DeepSeek V4: The Cheapest Frontier Model, Explained

#ai #llm #machinelearning #opensource

DeepSeek made its name in early 2025 by proving frontier capability didn't require frontier budgets, and V4 is that thesis at full maturity. Released as a preview on April 24, 2026 and now the mandatory platform for all DeepSeek users — the legacy chat and reasoner endpoints retire permanently on July 24 — V4 pairs a genuinely frontier benchmark profile with pricing that reads like a typo: eighty-seven cents per million output tokens for the flagship. If you're evaluating AI vendors in 2026, this model belongs in the conversation whether or not you end up using it, because it resets what 'expensive' means everywhere else.

Two models, one philosophy

DeepSeek V4 family — July 2026| Spec | V4-Pro | V4-Flash |
| --- | --- | --- |
| Total / active parameters | 1.6T / 49B active | 284B / 13B active |
| Context window | 1M tokens (384K max output) | 1M tokens (384K max output) |
| Price per 1M tokens (in / out) | $0.435 / $0.87 | $0.14 / $0.28 |
| Weights | Open, permissive license | Open, permissive license |
| Headline result | Matches GPT-5.5 / Claude Opus 4.7 on most agentic benchmarks | Frontier-adjacent at commodity pricing |
| Best for | Primary agentic workloads on a budget | High-volume routine steps |

The headline benchmark: the top V4-Pro configuration scores 80.6% on SWE-bench Verified — the highest open-weights entry on the leaderboard, tied with Google's Gemini 3.1 Pro. Across the broader agentic suite, V4-Pro matches GPT-5.5 and Claude Opus 4.7 — last generation's closed flagship and this generation's closed workhorse — at roughly ten to thirteen times lower output cost. That is the entire pitch, and it's a strong one: not the best model available, but plausibly the best capability-per-dollar ever shipped.

[Diagram available in the original article — view on softwarebuilding.ai]

That chart is the open-weight market in one image. A year ago 'open model' meant one price tier and a capability discount. Now it spans Kimi K3 at $15 per million output tokens playing in flagship territory, GLM-5.2 and MiniMax M3 in the mid-tier, and DeepSeek occupying the floor — with a SWE-bench score that embarrasses models charging thirty times more. V4-Flash deserves special mention: at $0.28 per million output tokens, running it a thousand times costs less than a single dinner, which changes what's economically thinkable for high-volume automation.

The July 24 migration — and what it signals

One operational note that's easy to miss in the benchmark noise: on July 24 at 15:59 UTC, DeepSeek permanently retires its legacy deepseek-chat and deepseek-reasoner API aliases. Anything still pointed at them stops working — migration to V4 is mandatory, with no long deprecation tail. If any of your systems touch DeepSeek's API, that's a this-week action item. It's also a useful reminder of the trade-off this vendor represents: extraordinary price-performance, moved at a pace that assumes you can move too. Closed vendors deprecate gently because enterprises pay them to; DeepSeek's cadence is closer to open-source infrastructure, where the price of the discount is keeping up.

When V4 is the right call — and when it is not

Reach for V4-Pro when agentic capability per dollar is the deciding factor: internal coding agents, data-processing pipelines, document automation, anything high-volume where an occasional retry is acceptable and a 10x cost difference compounds into real money.
Reach for V4-Flash for the commodity steps — classification, extraction, summarization, routing — where it is arguably the best pure value in the market right now.
Stay with a closed frontier model where the hardest reasoning matters (Fable 5 and Sol still clearly lead the top end), where vendor SLAs and gentle deprecation policies are what your ops team is buying, or where your compliance posture restricts Chinese-origin models regardless of where you host the weights — the same review we outlined for GLM and MiniMax applies unchanged.
And in most real systems: it's not either/or. The routing pattern — open value tier for volume, closed frontier for the hard 20% — is exactly where V4 slots in as the volume workhorse.

