DeepSeek V4 Review: The Cheapest Serious 1M-Token Models in AI?

DeepSeek released the V4 family on April 24, 2026 — two models, V4-Pro and V4-Flash, live on the API and open-sourced on day one — and quietly reset the floor for what serious AI capability costs. Both models carry a 1M-token context window as the default, both offer thinking and non-thinking modes, and the flagship V4-Pro tops out at $0.87 per million output tokens: a price that undercuts every comparable 1M-context model on the market, open or closed.

And July is decision month. In a notice dated June 29, DeepSeek confirmed the official version of V4 arrives mid-July 2026, promising feature and performance improvements — and introducing peak-hour pricing that doubles rates during Beijing business windows. Ten days after that, on July 24, the legacy model names `deepseek-chat` and `deepseek-reasoner` retire for good. This review covers what V4 actually is, what it costs today and after the official release, where it wins, where it clearly doesn't, and what to do before the deadlines.

Quick verdict

Consider DeepSeek V4 if you…

- Run high-volume production workloads where unit cost decides what's viable — V4's pricing is in a class of its own

- Need long context cheaply: 1M tokens standard, up to 384K output tokens per response, with aggressive context-caching discounts

- Want drop-in integration — the API speaks both OpenAI ChatCompletions and Anthropic formats, so existing tooling mostly just works

- Value open weights and self-hosting options, especially the smaller V4-Flash

Hold off (or route elsewhere) if you…

- Run long-horizon autonomous agents — on Z.ai's published cross-model table, V4-Pro trails both GLM-5.2 and the closed frontier badly on marathon-style benchmarks

- Need image input: V4 is text-only

- Depend on flat pricing around the clock — from the official release in mid-July 2026, peak-hour rates double during Beijing business windows (off-peak keeps today's rates)

What is DeepSeek V4?

V4 is DeepSeek's fourth-generation model family, shipped as two Mixture-of-Experts models. DeepSeek-V4-Pro is the flagship: 1.6 trillion total parameters with 49 billion active per token. DeepSeek-V4-Flash is the efficiency tier: 284 billion total, 13 billion active. Both run a 1M-token context window by default across DeepSeek's official services, and both expose dual modes — thinking for hard problems, non-thinking for speed — under the model IDs deepseek-v4-pro and `deepseek-v4-flash`.

Two positioning details matter. First, V4 launched as a preview that DeepSeek treats as production-usable — and per the company's June 29 notice, the official version lands mid-July 2026 with "feature optimizations and performance improvements." Second, it launched open-sourced, continuing DeepSeek's pattern of releasing weights publicly — which keeps self-hosting and fine-tuning on the table, realistically for the 284B Flash more than the 1.6T Pro.

What's new in DeepSeek V4?

A 1M-token context as the default, not an upsell. Every official DeepSeek service now runs the full million-token window as standard — no separate long-context SKU, no premium rate.

Enormous output headroom. Both models support up to 384K output tokens per response — triple what most 1M-context rivals allow — which matters for whole-file code generation, long structured extractions, and report writing.

Dual modes on one endpoint. Thinking mode for hard problems, non-thinking for cheap fast calls, switchable per request instead of per model.

Two tiers with one API shape. Pro for capability, Flash for volume — same context, same modes, same integration.

Dual API compatibility. V4 speaks both OpenAI ChatCompletions and Anthropic API formats natively, so most existing agent tooling connects without a rewrite.

Pricing: what it actually costs

This is the section that makes V4 famous. V4-Pro costs $0.435 per million input tokens and $0.87 per million output. V4-Flash costs $0.14 and $0.28. Cache hits on repeated context are listed at well under a cent per million tokens — effectively free for agent loops that re-read the same project state.

For scale: GLM-5.2, itself celebrated as the value play of the summer, lists $1.40 / $4.40. Claude Sonnet 5 lists $3 / $15, Claude Opus 4.8 $5 / $25. V4-Pro's output price is a fifth of GLM-5.2's and roughly 3% of Opus 4.8's. Even if V4-Pro loses some head-to-head quality matchups — and it does — the economics change which questions are worth asking a model at all.

