DeepSeek released V4.5 on July 5, 2026 the way it always does: weights on Hugging Face under an MIT license, an API priced low enough to make everyone else's spreadsheet look wrong, and a technical report that followed a few days later. Hosted pricing is $0.50 per million input tokens and $1.10 per million output — an order of magnitude below the proprietary frontier. The reasoning scores land in the frontier band: GPQA Diamond 91.8%, competition-math results sitting with models that cost twenty times as much.
The interesting thing about DeepSeek is no longer that it is cheap. It is that the gap between "cheap open model" and "frontier proprietary model" has narrowed to the point where, for a growing set of workloads, the open model is simply the correct choice and the only real question is whether you host it yourself.
The short version
| Model | License | GPQA Diamond | Input / Output (per 1M) | Weights? |
|---|---|---|---|---|
| Claude Opus 4.8 | Proprietary | 91.5% | $15 / $75 | No |
| GPT-5.6 | Proprietary | 91.0% | $5 / $30 | No |
| DeepSeek V4.5 | MIT | 91.8% | $0.50 / $1.10 | Yes |
| DeepSeek V4 Pro | Apache 2.0 | 89.9% | $0.44 / $0.87 | Yes |
| Llama 5 | Community | 88.0% | $0.80 / $2.40 | Yes |
Read the price column against the GPQA column. On this particular reasoning benchmark, DeepSeek V4.5 is level with the proprietary frontier while charging roughly a fifteenth of the price and handing you the weights. That is the pitch, and on reasoning-heavy tasks the pitch holds up.
Full specs are on the DeepSeek V4.5 model page; see it ranked on the AI model leaderboard.
The honest caveat, first
DeepSeek publishes strong numbers on launch day and the independent replications arrive later. The GPQA and math scores above are the vendor's, and the early third-party runs have tracked close to them — but "close" is doing some work, and benchmark figures from any lab on release day deserve a discount until the community reproduces them. On coding specifically, DeepSeek tends to score a little better on the public benchmarks than it feels in day-to-day use, where the proprietary models are still smoother.
So treat V4.5 as "frontier-band on reasoning and math, strong-but-not-leading on coding, and unbeatable on price." That framing has held for the last two DeepSeek releases and it holds here.
Where it wins
Reasoning and math per dollar. This is DeepSeek's genuine strength. On problems that are about working through a chain of logic correctly — math, science, structured analysis — V4.5 performs with the frontier at a rounding-error price. If your workload is reasoning-heavy and not dependent on the last point of coding polish, nothing else on the market comes close on value.
The MIT license. This matters more than people outside legal and procurement realise. MIT means you can run the weights in production, fine-tune them, ship them inside a product, and not have a lawyer flag a clause six months later. It is more permissive than Llama's community license and vastly more flexible than "call our API and hope the price holds." For regulated industries and anyone building a product on top of a model, the license is often the deciding factor, not the benchmark.
Self-hosting economics. Because you can download the weights, the marginal cost of a token self-hosted is infrastructure, not API price. At high volume, that flips the whole cost equation: a workload that costs five figures a month on a proprietary API can run on your own GPUs for the cost of the hardware and power. We worked the numbers in open source LLMs: 86% AI cost savings, and DeepSeek is the model that pushed that figure.
Where it falls short
Coding is good, not leading. On SWE-bench Pro, V4.5 reaches 62.1% — respectable, and well ahead of most open weights, but a clear step behind GPT-5.6 and Opus 4.8. For a coding product where quality is the whole value, the proprietary models still justify their price.
Self-hosting is real work. The weights are free; running them well is not. A frontier-scale model needs serious GPU capacity, an inference stack (vLLM, TGI, or similar), and someone who understands batching and quantisation to get the throughput economics to work. If you do not have that, the hosted API is fine — but then you are giving up part of the cost advantage that made you look at open weights in the first place.
Support is the community, not a vendor. When something breaks at 2am, there is no account manager. For teams that need an SLA and a phone number, that is a genuine cost that the sticker price does not show.
What it actually costs
Two numbers, depending on how you run it.
Hosted API: $0.50 per million input, $1.10 per million output. A reasoning workload of 300M input and 60M output tokens a month runs to $150 + $66 = about $216/month — a figure that would be well into four figures on a proprietary frontier model.
Self-hosted: token cost drops to zero and you pay for GPUs instead. This only makes sense above a volume threshold — roughly the point where your monthly API bill exceeds the amortised cost of the hardware plus an engineer's time to run it. Below that line, the API is cheaper and simpler. Above it, self-hosting can cut the bill by the 80%-plus figures the open-weight case is built on.
