DeepSeek V4 and the Price-Performance Question: What the Numbers Actually Show

On April 24, 2026, DeepSeek released V4: two models, V4 Pro (1.6 trillion total parameters, 49 billion active) and V4 Flash (284 billion total, 13 billion active), both Mixture-of-Experts models with a 1-million-token context window. A month later, DeepSeek made a promotional 75 percent price cut on V4 Pro permanent, and independent benchmarking firm Artificial Analysis ranked it among the best models in the world on "intelligence per dollar," ahead of the flagship models from OpenAI and Anthropic on that specific measure (South China Morning Post, May 24, 2026).

That headline is true and also easy to misread. "Cheap per token" and "cheap to get a task done" are different numbers, and the gap between them is where most of the real economics of choosing a model actually live. This is a look at what V4 costs, what it benchmarks at, and where the cost-per-token story stops telling the whole truth.

What actually shipped

Both V4 models use a hybrid attention design DeepSeek calls Compressed Sparse Attention and Heavily Compressed Attention, plus a structural change to residual connections (Manifold-Constrained Hyper-Connections) intended to keep signal propagation stable across a very deep network. Training used the Muon optimizer over more than 32 trillion tokens. The headline efficiency claim: at a 1-million-token context, V4 Pro needs only 27 percent of the single-token inference FLOPs and 10 percent of the KV cache that DeepSeek's own V3.2 required at the same context length (DeepSeek-V4 model card and technical report, Hugging Face, arXiv:2606.19348). That efficiency gain is the thing that makes the pricing below possible: a MoE model this large is only cheap to run if most of its 1.6 trillion parameters stay dormant on any given token, and if the attention mechanism doesn't blow up memory at long context.

One Agent's Opinion Is Still Just One Opinion

Isolation is the easy part. Give Codex and Claude Code each their own git worktree, a separate branch and checkout of the same repo, and their edits physically cannot collide while they work, because they're writing to different directories on disk. Anthropic ships this natively for Claude sessions and subagents; nothing does it natively across both providers, which is why cross-provider setups today are either two terminals you track by hand or, per MindStudio's comparison, a Codex plugin living inside Claude Code as a review layer rather than two independent sessions. Unstoppable does all of that in the background for you.

But isolation only solves the problem of two agents stepping on each other's files. It doesn't solve the harder problem: after both diffs land, whose read on the code do you trust? Ask Codex to review a change and you get Codex's opinion. Ask Claude Code and you get Claude's. Neither one is checking the other's blind spots. That's the gap pipelines are built to close.

Pipelines: Model Fusion Instead of One Agent's Opinion

Unstoppable's Code Review (Multi-Provider) pipeline doesn't ask one model for an opinion on a diff. It runs six review tracks in parallel, split across both providers: a Sonnet-based repo-compliance check and a GPT-based repo-compliance check (each reading every CLAUDE.md and AGENTS.md in scope and flagging only quotable, unambiguous rule violations), an Opus "methodical reviewer" and a GPT "methodical reviewer" doing a line-by-line pass, and two dedicated security tracks: one on Codex, built on OpenAI's own curated codex-security skill, and one on Claude/Opus.

Every track walks in with the same context already loaded: the scoped diff, the PR description, the linked Linear ticket, and the last round of human or AI review comments, so no track is guessing at intent. Every track is also told, explicitly, what not to report: no subjective concerns, no style nitpicks that aren't a documented CLAUDE.md/AGENTS.md rule, nothing a linter would already catch, nothing that requires a guess. If a track isn't certain a finding is real, it's instructed to drop it.

Then a final Opus pass reconciles all six tracks into one verdict: which findings multiple models independently flagged, which are single-source, and one risk score from 10 down. That score is what decides whether a human needs to look at all: critical and high scores require human review, medium recommends it, low makes it optional, and a 10 means no validated findings and the pipeline is telling you not to bother. That's not a hypothetical policy, it's the literal scoring rubric the pipeline runs on. The result posts as a single GitHub review with inline comments on the individual findings, and a rerun on the same PR dismisses the previous AI review instead of piling on.

That's the actual point of running two providers on the same repo: not "pick Codex or pick Claude," but never having to trust either one's review by itself. If you want to wire something like this into your own repo instead of running the bundled preset as-is, here's how pipeline automation for coding agents works.