One change is coming with the official release. Per DeepSeek's June 29 notice, from mid-July all token prices double during peak windows — 09:00–12:00 and 14:00–18:00 Beijing time — while off-peak hours keep today's rates. Even doubled, V4-Pro's peak output price (about $1.74 per million) still undercuts GLM-5.2's standard rate, but the flat-price era ends with the preview: if your traffic peaks during Chinese business hours, model the swing — or schedule and route around it.

Review score

The scores below are our editorial assessment based on DeepSeek's official documentation and Z.ai's published cross-model benchmark table — treat the cross-vendor numbers as third-party context rather than DeepSeek's own claims.

- Coding: 8/10 — capable at everyday engineering; not the open-weight leader

- Agentic / tool use: 7/10 — solid tool orchestration, weak on marathon-length autonomy

- Reasoning: 8/10 — strong math and science scores for the price class

- Cost efficiency: 10/10 — nothing comparable comes close

- Context and long documents: 9/10 — 1M in, 384K out, near-free cache hits

- Speed: 8/10 — lean active-parameter counts, plus a non-thinking mode for fast paths

Overall: ~8.3/10 — the price-performance king of mid-2026, with a long-horizon asterisk.

Pros

- Pricing that resets the curve: $0.14–$0.87 per million tokens across the family

- 1M context standard with 384K output — the biggest output ceiling in its class

- Thinking and non-thinking modes on one endpoint, switchable per request

- OpenAI and Anthropic API compatibility means near-zero migration friction

- Open-sourced weights keep self-hosting and fine-tuning on the table

Cons

- Long-horizon agentic work is the clear weakness — on Z.ai's published table, V4-Pro scores 29.0 on FrontierSWE where GLM-5.2 posts 74.4 and Claude Opus 4.8 75.1

- Text-only: no image input in the V4 API

- Peak-hour pricing arrives with the mid-July official release — rates double during Beijing business windows, so budgets stop being flat

- Still a preview until mid-July, so behavior may shift with the official version

- The July 24, 2026 retirement of deepseek-chat and deepseek-reasoner forces legacy users to migrate on DeepSeek's schedule

- Self-hosting the 1.6T-parameter Pro is impractical for almost everyone (Flash, at 284B, is the realistic target)

How DeepSeek V4 compares

V4-Pro vs V4-Flash. Same context, same modes, same API — the split is quality versus throughput at a 3x price gap. A sensible default: Flash for summaries, extraction, classification, and chat; Pro for code, analysis, and anything a human reviews.

vs GLM-5.2. The open-weight fight of the summer, and it's genuinely close on value. GLM-5.2 is the stronger coder — on Z.ai's published table it leads V4-Pro 81.0 to 64.0 on Terminal-Bench 2.1 and dominates long-horizon work (74.4 vs 29.0 on FrontierSWE) — while V4 counters with prices 3–5x lower and triple the output ceiling. Volume pipelines lean V4; coding agents lean GLM-5.2.

vs the closed frontier. Claude Opus 4.8 and GPT-5.5 remain clearly stronger models across hard benchmarks. But at a 30–60x output-price gap versus Opus 4.8, V4 isn't trying to win that fight — it's trying to make 90% of your traffic too cheap to justify sending anywhere else.

The July 24 deadline: migrating off legacy names

If your integration still calls `deepseek-chat` or `deepseek-reasoner`, those names stop working after July 24, 2026, 15:59 UTC. They currently route to V4-Flash, which means your traffic is already running on V4 economics — but after the cutoff, requests fail outright. The migration itself is one line (`model: "deepseek-v4-pro"` or `"deepseek-v4-flash"`), so the real work is re-testing prompts against V4 behavior and deciding, endpoint by endpoint, which tier each workload deserves — or putting a router in front so the next deprecation is a config change instead of a code change.

Who should use DeepSeek V4?

High-volume products — support, summarization, extraction, classification — where V4 pricing turns marginal features profitable

Long-document and RAG-heavy workloads that fill 1M-token windows daily and profit from near-free cache hits

Teams generating long outputs: codebases, reports, structured data — 384K output is the class ceiling

Cost-conscious builders who want a capable default and are happy to escalate the hard 10% elsewhere

Who should look elsewhere?

