DeepSeek V4: The Complete Guide to Pro, Flash and Pricing
I've been running DeepSeek V4 since the April preview — both models, mostly for coding and long-document work. The short version, since that's what most people want: Flash is my default for everything. Pro is what I switch to when Flash measurably isn't enough. And if you can push your heavy jobs to after 6pm Beijing time, everything costs half as much. The rest of this guide is those three sentences unpacked, with a date and a caveat attached to every number — V4 changed under everyone's feet twice this year, and a fair chunk of what you'll read about it online quietly expired months ago.
What Is DeepSeek V4?
A family of open-weight mixture-of-experts models, MIT-licensed — download the weights, run them, fine-tune them, ship them in a product, no strings. Two members: V4-Pro, the 1.6-trillion-parameter flagship, and V4-Flash, the smaller and much cheaper one. Both give you a 1M-token context window by default (not as a premium tier — as the default), up to 384K tokens of output in one go, Thinking and Non-Thinking modes, and three reasoning-effort levels (low, high, max) to trade depth against latency.
If you're coming from earlier DeepSeek models, the migration had a sharp edge: the old deepseek-chat and deepseek-reasoner endpoints were fully retired on July 24. If your code broke that week and you hadn't touched it, that was why.
The timeline explains most of the contradictory claims you'll meet online:
| Date (2026) | What happened |
|---|---|
| Apr 24 | Preview release: Pro and Flash ship together, open weights, API live same day |
| Jul 24 | Legacy endpoints deepseek-chat / deepseek-reasoner retired |
| Jul 31 | Flash reaches GA (checkpoint V4-Flash-0731) |
| Aug 13 | Pro GA release: agent upgrades, effort levels, native Responses API |
| Aug 16 | New pricing takes effect — peak and off-peak rates |
| Aug 21 | deepseek-v4-flash-vision-exp multimodal model + free Files API |
An April blog post and an August one describe materially different products — different checkpoints, different prices, different benchmark claims. So whenever a number matters below, I'll pin it to its date. That's the only honest way to write about this model.
Pro vs Flash: Which One Do You Actually Need?
Here's the spec sheet, then my take.
| Spec | Pro | Flash |
|---|---|---|
| Total / active params | 1.6T / 49B | 284B / 13B |
| Context window | 1M (default) | 1M (default) |
| Max output | 384K | 384K |
| Modes / effort | Thinking + Non-Thinking; low/high/max | Same |
| Current checkpoint | Pro-0813 (Aug 13) | Flash-0731 (Jul 31) |
| Download size | 865 GB | 160 GB |
| License | MIT | MIT |
Pro activates 3.8× more parameters per token — that's where its reasoning edge lives. Flash costs about a third of Pro's price on every line item. And on Artificial Analysis's independent Intelligence Index, the two GA checkpoints land one point apart: 53 for Pro, 52 for Flash.
One point. That gap is smaller than the gap between either of them and everything else at their price.
So my rule, after months of both: Flash by default — coding loops, batch work, extraction, summarization, agents. Switch to Pro when you can measure that Flash is the bottleneck, not when you feel like it. The official positioning agrees, in marketing language: Flash's reasoning "closely approaches" Pro, and it performs on par with Pro on simple agent tasks. You're not losing much on the typical job; you're losing 3× on every atypical one if you default to Pro.
Why 1M Context Got Cheap
A million tokens used to be a luxury tier that providers charged extra for. Making it the default took structural work, and the official one-line summary is: token-wise compression plus DeepSeek Sparse Attention. The technical report opens it up, and it's worth understanding in plain terms, because it's why long sessions on V4 don't cost what they logically should.
Picture an archive that refuses to build a bigger building. CSA (Compressed Sparse Attention) binds older documents into condensed volumes — roughly 4× compression along the sequence — while a shelf of recent tokens stays uncompressed, word for word. HCA (Heavily Compressed Attention) goes much further: think microfilm, about 128× compression, no selective retrieval at all. And a lightning-fast indexer — the librarian who, given any question, instantly pulls the most relevant 1,024 condensed volumes — decides what the model actually reads each step. Pro interleaves these strategies across its layers and stores most of the compressed archive in FP8.
Two more upgrades ride along: Manifold-Constrained Hyper-Connections (cleaner signal through the residual paths) and the Muon optimizer (faster, more stable training). Pre-training ran on 32T+ tokens.
The payoff, measured against V3.2 at a full 1M context: 27% of the inference FLOPs per token, and about 10% of the KV cache. That's not an incremental improvement — it's the difference between 1M context as a demo and 1M context as something you can actually run all day.