Where Unstoppable Fits

Unstoppable runs Claude Code and Codex as bundled native runtimes against the same repo, in parallel isolated worktrees, and it does all of that in the background for you. Supporting both natively meant building two separate integration paths internally rather than one shared shim bolted over both CLIs. The boring engineering reason cross-provider fusion is rare, not a slogan.

It's bring-your-own-subscription on both sides: Claude through your Anthropic account or API key, Codex through ChatGPT login, an OpenAI key, or Bedrock. You're not paying twice for access you already have. On top of that, Unstoppable also hosts cheaper models directly, DeepSeek among them, billed through AI Credits instead of a separate provider subscription, so a task that doesn't need a frontier model doesn't have to cost frontier-model money. Every chat and every pipeline step picks its model per task, so switching from Opus to GPT to a hosted DeepSeek run is a dropdown, not a migration.

The worktree isolation is table stakes for running two agents at once without them clobbering each other's files. The pipelines (like the multi-provider code review above) are what turn "two agents, one repo" into something more than a parlor trick: six review tracks across two providers reconciled into one verdict, with a score that tells you when to actually look.

We are staying honest on two counts. We have no adoption numbers to wave at you, so we are not going to invent a "10,000 developers" line. And we are not claiming Claude beats Codex or the reverse, because the product only makes sense if you don't have to choose.

Match the Tool to the Task

Start with the task in front of you. Overnight, mechanical, terminal-heavy, or a review pass: reach for Codex. Unfamiliar codebase, long steering session, custom skills already in place: reach for Claude Code. Reviewing a finished change before it merges: don't reach for either one alone. A pipeline that fuses both providers' findings catches more and argues with itself less than a single model's opinion does.

Blake Crosley, credited with the blind-test rounds in Jeong's survey, landed the part that matters after the agents finish: "The real output isn't the winner. It's the synthesis step that extracts the strongest elements from both." That's a manual version of what a fusion pipeline does automatically: reconcile two (or six) independent takes into one you can actually act on.

Run both against the same repo today at unstoppable.ai.

How do they compare on SWE-bench Verified? Claude Opus 4.6 80.8, Gemini 3.1 Pro 80.6, DeepSeek V4 Pro (Max) 80.6, Kimi K2.6 80.2, GPT-5.4 not reported.

On SWE-bench Pro? Kimi K2.6 58.6, GPT-5.4 57.7, Claude Opus 4.6 57.3, DeepSeek V4 Pro (Max) 55.4, Gemini 3.1 Pro 54.2.

On LiveCodeBench? DeepSeek V4 Pro (Max) 93.5, Gemini 3.1 Pro 91.7, Kimi K2.6 89.6, Claude Opus 4.6 88.8, GPT-5.4 not reported.

On Codeforces rating? DeepSeek V4 Pro (Max) 3206, GPT-5.4 3168, Gemini 3.1 Pro 3052, Claude Opus 4.6 and Kimi K2.6 not reported.

On GPQA Diamond? Gemini 3.1 Pro 94.3, GPT-5.4 93.0, Claude Opus 4.6 91.3, Kimi K2.6 90.5, DeepSeek V4 Pro (Max) 90.1.

On MMLU-Pro? Gemini 3.1 Pro 91.0, Claude Opus 4.6 89.1, GPT-5.4 87.5, DeepSeek V4 Pro (Max) 87.5, Kimi K2.6 87.1.

On Humanity's Last Exam? Gemini 3.1 Pro 44.4, Claude Opus 4.6 40.0, GPT-5.4 39.8, DeepSeek V4 Pro (Max) 37.7, Kimi K2.6 36.4.

On Terminal-Bench 2.0? GPT-5.4 75.1, Gemini 3.1 Pro 68.5, DeepSeek V4 Pro (Max) 67.9, Kimi K2.6 66.7, Claude Opus 4.6 65.4.

