the rest of the field, not for budgeting your own invoice.

How to Try DeepSeek V4 Pro on Writingmate

DeepSeek V4 Pro is live in the model picker now. Open a new chat, click the model dropdown, and search for "DeepSeek V4 Pro," or start from the DeepSeek V4 Pro model page directly.

The fastest way to see whether it actually beats your current default is to run your own backlog through it rather than trust this write-up. Open the DeepSeek V4 Pro vs. Claude Sonnet 5 comparison page and paste in a real stack trace or a real messy file — that's a much better test than a demo prompt, and you can rerun it every time a new model shows up in the Writingmate models catalog.

If you're wiring V4 Pro into an existing agent framework instead of testing it in chat, the OpenAI-compatible API docs cover how to point tools like OpenCode, Continue, or a custom agent loop at it using the same request format you'd use for any other model in the catalog.

Best Use Cases for DeepSeek V4 Pro

• High-volume coding work where the per-token cost of a flagship model doesn't pencil out — bug fixes, scoped refactors, test generation at scale
• Multi-step agent loops with sequential tool calls, since the reasoning trace now survives across tool-result rounds instead of resetting each turn
• Full-codebase analysis and long-document work that needs the 1M-token context window without paying flagship prices for it
• Teams that want to dial reasoning effort up for judgment-heavy tasks and down for routine ones, rather than paying always-on reasoning tax on everything

Where I'd still reach for something else: fully unattended, hours-long agent runs where error recovery matters more than any single task's correctness, and anything that needs vision or audio input, since V4 Pro is text-only.

Bottom Line on DeepSeek V4 Pro

DeepSeek V4 Pro earned its comparison-baseline reputation honestly. It matched Claude Sonnet 5's root-cause accuracy on a real debugging task, held its own on a refactor with only a minor style miss, and closed out a multi-step tool-calling loop in one follow-up thanks to the fix that actually matters most for agentic work — reasoning that survives between tool calls instead of resetting. None of that required paying Claude or GPT-5.6 Sol Pro pricing to get.

It's not a strict upgrade over either of those two on the hardest, longest agent loops, and the text-only input means it's the wrong pick the moment a task needs a screenshot or a voice note. But for the bulk of day-to-day coding work — the broken function, the messy module, the scoped agent job — DeepSeek V4 Pro belongs in your rotation, not just in someone else's comparison table.

See you in the next one!

Artem