The Real Cost of AI in 2026: Self-Hosting vs. API, Re-Run With Today's Numbers

About eighteen months ago we published a cost breakdown comparing self-hosting DeepSeek R1 against paying for the DeepSeek and OpenAI APIs. It was one of our most-shared posts, and people still email us about it.

The problem is that almost every number in it is now wrong.

R1 has been retired. OpenAI's "o1" is a legacy line item. H100s got cheaper to rent, then the models got so much bigger that you need more of them anyway. And API prices for the same class of work dropped so far that the monthly figures in our old post look like typos.

So here is the same analysis, re-run with September 2026 pricing. Same structure, same question: if you're building a product on top of a large language model, what does it actually cost to run — and should you host it yourself or rent it by the token?

What Changed Since Our Last Breakdown

Three shifts matter more than any single price:

• •The open-weight frontier moved to DeepSeek V4. V4-Pro (1.6T total parameters, 49B active) and V4-Flash (284B total, 13B active) both ship under the MIT license. R1 and its distilled variants are no longer the models you'd deploy.
• •OpenAI's flagship is now GPT-6 Astra, and the GPT-5.x family became the workhorse tier. The o-series reasoning models still exist on the price list, but they are not where new workloads land.
• •DeepSeek introduced peak-hour pricing. Since August 2026, calls made during 01:00–04:00 and 06:00–10:00 UTC on weekdays cost double the off-peak rate. If your users are in the US, most of your traffic is already off-peak.

Everything below reflects those changes.

Self-Hosting Costs in 2026

Our 2025 post priced a 16× H100 cluster at $2.50 per GPU-hour. Today the on-demand median for an H100 is closer to $3.25, and the more relevant problem is memory: DeepSeek V4-Pro's native checkpoint is roughly 865 GB, which does not fit on 8× H100 (640 GB). You need 8× H200 (1,128 GB) or Blackwell hardware.

V4-Flash is the practical self-hosting target for most teams. It delivers most of V4-Pro's quality, and its ~167 GB checkpoint fits on a single B200 with room for KV cache.

For comparison, our 2025 numbers were $28,800/month for R1 on 16× H100 and $3,600/month for a distilled model on 3 GPUs. The "big model" tier got slightly cheaper thanks to H200 pricing; the "small model" tier got slightly more expensive because V4-Flash is a far more capable model than the 2025 distillations and needs a bigger single GPU.

Two caveats that were true then and are still true now:

• •These are compute-only numbers. A production deployment needs at least two nodes for redundancy, someone to own vLLM/SGLang upgrades, monitoring, and on-call. Budget an engineer's time, not just GPU-hours.
• •Reserved and spot pricing can cut these by 30–50%. But reserved capacity is a commitment, and spot capacity can disappear under you. The on-demand numbers are the honest baseline.

API Costs in 2026

This is where the story flipped. To keep it apples-to-apples we assume an average query of 2,000 input tokens and 1,000 output tokens (a typical RAG or assistant call with some context), and 10,000 queries per month, the same volume as our original post.

In 2025 we quoted $800/month for DeepSeek and $1,800/month for OpenAI at the same 10,000 queries. Today the equivalent tiers are $9 and $200. That is not a rounding error; it is a roughly 90–99% reduction in the cost of a unit of AI work in eighteen months.

A few things worth noticing in that table:

• •Prompt caching changes the math further. DeepSeek bills cache-hit input tokens at $0.003 per million off-peak, and OpenAI discounts cached input by 90%. If you have a long, stable system prompt, your real per-query cost can be well below the figures above.
• •The spread between providers is now wider than the spread between self-hosting options. GPT-6 Astra costs about 78× more per query than V4-Flash. Choosing the right tier for each task is now the single biggest cost lever you have.
• •Reasoning-heavy work costs more than the headline. Models that "think" bill their reasoning tokens as output. For agentic or multi-step tasks, expect 3–10× the base per-query cost.