Two things stand out. First, on SWE-bench Verified, the benchmark most directly tied to "can it fix a real GitHub issue," V4 Pro's 80.6 percent is functionally tied with Gemini 3.1 Pro and just 0.2 points behind Claude Opus 4.6, making it the highest-scoring open-weight model on that specific test. Second, it is not uniformly ahead: on the harder SWE-bench Pro set, on GPQA Diamond, and on Terminal-Bench 2.0, it trails several closed models by a few points, and its verbosity (below) means those wins come with a different token bill than the raw benchmark score implies. It leads decisively on LiveCodeBench and Codeforces, which reward raw competitive-programming skill, and it is closer to but not at the frontier on open-ended reasoning and long-horizon agentic tasks.

The "intelligence per dollar" number, and its asterisk

Artificial Analysis, which runs a standardized battery of nine evaluations (its Intelligence Index) across hundreds of models and prices each run using the vendor's own token rates, scores V4 Pro in its maximum-effort reasoning mode at 44 on that index, roughly third among the 93 comparably-sized open-weight models it tracks and well above the class median of 25 (Artificial Analysis). Running its full evaluation suite against V4 Pro cost Artificial Analysis $176 as of this writing; at the time of the SCMP report in May, that figure was quoted at $268, against roughly 12 times more for GPT-5.5 and roughly 19 times more for Claude Opus 4.7 to run the identical suite. Both figures are legitimate; they're snapshots from different points as Artificial Analysis has updated its Intelligence Index methodology, and either way the multiplier is large.

V4 Pro is also unusually verbose, a detail that rarely makes the per-token headline. Artificial Analysis notes it generated 180 million output tokens running its Intelligence Index, against a class median of 92 million, or roughly double. A model that reasons at length before answering can absolutely be worth it, but "double the output tokens" on a metric where output is billed per token quietly erodes some of the per-token price advantage before you've written a line of application code. The headline price-per-token comparison and the actual price-per-answer comparison are not the same number, and V4 Pro is a clean example of why.

Cost per finished task, not cost per token

What matters more than token price is cost per resolved task, because a cheaper model that fails more often and retries can end up costing more per finished feature than a pricier model that succeeds on the first attempt. SSOJet ran exactly this calculation in June 2026 across eight coding models, using a fixed hypothetical token budget (50,000 input tokens, 12,000 output tokens per attempt) divided by each model's published SWE-bench Verified solve rate, with every price pulled directly from vendor documentation (SSOJet, "8 AI Coding Models Ranked by Cost-per-Task", verified June 7, 2026):

DeepSeek V3.2, cost per resolved task? $0.28 input / $0.42 output per million tokens, 67.8% on SWE-bench Verified, about $0.028 per resolved task.

Qwen3-Coder 480B? $0.22 / $1.00, 69.6% solve rate, about $0.033 per task.

GLM-4.6? $0.39 / $1.74, 68.0% solve rate, about $0.059 per task.

Kimi K2.6? $0.60 / $2.50, 80.2% solve rate, about $0.075 per task.

Gemini 3 Flash? $0.50 / $3.00, 78.0% solve rate, about $0.078 per task.

Claude Haiku 4.5? $1.00 / $5.00, 73.3% solve rate, about $0.150 per task.

GPT-5.1? $1.25 / $10.00, 76.3% solve rate, about $0.239 per task.

Claude Opus 4.5? $5.00 / $25.00, 80.9% solve rate, about $0.680 per task.

That set of numbers predates V4 (it covers V3.2), but the same arithmetic applies cleanly to V4 Pro's verified numbers. Using SSOJet's stated method (their token budget, DeepSeek's official price, and the 80.6 percent SWE-bench Verified score from DeepSeek's own model card), V4 Pro works out to: (50,000 ÷ 1,000,000 × $0.435) + (12,000 ÷ 1,000,000 × $0.87) = $0.0322 per attempt, divided by an 80.6 percent solve rate, for about $0.04 per resolved task. That would land it essentially tied with Claude Opus 4.5 on accuracy (80.6 versus 80.9 percent) at roughly a seventeenth of Opus's per-task cost, and ahead of Kimi K2.6 on both price and solve rate under this same fixed-budget assumption. Two caveats belong right next to that number: it's a back-of-envelope estimate using a published methodology, not a benchmark anyone has run end-to-end on V4 specifically, and it doesn't account for V4 Pro's above-average verbosity, which would push real-world output token counts, and therefore real cost, above this estimate on tasks that invite long reasoning traces.