Most teams that adopt DeepSeek do not go all-in on it. They route the reasoning and high-volume work to V4.5 and keep a proprietary model for the coding and agent steps that need it. Swfte Connect runs exactly that split from one API and one bill, including routing to a self-hosted DeepSeek endpoint if you stand one up. See multi-model AI strategy for the reasoning.
The self-hosting break-even, worked out
The open-weight cost story only pays off above a certain volume, and it is worth doing the arithmetic rather than assuming self-hosting is always cheaper. It is not.
Start with the hosted API as your baseline: $0.50/$1.10 per million tokens. For a workload of 300M input and 60M output tokens a month, that is about $216. To beat that by self-hosting, your infrastructure plus the engineering time to run it has to cost less than $216 a month — and at that volume it almost certainly does not. A single capable GPU instance runs several hundred dollars a month before you have paid anyone to operate it. Below a few hundred million tokens a month, the hosted API wins on both cost and simplicity, and self-hosting is a hobby, not a saving.
The picture flips once you are running billions of tokens a month. At, say, 5B input and 1B output tokens, the hosted API bill is roughly $3,600 a month and climbing linearly. A self-hosted cluster sized for that throughput — a handful of GPUs kept busy with continuous batching — costs largely the same whether it serves 3B tokens or 8B, because you are paying for capacity, not usage. That is where the 80%-plus savings the open-weight case advertises actually come from: high, steady volume where you can keep expensive hardware saturated. We laid out the full model in cloud vs on-prem AI: TCO analysis.
So the honest rule is: use the hosted DeepSeek API until your bill is consistently into the low thousands per month and your volume is steady, then evaluate self-hosting. Do not stand up a GPU cluster to save $216 — you will spend far more than that in the time it takes to run it well.
Why the license is the real headline
For an engineer, the benchmark is the interesting number. For anyone shipping a product, the MIT license is the more important one, and it deserves more attention than it usually gets.
Most "open" models are not MIT. Llama ships under a community license with an acceptable-use policy and a scale clause. Some models restrict commercial use or field-of-use. Every one of these is a clause your legal team has to read, interpret, and sign off on before you build a product on top of the model — and a clause that can change how you are allowed to deploy six months in, after you have already committed. MIT has none of that. You can run it, modify it, fine-tune it, embed it in a commercial product, and redistribute it, with essentially one obligation: keep the copyright notice. For procurement and legal, that difference is enormous — it turns a weeks-long review into a formality.
This is why DeepSeek keeps winning in regulated industries even when a proprietary model scores a point or two higher. A bank or a hospital cannot send patient or transaction data to a third-party API without a mountain of compliance work, and it cannot build a product on a license it does not fully control. An MIT-licensed model it can run inside its own perimeter removes both problems at once. The benchmark gets DeepSeek onto the shortlist; the license is often what wins the deal.
Who should switch, and who shouldn't
Switch if your workload is reasoning-heavy, high-volume, or license-sensitive, and cost is a first-order concern. DeepSeek V4.5 is the best value on the board and MIT removes the usual open-weight legal friction.
Switch if you already self-host DeepSeek V4 Pro. V4.5 is a straight upgrade on the same license — better reasoning, same deployment story.
Skip it if your product is a coding assistant competing on quality, or if you need vendor support and an SLA and do not want to run inference infrastructure. In those cases the proprietary premium is buying you something real.
How to read a launch-day benchmark
One habit worth building, with DeepSeek specifically but with any lab: treat release-day numbers as a claim, not a result. It is not that labs lie — it is that the eval that produces a headline number is chosen by the people who want the number to look good, and the harder, more adversarial community runs come later. DeepSeek's last two releases held up under that scrutiny, which is why its numbers carry more credibility now than a first-time lab's would, but "held up" still meant a point or two of drift once independent harnesses got hold of them.
The practical move is to weight the benchmarks by how much you can verify them on your own data. GPQA and competition-math are hard to game and easy to reproduce, so DeepSeek's strong showing there is trustworthy. Coding benchmarks are noisier and more sensitive to harness details, so discount those a little and run the model on your actual repository before you believe the SWE-bench figure. The whole point of an open-weight model is that you can do this — download it, run your own eval, and trust what you measure over what anyone published. That option is worth as much as any single benchmark on the page.
The bottom line
DeepSeek V4.5 is the clearest sign yet that the open-weight tier has caught the frontier on everything except coding polish and hand-holding. It matches proprietary reasoning scores, ships under a license you can actually build on, and costs a fraction of everything above it. It will not replace GPT-5.6 or Opus 4.8 for the hardest coding work, and it asks you to bring your own infrastructure to unlock its best economics. But for reasoning at scale, it has quietly become the model to beat — and the one that sets the price everyone else is measured against.
Compare it against the full open-weight field in open weights vs proprietary AI models, or see the live ranking on the AI model leaderboard.