One honesty note: the 4×/128× figures and the indexer internals come from third-party analysis of the model (the official report states the mechanisms but not those specific ratios), so treat the fine details as well-informed reconstruction rather than gospel.
For contrast with the generation before:
| Dimension | V3 | V4-Pro |
|---|---|---|
| Total / active params | 671B / 37B | 1.6T / 49B |
| Context | 128K | 1M |
| Max output | 8K | 384K |
| Attention | MLA | Token-wise compression + DSA |
| API model names | deepseek-chat / deepseek-reasoner | deepseek-v4-pro / deepseek-v4-flash |
The Benchmarks, Honestly
Every benchmark number you'll see about V4 comes from one of three places, and they tell three different stories. You need all three.
The official claims (April, preview era). Open-source SOTA for agentic coding; world-knowledge second only to Gemini 3.1 Pro among open models. Pro-Max posted MMLU-Pro 87.5, GPQA Diamond 90.1, Terminal-Bench 2.0 at 67.9. Impressive — and self-reported, on benchmarks the lab chose.
Independent tracking (June, llm-stats). On SWE-bench Verified, V4-Pro-Max lands at 80.6% — the best open-weight score, tied with Gemini 3.1 Pro:
| Model | SWE-bench Verified | Output $/1M |
|---|---|---|
| Claude Fable 5 | 95.0% | $50 |
| Claude Opus 4.8 | 88.6% | $25 |
| Claude Opus 4.6 | 80.8% | $25 |
| DeepSeek V4-Pro-Max | 80.6% | $1.98 (off-peak) |
| Gemini 3.1 Pro | 80.6% | $12 |
| MiniMax M3 | 80.5% | $1.20 |
The contamination-resistant test. DeepSWE — writing projects from scratch across 91 fresh repos, no chance of training leakage — told a colder story in April: preview-era V4-Pro managed 8% pass@1, against 70% for GPT-5.5 and 54% for Opus 4.7. Then the GA release (Aug 13) self-reported DeepSWE jumping from 12.8 to 62.7, alongside big gains on Terminal-Bench and NL2Repo. Those GA numbers have no independent replication yet, and Scale's SEAL leaderboard still carries no V4 entry at all.
My read: the 80.6% is real but sits on a lenient verifier. The GA gains are directionally credible — the checkpoint genuinely changed — but I'd wait for independent numbers before quoting the jump as settled fact. If a task of yours lives in that gap between 80.6% and 95%, the cheap model will burn your time instead of your money. That's the actual trade, and no benchmark table will make it for you.
What It Costs — and When It's Half Price
Pricing changed on August 16 (16:00 UTC). What follows is the current structure; anything you read showing different numbers is either older than that or wrong.
| Per 1M tokens | Pro off-peak / peak | Flash off-peak / peak |
|---|---|---|
| Input (cache miss) | $0.66 / $1.32 | $0.22 / $0.44 |
| Input (cache hit) | $0.022 / $0.044 | $0.007 / $0.014 |
| Output | $1.98 / $3.96 | $0.66 / $1.32 |
Peak hours are UTC 01:00–04:00 and 06:00–10:00 — seven hours a day, weekdays. In Beijing time that's 9:00–12:00 and 14:00–18:00: standard office hours. Everything else is off-peak, at half price. So if you're in China (or anywhere UTC+8), your evening work sessions run at off-peak rates automatically. Europe gets a similar deal; the US mostly sleeps through the cheap window.
You'll still see "17× cheaper than Claude" floating around. That was April math. Current off-peak output pricing: about 12.6× cheaper than Opus 4.8 ($1.98 vs $25), 7.6× cheaper than GPT-5.4/Sonnet 4.6. Still remarkable; no longer the headline number.
The multiplier most people miss is the cache. Input that hits the cache costs 1/30th of a miss, and agentic workloads — where the same system prompt and files get resent every turn — land most of their input in the cache. Practical effect: your real bill runs well below the sticker miss-price. Back-of-envelope for a daily workload of 1M input + 200K output tokens, off-peak, zero cache hits: Flash ~$0.35/day (~$11/month), Pro ~$1.06/day (~$32/month). The same volume on Opus 4.8 is ~$300/month. Cache hits pull the DeepSeek numbers down further from there.
Two footnotes for completeness: April's launch prices ($1.74/$3.48 for Pro) are dead, and mid-2026 blog posts still quoting them — or the CNY prices from that era — are describing a product that no longer exists. Third-party resellers like OpenRouter carry both models at flat rates without the peak/off-peak swing, if you prefer predictable billing over hunting the cheap hours.
