The Race to the Bottom: What Plunging LLM API Costs Mean for SaaS

Why aren't falling LLM API costs making SaaS companies richer?

Plunging API costs mostly do not help SaaS companies that rent a thin AI layer. Model providers are locked in a price war, with per-token costs down roughly 10x every year since GPT-3 launched. That sounds like great news for any SaaS company with an AI feature. It mostly isn't. Cheaper tokens erase the cost advantage that used to justify a markup, so the wrapper layer around the model is the part getting squeezed, not the part getting richer.

How fast are prices actually falling?

The numbers are not subtle. Andreessen Horowitz found that for a model of equivalent performance, inference cost fell by a factor of 1,000 between GPT-3's 2021 launch and late 2024, or roughly 10x per year. Epoch AI's independent analysis found the rate of decline varies by task, from 9x to 900x per year depending on the benchmark, with the fastest drops concentrated in the last year of their study.

2026 gave SaaS buyers two more data points to watch:

• OpenAI's GPT-5.6 family. OpenAI cut GPT-5.6 Luna's price by 80% and Terra's by 20% in July 2026, bringing Luna to $0.20 per million input tokens and $1.20 per million output tokens. Separately, GPT-5.6 Sol, the flagship model, has been running at a 50% promotional discount on OpenRouter and Vercel's AI Gateway, down to $2.50 input and $15 output per million tokens.
• Zhipu's GLM line. GLM-4.6, a model competitive on coding benchmarks, lists at $0.60 per million input tokens and $2.20 per million output tokens directly from Z.AI, and third-party routers have pushed effective GLM-4.6 pricing as low as $0.27 input.

Put those numbers next to GPT-4's 2023 launch price of $30 to $60 per million tokens and the trend is obvious: intelligence that used to cost real money now costs pocket change.

Why do cheaper tokens not mean fatter SaaS margins?

Here is the part that trips people up. If the model gets cheaper, shouldn't the wrapper's margin go up? In practice, no, for three reasons.

Everyone gets the discount at the same time. A price cut from OpenAI or Zhipu is not a competitive edge you earned. Every other company calling the same API gets it automatically. Pricing power that comes from a vendor's roadmap, not your own product, evaporates the moment a rival ships the same feature at the same lower cost. This is exactly the dynamic behind the thin wrapper panic among investors: applications that are just a UI on top of someone else's model have no moat once the model itself is commoditized.

Model prices dropping does not mean total AI spend drops. Cheaper tokens make more use cases economical, so usage grows to fill the gap. Companies that once budgeted for light AI assist now run agentic, multi-step workflows that burn far more tokens per task. Net effect: many SaaS companies report their AI infrastructure costs rising even as per-token prices fall, because volume outpaces the discount.

Wrapper economics were never that good to begin with. Traditional SaaS runs at 70 to 80% gross margin because serving one more user costs almost nothing. AI-native features do not get that for free. Analysis of AI-native product economics puts blended AI gross margins closer to 50 to 60%, with some AI-only products landing near 17% once every user interaction requires a live model call. Layering a basic AI feature onto an existing SaaS product typically costs 12 to 17 points of gross margin, turning an 80%-margin seat into something closer to 65%.

Who actually wins from cheap tokens?

The winners are not the thinnest wrappers. They are companies on either end of the spectrum.

Companies with a real moat above the model. Harvey, a legal AI product built entirely on other companies' foundation models, reached $100 million in annual recurring revenue about 28 months after launch by embedding deeply into legal workflows and charging $1,200 per lawyer per month, not by having a cheaper API bill. The wrapper label doesn't hurt them because the value is in the workflow integration and data, not the raw model call.

Companies that build and own the layer around the model. When token costs keep sliding, the fixed costs of running your own inference pipeline, prompt infrastructure, evaluation harness, and routing logic become the more attractive bet relative to paying a SaaS vendor's markup on top of the same falling API prices. This is the build-vs-buy calculus shifting in real time: as the underlying compute gets cheaper, the case for owning the orchestration layer instead of renting someone else's thin version of it gets stronger, not weaker. Teams thinking hard about that tradeoff, rather than defaulting to whatever tool showed up first in a Slack thread, are the ones positioned to capture the savings instead of just passing them through to a vendor's gross margin. That is the same logic behind treating your software stack as an asset like the software you build in-house with Remy, rather than a subscription you keep renewing regardless of what it actually costs to replicate.

Companies that route intelligently. OpenAI's own customer quotes show the pattern: Ramp uses GPT-5.6 Terra and Luna as defaults for coding and background agent automation specifically because they lead on cost-efficiency, while reserving the expensive flagship model for harder reasoning steps. Cognition's Devin Fusion mixes a cheaper model in as a pair programmer for routine work and saves the premium model for what actually needs it. The savings go to teams that treat model choice as an engineering decision, not a subscription default.

What is the pattern to watch?

None of this means AI features are a bad bet. It means the economics reward specificity and ownership over convenience. A generic chatbot wrapper competes purely on price, and price is falling toward zero for everyone at once. A tool that owns proprietary workflow data, is deeply embedded in a customer's process, or controls its own inference pipeline keeps the upside of falling compute costs instead of handing it back to competitors who copy the same feature a quarter later.