Where Self-Hosting Actually Breaks Even

At 10,000 queries a month, self-hosting is not a cost decision; it's a control decision. The numbers only start to converge at serious scale. Here is the same comparison at 1,000,000 queries per month:

The break-even points that fall out of this:

• •Self-hosting V4-Flash beats the GPT-5.4 API at roughly 245,000 queries per month. Beyond that, the single-B200 deployment wins on pure compute cost.
• •Self-hosting V4-Flash never beats the V4-Flash API on cost until you pass about 5.4 million queries per month — and that's before you add redundancy and engineering time.

In other words: if you are cost-optimising and happy with an open-weight model, the cheapest path is almost always DeepSeek's own API. Self-hosting wins when you need something the API cannot give you.

Cost-Benefit Analysis: Who Wins on What

The last column is the one that decides most real-world projects. If you are in healthcare, finance, legal, or handling anything under a data residency requirement, the question is rarely "which is cheaper" — it is "which one can we get through compliance review." DeepSeek's API in particular processes requests on infrastructure in China, which is a hard stop for many regulated US buyers regardless of price. Self-hosting the same open weights on your own cloud account removes that objection entirely, which is exactly why V4's MIT license matters.

SWOT Analysis: Building on APIs in 2026

Strengths

• •Zero hardware investment, no maintenance, no upgrade cycles to manage.
• •Immediate access to the frontier: GPT-6 Astra and DeepSeek V4-Pro were available via API on launch day.
• •Per-query costs have fallen far enough that AI is now a rounding error for most products at moderate volume.

Weaknesses

• •Vendor dependency: model deprecations happen fast. Everything we recommended in 2025 is retired today.
• •Time-of-day pricing (DeepSeek) and promotional rates (OpenAI's GPT-5.6 Sol is promotional through November 2026) make forecasting harder than it looks.
• •Data leaves your environment on every call.

Opportunities

• •Model routing: send simple queries to a $0.001-class model and reserve frontier models for the 5% of requests that need them. Teams doing this well are cutting AI spend by 60–80% with no quality loss users notice.
• •Prompt caching turns long system prompts and retrieved context from a cost centre into a near-free input.

Threats

• •Regulatory exposure: cross-border data flows to model providers are increasingly scrutinised, and a provider's compliance posture becomes your compliance posture.
• •Outages and rate limits are inherited. A provider incident is your incident.
• •Price increases at the top tier: while mid-tier pricing has fallen, the flagship models cost more per token than their predecessors did.

So, What's the Best Approach?

The honest answer has not changed — it depends on your volume, your data sensitivity, and how much engineering you want to own. But the practical guidance has sharpened:

• •Under ~250,000 queries per month: use an API. Pick the cheapest model tier that passes your quality bar, add prompt caching, and revisit in six months. Self-hosting at this scale is paying $5,000/month to avoid a $10 bill.
• •250,000 to a few million queries per month, no data constraints: still an API, but now model routing across tiers is worth real money. This is where most growing SaaS products sit.
• •Any volume with regulated or sensitive data: self-host V4-Flash on a single B200 (or two for redundancy) inside your own cloud account. The MIT license makes this legally clean, and the cost is now comparable to one mid-level engineer, not a data-centre line item.
• •Multi-million queries per month: run the numbers on reserved GPU capacity. At that scale a dedicated deployment of V4-Pro can undercut every API on the list, and you gain latency control as a bonus.

Whichever path you take, the decision is not permanent. Design your application so the model sits behind an interface you control, and you can move between providers — or bring the model in-house — when the numbers shift again. And based on the last eighteen months, they will.

Need Help Deciding?

We help startups and mid-sized companies build, deploy, and cost-optimise AI features, from choosing the right model tier to running open-weight models in a compliant private environment. If you'd like a cost model built around your actual traffic and data requirements, book a consultation with our team and we'll walk through it together.

Pricing sources (September 2026): DeepSeek API pricing, OpenAI API pricing, Cloud GPU Rental Price Index, Thunder Compute B200 pricing, Jarvislabs H200 pricing, DeepSeek V4 deployment requirements. GPU prices are on-demand medians and vary by provider and region. All per-query figures assume 2,000 input and 1,000 output tokens.