How to Use It: Web, API, Agents, Self-Hosting
Web and app — the free way. chat.deepseek.com runs two modes: Expert Mode (Pro) and Instant Mode (Flash). No API key, no billing. If you just want to feel out what the model can do, start here.
API — base URL stays https://api.deepseek.com; change the model name and you're migrated:
from openai import OpenAI
client = OpenAI(api_key="your-key", base_url="https://api.deepseek.com")
resp = client.chat.completions.create(
model="deepseek-v4-flash", # or "deepseek-v4-pro"
messages=[{"role": "user", "content": "Review this patch..."}],
extra_body={"thinking": {"type": "enabled"}}, # optional: Thinking mode
)Since the Aug 13 GA, the API also speaks the OpenAI Responses format natively, so Codex-style tools connect without an adapter. The official pricing page tracks current rates.
Agents — this is where V4 has been most aggressively positioned: the GA announcement name-checks integration with Claude Code, OpenClaw, and OpenCode. For the Claude Code route, DeepSeek exposes an Anthropic-compatible endpoint:
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_API_KEY=your-deepseek-keyTwo env vars and Claude Code runs on V4. Full walkthrough in the Claude Code × DeepSeek guide.
Self-hosting — the weights are MIT. Flash at 160GB is a realistic target for a serious workstation or a rented box; Pro at 865GB means multiple nodes, and you'll know if you need it. Sampling guidance from the model card: temperature 1.0, top_p 0.95 for agentic work, and give high/max-effort requests room for the full 384K output.
Images — as of Aug 21 there's deepseek-v4-flash-vision-exp: matches Flash's text ability, handles multimodal agent tasks near Opus-4.8 level, charges ≤384 tokens per image at Flash prices, and the Files API is free. Experimental, but no longer a gap in the lineup.
Where It Still Falls Short
The gap to the frontier is real. 80.6% versus 88.6% (Opus 4.8) versus 95% (Fable 5) on SWE-bench Verified — most days that gap is invisible, and then one hard task lands inside it and you spend an evening on retries. Know which mode you're in.
The GA benchmark jumps are self-reported. DeepSWE going from 12.8 to 62.7 in one checkpoint is a big claim; nobody independent has confirmed it yet, and SEAL doesn't list V4 at all.
Pricing has moved twice in four months (April launch prices → Aug 16 peak/off-peak). Budget accordingly: treat any cost projection older than a quarter as fiction.
Community-reported rough edges, your mileage may vary: sessions past ~500K tokens have shown drift; Pro sometimes ignores instructions in CLAUDE.md and needs watching on long agent runs. I keep an eye on the trajectory view when a run matters. And if you're waiting for V4 news, the team's own warning applies — rely only on official accounts, because the fake-site problem around DeepSeek is genuinely bad.
DeepSeek V4 FAQ
When was DeepSeek V4 released? Preview April 24, 2026 (both models, open weights). Flash hit GA July 31, Pro August 13. If you see a single "release date," it's probably one of these three — check which.
Is DeepSeek V4 open source? Is it free? Open weights under MIT — free to download and use yourself. The web and app versions are free to use. API usage is pay-as-you-go; new accounts have sometimes come with trial credits, but check the platform page rather than any article's numbers, mine included.
How big is the context window? 1M tokens by default, with up to 384K output. That's roughly a million English words in, a small book out.
What does the API cost? Off-peak: $0.22/$0.66 per 1M in/out for Flash, $0.66/$1.98 for Pro (Aug 16 pricing). Peak hours — Beijing time 9–12 and 14–18 on weekdays — double those. Cache hits cost 1/30th of a miss.
Pro or Flash for daily work? Flash, unless you can measure the difference. See the comparison above — the independent index has them one point apart.
Can it compete with GPT-5.5 or Claude Opus? On coding benchmarks, Pro-Max sits with Gemini 3.1 Pro at 80.6% — behind Opus 4.8 (88.6%) and Fable 5 (95%), at a third to a twenty-fifth of the price. Whether that trade works depends on whether your tasks live in the gap.
Does it support images? Since Aug 21: deepseek-v4-flash-vision-exp, experimental, ≤384 tokens per image. Writing better prompts for any of these modes is its own topic — the prompt guide covers it.
That's the whole picture of DeepSeek V4 as of late August 2026. If you're starting out: try Expert Mode on the web for free, grab an API key when you want it in your own tooling, and point your agent of choice at the Anthropic-compatible endpoint when you're ready to go further. The DeepSeek hub has the rest of the ecosystem — pricing deep-dives, Harness, tool integrations — as it ships.