Speed isn't the tradeoff you'd expect

A 1.6-trillion-parameter model sounds like it should be slow. It mostly isn't, because Mixture-of-Experts routing only activates 49 billion of those parameters on any given token. Artificial Analysis measured V4 Pro (reasoning, max effort) at 71.7 output tokens per second, above the open-weight class median of 60.4, with a time-to-first-token of 1.77 seconds against a median of 2.01 seconds (Artificial Analysis). The hybrid attention architecture's long-context efficiency gains appear to be doing real work here, not just on the FLOPs and KV-cache numbers DeepSeek publishes itself.

The self-hosting option

Because V4 ships under MIT with weights posted on Hugging Face and direct support in vLLM and SGLang, self-hosting is a real lever, not a theoretical one. That converts a per-token API bill into a fixed hardware and operations cost, which wins decisively above some throughput threshold for steady, high-volume internal workloads, and it also sidesteps the data-residency question below entirely for teams that need to keep code off any third-party's servers. The tradeoff is the usual one for self-hosting any large MoE model: you need memory and orchestration for a model with 1.6 trillion total parameters even if only 49 billion are active per token, and someone on your team owns the operational burden that the hosted API otherwise absorbs.

The cost that doesn't show up on the pricing page

Where a model's inference actually runs matters too, separately from the per-token math above, especially for DeepSeek's hosted API. DeepSeek's own privacy policy states that user data, including chat content, is stored on servers in the People's Republic of China. Because of that, DeepSeek's hosted service is subject to China's 2017 National Intelligence Law, which can compel organizations operating there to support state intelligence work. As Northwestern University's Eddy Borges-Rey put it to Al Jazeera, "virtually all major tech companies... exploit user data to some extent," but Western governments treat Chinese platforms differently because of the jurisdiction they operate under, "whether this is happening in practice is debatable, but the mere possibility is enough to justify bans from a national security perspective" (Al Jazeera, February 6, 2025).

That's not a hypothetical concern for procurement teams. Italy's data protection authority ordered a limitation on DeepSeek's processing of Italian users' data in January 2025 over insufficient transparency about data handling. Australia barred DeepSeek from all government devices and systems in February 2025, citing national security risk. Taiwan prohibited its use across public-sector organizations, state-owned enterprises, and public schools. South Korea's industry ministry temporarily restricted it on official devices while its privacy regulator sought details on data handling, and the app later returned to South Korean app stores after roughly two months. The US Navy and NASA separately restricted DeepSeek on their own networks and devices around the same period. These restrictions have generally persisted into 2026 rather than being broadly lifted. For enterprises evaluating the API directly, DeepSeek also does not publish the kind of third-party compliance attestations (SOC 2, HIPAA business associate agreements) that OpenAI, Anthropic, and Google routinely provide, which is one reason many teams that want the model's cost and quality profile access it through US-hosted third-party inference providers under a standard enterprise agreement rather than DeepSeek's own API, or self-host the open weights entirely, rather than sending production data directly to DeepSeek's infrastructure.

This doesn't change the token-price math above, but it does mean "cheapest model" and "lowest total cost of adoption" can be different questions once data governance, compliance, and jurisdiction enter the calculation. A serious evaluation should treat that as a separate line item, not something the price-per-token comparison already accounts for.

What this means for choosing a model

Three things fall out of all this that are more useful than "DeepSeek V4 is cheap":

Sticker price and cost-per-finished-task are different numbers, and the gap between them is driven by solve rate and verbosity, both of which vary by task type. Compute the second number on your own workload before treating the first one as decisive.

Benchmark leadership is uneven, not universal. V4 Pro genuinely leads on some coding-specific evaluations and trails on others, and self-reported scores from any lab, including DeepSeek's, are worth validating against your own task distribution before you route production traffic based on them.

The cheapest defensible strategy for most teams is routing, not a single default model: send high-volume, low-failure-cost work to whichever model wins on cost-per-task for that work, and reserve frontier-tier spend for the tasks where a wrong answer costs more than the benchmark gap is worth. That decision won't stay settled for long: V4 Pro's own list price moved 75 percent in a single month, and whatever undercuts it next is probably already being benchmarked.