Mercury 3 Pricing: API Costs, Discounts, and Where to Access It
Mercury 3's per-token pricing, launch discount, context window, and where to try Inception's diffusion LLM, including OpenRouter and Vercel.
What does Mercury 3 cost to use?
Mercury 3, the diffusion language model from Inception Labs, launches at 10 cents per million input tokens and 35 cents per million output tokens, with that rate reflecting a 50% launch discount. Cached inputs are discounted a further 90%, which matters a lot for agentic workloads that resend the same context on every step. The model is fully OpenAI-compatible, so switching an existing integration over is largely a matter of changing the base URL.
TL;DR
- Launch pricing for Mercury 3 is 10 cents per million input tokens and 35 cents per million output tokens, discounted 50% off the standard rate.
- Cached input tokens get a 90% discount, which rewards agent setups that repeatedly resend the same context window.
- The core pitch is speed, not cost savings. At full price, Mercury 3 may end up priced closer to or above comparable models like GPT-6 Luna or Claude Haiku 5.5, so you’re paying for throughput.
- Throughput is the headline number: Inception reports over 1,000 tokens per second on Nvidia Blackwell hardware, several times faster than similarly sized autoregressive models.
- Access points include Inception’s own API, OpenRouter, Base Ten, and Vercel, plus a free tier for the companion model, Mercury Decide, on OpenRouter.
- Context window jumped to 1 million tokens, up from 260K in the prior Mercury release, and the API supports tool calling and structured outputs.
- Mercury Decide, the judgment-only companion model, is priced well below comparable “LLM-as-judge” setups, reportedly around a quarter of the cost of similar decision models.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
How is Mercury 3 priced compared to similar models?
Inception positions Mercury 3 against models like GPT-6 Luna and Claude Haiku 5.5 on intelligence benchmarks, and the launch pricing (10 cents in, 35 cents out per million tokens) sits in a comparable band once the discount is applied. The catch is that this is a temporary launch discount. Once it expires, Inception’s own framing suggests Mercury 3 could land somewhat above Luna or Haiku on a pure per-token basis.
That matters because the value proposition isn’t “cheaper tokens.” It’s fewer seconds of wall-clock time per response. Inception reports Mercury 3 generating over 1,000 tokens per second on Blackwell hardware, a multiple of what Luna and Haiku produce on comparable benchmarks. For latency-sensitive products, that speed difference can be worth paying a premium for, even if the headline per-token price isn’t the cheapest option in its class.
The 90% discount on cached inputs is the other lever worth paying attention to. Agentic systems that repeatedly resend long context (conversation history, retrieved documents, tool outputs) will see a meaningful chunk of their token spend hit that cached rate rather than the full input price, which changes the real-world economics compared to looking at the sticker price alone.
Where can you access Mercury 3?
Mercury 3 is available through several channels rather than being locked to a single provider:
- Inception’s own API, which is the primary access point and where the OpenAI-compatible endpoint lives.
- OpenRouter, useful for teams already routing between multiple model providers.
- Base Ten, a hosting option for teams running inference infrastructure.
- Vercel, which extends access to teams building within that deployment ecosystem.
Because the API is OpenAI-compatible, including support for tool calling and structured outputs, most existing integrations built against OpenAI-style endpoints can point at Mercury 3 with a base URL change rather than a rewrite.
What do you get for the price: context window and capabilities?
Mercury 3 ships with a 1 million token context window, a sizable jump from the 260K window on Inception’s previous Mercury release. Combined with tool calling and structured output support, that makes it usable for tasks like processing large codebases or long documents in a single pass rather than chunking them across multiple calls.
On benchmarks Inception has published, Mercury 3 performs well on reasoning, instruction following, and search-oriented tasks (including strong GPQA Diamond and Aether search results), while sitting a step behind top-tier coding models like Opus 5.5 on the harder end of the Aether coding benchmark. That puts it in a sensible spot as a default model for coding sub-agents or for workloads that spread across many parallel agents rather than one heavyweight reasoning pass.
What is Mercury Decide and how is it priced?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Mercury Decide is a separate, smaller model built specifically for yes/no or classification-style judgments rather than open-ended text generation. You pass it a state (which can be text or images) and a typed question, and it returns an answer plus a probability score in a single call, without generating any free text.
Independent evaluations cited by Inception (from Vals AI) show Mercury Decide matching comparable judgment models on claim verification tasks and outperforming them on LegalBench, while costing roughly a quarter as much per call. On Artificial Analysis’s Decision Index, it ranks third for low hallucination rate. Speed numbers from the same testing put a single judgment at roughly 0.16 seconds, with 64 parallel judgments completing in around 0.4 seconds, which suits use cases where one context needs to be checked against many questions at once.
At the time of recording, Mercury Decide had a free tier available on OpenRouter, making it an easy way to test the judgment-model pattern (fast, cheap system-1 style checks wrapped around a slower, more expensive system-2 model) without committing to paid API usage first.
Is Mercury 3 worth it for agentic workloads?
For workloads where latency compounds, long chains of sub-agent calls, recursive self-improvement loops, or real-time interfaces like smart glasses, the speed advantage can outweigh a higher per-token price. Inception’s own example involved a search product on smart glasses where the model’s generation time dropped to under a second, keeping pace with a search backend that was already responding in milliseconds. In that kind of setup, shaving response time matters more than shaving fractions of a cent off each token.
For simpler, single-call workloads where latency isn’t the bottleneck, the pricing math is less obviously in Mercury 3’s favor once the launch discount ends, since comparable autoregressive models may end up cheaper per token. The practical takeaway is to treat Mercury 3’s pricing as a speed premium rather than a cost-cutting move, and to weigh it against how much your specific workload benefits from faster generation and a larger context window.
Frequently Asked Questions
How much does Mercury 3 cost per million tokens?
At launch, Mercury 3 costs 10 cents per million input tokens and 35 cents per million output tokens, which includes a 50% launch discount. Cached input tokens are discounted an additional 90%.
Is Mercury 3 cheaper than Claude Haiku or GPT-6 Luna?
At the discounted launch price, Mercury 3 is competitive with those models. Once the discount ends, it may be priced somewhat higher per token, with the tradeoff being substantially faster generation speed rather than lower cost.
Where can I try Mercury 3 for free or test it before committing?
Mercury 3 itself is a paid API product across Inception’s own platform, OpenRouter, Base Ten, and Vercel. Its companion model, Mercury Decide, has had a free tier available on OpenRouter.
What context window does Mercury 3 support?
Mercury 3 supports a 1 million token context window, up from the 260K window on Inception’s prior Mercury model release.
Does switching to Mercury 3 require rewriting my application?
No. Mercury 3’s API is OpenAI-compatible, including support for tool calling and structured outputs, so most integrations can switch by changing the base URL rather than rewriting application logic.