DeepSeek V4 Brings a 1.6-Trillion-Parameter Open Model and a Price That Undercuts GPT-5.5 by 9x
The MIT-licensed V4-Pro and V4-Flash carry million-token context and trail the best closed models by only a few months, DeepSeek says.
Key takeaways
- V4-Pro has 1.6T total and 49B active parameters; V4-Flash has 284B and 13B. Both offer 1M-token context under an MIT license.
- V4-Pro lists at $1.74 input and $3.48 output per million tokens versus $5 and $30 for GPT-5.5.
- DeepSeek's 80.6% SWE-bench claim measured 74.0% in an independent CAISI test, a reminder to check vendor scores.
DeepSeek released a preview of V4 on April 24, 2026: two open-weight mixture-of-experts models, V4-Pro and V4-Flash, both with a one-million-token context window and an MIT license. The company says the Pro model trails the best closed systems by roughly three to six months, at a sliver of the price.
Two models, one architecture
V4-Pro has 1.6 trillion parameters in total, with 49 billion active for each token. V4-Flash is smaller at 284 billion total and 13 billion active. Weights are downloadable from Hugging Face and ModelScope; Pro is about 865 GB and Flash about 160 GB, so self-hosting the larger model takes a serious cluster. DeepSeek claims major efficiency gains in long context: at one million tokens, V4-Pro needs 27% of the per-token inference compute and 10% of the key-value cache that V3.2 required. Those savings are what make a million-token default economically plausible.
Benchmarks and the verification gap
DeepSeek's own scorecard for V4-Pro-Max lists 87.5% on MMLU-Pro, 90.1% on GPQA Diamond, 67.9% on Terminal-Bench 2.0, 55.4% on SWE-Bench Pro and 83.5% on a one-million-token retrieval test. On the same comparison, GPT-5.5 scores 82.7% on Terminal-Bench 2.0 and Claude Opus 4.7 scores 69.4%, so on terminal agent work the open model sits close to Anthropic and well behind OpenAI. On SWE-Bench Pro, Pro's 55.4% compares with 58.6% for GPT-5.5 and 64.3% for Opus 4.7.
Independent testing tempers the claims. The US government's CAISI measured V4-Pro at 74.0% on SWE-bench rather than the 80.6% DeepSeek reported, a 6.6-point gap that DeepSeek attributed to scaffolding and token budgets. That kind of difference is routine for coding benchmarks, where harness details move scores, but it means buyers should run their own evaluations on their own repositories.
Why it matters
Price is the sharpest edge. At $3.48 per million output tokens, V4-Pro costs about one-ninth of GPT-5.5 and about one-seventh of Opus 4.7, and V4-Flash at $0.28 is cheaper still. For agent workloads that generate millions of tokens, a model that is somewhat worse but an order of magnitude cheaper changes what is worth automating. It also lets companies keep sensitive data in-house, since the weights can run on private infrastructure.
There is a ceiling on that argument. Closed labs lead on the hardest agentic tests, release cadence is now measured in weeks, and the three-to-six-month gap DeepSeek cites keeps moving as the frontier advances. Open weights also raise familiar policy questions about misuse and export controls on the chips used for training.
What the efficiency claims mean in practice
Long context is notoriously costly. In a standard transformer, the key-value cache grows with every token held in memory, so serving a million-token prompt can demand far more GPU memory than the weights themselves. A tenfold reduction in cache size, if it holds, means one server can handle many more simultaneous long sessions, which is what drives the low API prices. DeepSeek has a history of this pattern: it builds models around inference efficiency, then passes the savings to customers, forcing competitors to justify their margins.
The preview label is worth noting. DeepSeek positioned this as an early release, and third-party platforms were still publishing their own evaluations in the days after launch. A later point release, V4.1 Flash, was listed by Vals.ai with a September date, although that page could not be fully confirmed, so the family appears to be iterating quickly.
What to watch
- Independent leaderboard placements for V4-Pro against Opus 4.7 and GPT-5.5 on private test sets.
- Whether the preview label is dropped and a full release adds the missing capabilities.
- Price responses from closed labs, particularly for their smaller, faster models.
- Adoption by enterprises willing to self-host a 865 GB model.
Sources
- DeepSeek V4 — DataCamp
- DeepSeek Releases V4: Open-Source 1.6T MoE with 1M Context — NYU Shanghai RITS
- DeepSeek V4 preview: open weights rival closed models — Kollab
- OpenAI's new GPT-5.5 arrives with advanced agentic coding and computer use — Neowin