Teams whose core workload is long-running autonomous agents — GLM-5.2 or a closed flagship is the safer lane

Vision-dependent workflows (V4 takes no images)

Anyone who needs a contractual 'stable' label today rather than a preview

Is DeepSeek V4 worth it?

For the price, emphatically yes — as a lane, not a religion. V4 doesn't beat the best open-weight coder (GLM-5.2) or the frontier flagships on quality, and its long-horizon agent numbers are genuinely weak. What it does is make enormous swaths of everyday AI work — the summaries, extractions, drafts, and chat turns that dominate real token bills — cost almost nothing. The winning pattern is V4 as the volume floor, with escalation lanes above it.

Final verdict

DeepSeek V4 is the price-performance benchmark every other model now gets measured against — our score: ~8.3/10. Run Flash where volume rules, Pro where quality matters but budgets are real, and route the summit-grade work to a stronger model. Re-run your cost math when peak pricing lands mid-July — cheap stays cheap, but it stops being flat. And if you're still on deepseek-chat or deepseek-reasoner: you have until July 24. Move.

Using DeepSeek V4 through OrcaRouter

A three-lane strategy — V4 for volume, a stronger open or closed model for hard tasks, a fallback for reliability — is exactly the setup that's painful to run by hand and trivial behind a gateway. OrcaRouter serves DeepSeek V4 alongside GLM-5.2, Claude, GPT, Gemini, and 200+ other models through one OpenAI-compatible endpoint, at provider list prices with zero token markup: smart routing picks the lane per request, automatic failover covers preview-era hiccups, and per-model observability shows what each lane actually costs per completed task. It also defuses provider-side changes like the July 24 name retirement or mid-July's peak-hour pricing — when a model name dies or a price window opens, you update a routing rule, not your codebase.

FAQ

Is DeepSeek V4 available now?

Yes — V4-Pro and V4-Flash have been live on the DeepSeek API since April 24, 2026, under the model IDs deepseek-v4-pro and deepseek-v4-flash, with open-sourced weights. Officially it's a preview; DeepSeek's June 29 notice puts the official version at mid-July 2026.

What is DeepSeek's peak-hour pricing?

Announced for the official release: from mid-July 2026, all token prices double during Beijing business windows (09:00-12:00 and 14:00-18:00 Beijing time), while off-peak hours keep current rates. Traffic that peaks outside Chinese business hours — most Western workloads — largely dodges the surcharge.

What happens to deepseek-chat and deepseek-reasoner?

They currently route to V4-Flash automatically, but after July 24, 2026 (15:59 UTC) both names are fully retired and requests will fail. Update the model parameter explicitly before the deadline.

Should I use V4-Pro or V4-Flash?

Flash for volume work a human never reviews line-by-line — summaries, extraction, classification, chat. Pro for code, analysis, and drafting, at roughly 3x Flash's price but still far below every comparable model.

Is DeepSeek V4 better than GLM-5.2?

Cheaper, not better. On Z.ai's published table GLM-5.2 leads clearly on coding and dominates long-horizon agent work; V4 counters with 3–5x lower prices, a 384K output ceiling, and near-free cache hits. Many teams run both: V4 for volume, GLM-5.2 for coding agents.

Can DeepSeek V4 handle images?

No — the V4 API is text-only. Route vision tasks (screenshots, PDFs as pixels, charts) to a multimodal model like Claude or Gemini instead.

Can I self-host DeepSeek V4?

The weights are open-sourced, but be realistic about scale: V4-Pro is a 1.6T-parameter MoE — data-center territory — while the 284B V4-Flash is the practical self-hosting target for well-resourced teams.

Does V4 work with my existing OpenAI or Claude tooling?

Mostly yes — the API natively supports both OpenAI ChatCompletions and Anthropic formats, so agent frameworks and SDKs built for either usually connect with a base-URL and model-name change.

Can I use DeepSeek V4 through OrcaRouter?

Yes — DeepSeek models are available on OrcaRouter at provider list prices with zero markup, next to GLM, Claude, GPT, and Gemini, so you can run the volume-plus-escalation split behind one endpoint.

Ready to make your token bill boring? Create your OrcaRouter account, point your OpenAI-compatible client at one endpoint, and route the volume to DeepSeek V4 while stronger models handle the summit — with zero token markup and a July 24-proof integration.