Note: The uncomfortable question V4 poses to every AI budget: if an open model matches your current workhorse at a tenth of the price, what exactly is the other 90% buying? Sometimes the answer is real — support, SLAs, the last few points of reliability, compliance cover. Sometimes it's inertia. Running twenty of your real tasks through V4-Pro costs almost nothing and tells you which.

DeepSeek V4 — common questions

Is DeepSeek V4 really as good as GPT or Claude?

On agentic benchmarks, V4-Pro genuinely matches GPT-5.5 and Claude Opus 4.7 — a closed flagship one generation back and the current closed workhorse — and its top configuration's 80.6% on SWE-bench Verified is the best open-weights score recorded, tied with Gemini 3.1 Pro. What it does not match is the current closed frontier: Claude Fable 5 and GPT-5.6 Sol still hold a clear lead on the hardest reasoning and the messiest real-repository work (Fable's 80.3% on the much harder SWE-bench Pro is a different tier of result). So the accurate framing: V4 delivers last year's flagship capability, verified, at roughly a tenth of the price — which for a large share of production workloads is exactly the right trade. For the frontier-hard 20%, pair it with a stronger model behind a router rather than forcing either to do the other's job.

What is the difference between DeepSeek V4-Pro and V4-Flash?

Scale and price, on a shared foundation. V4-Pro is the capability play: 1.6 trillion total parameters with 49 billion active per token, at $0.435 per million input tokens and $0.87 per million output — the one that matches closed workhorse models on agentic benchmarks. V4-Flash is the efficiency play: 284 billion total parameters with 13 billion active, at $0.14 in and $0.28 out — roughly a third of Pro's price for a model that stays frontier-adjacent on routine work. Both carry the same 1M-token context window, the same unusually large 384K max output (useful for long document generation and big code diffs), and the same permissive open-weights license. The practical split in systems we build: Flash handles classification, extraction, summarization, and templated drafting; Pro takes the genuinely agentic steps — multi-turn tool use, code changes, complex document work — and the two together often cover 80% of a pipeline's volume for single-digit percent of a closed-stack budget.

What happens on July 24 for existing DeepSeek users?

At 15:59 UTC on July 24, 2026, DeepSeek permanently retires the legacy deepseek-chat and deepseek-reasoner API aliases — requests to them stop working entirely, and migration to the V4 family is mandatory. If you or a vendor you depend on integrates DeepSeek, the checklist is short but urgent: inventory anything calling the old aliases, repoint to the V4 endpoints, and re-run your evaluation set, because V4's behavior differs from the models the aliases fronted — outputs, latencies, and token counts all shift, and prompts tuned to the old models deserve a quick re-check. The hard-cutoff style is worth registering as a vendor-management signal, too: DeepSeek gives you extraordinary economics and expects you to keep pace with its release cadence. Budget a small amount of recurring engineering attention for that, and price it into the comparison against slower-moving closed vendors — it's usually still an overwhelming win, but it isn't free.

Should we build our AI product on DeepSeek V4?

Build on it, yes, for the right layer of the system — bet everything on it, no, for the same reason we give about every model in this series. V4's economics make it close to unbeatable as the volume tier of a routed architecture: open weights mean you can self-host if data residency demands it or switch hosted providers freely, the permissive license removes most legal friction, and the benchmark profile is verified rather than vendor-claimed. The cautions: the hardest reasoning still belongs to closed frontier models; the July 24 alias retirement shows this vendor moves fast and expects you to; and organizations with restrictions on Chinese-origin models need the compliance review regardless of hosting choices. Structure the system so V4 is a slot, not a foundation — model-agnostic tool interfaces, per-step routing, evals that run against any backend — and you get the 10x economics while staying one config change from whatever ships next quarter. Deciding where each model belongs in that structure is exactly the scoping conversation we have with clients.

Sources and further reading

Originally published at https://softwarebuilding.ai/blog/deepseek-v4-explained.