Mercury 3 Explained

Mercury 3 is Inception Labs' flagship diffusion language model, announced on October 19, 2026. Inception reports that the model generates text in parallel at 1,068 tokens per second, provides a 1M-token context window, and costs $0.20 per million input tokens and $0.70 per million output tokens, with a 50% discount at launch.

Unlike standard large language models (LLMs) that generate text left to right one token at a time, Mercury 3 generates tokens in parallel and refines them over a small number of steps. Inception released the model on its API platform, OpenRouter, Baseten, and Vercel for text-only workloads.

What Mercury 3 Is

Mercury 3 is a diffusion language model. Inception Labs announced the flagship system on October 19, 2026, describing it as the most capable diffusion-based language model released to date.

Mercury 3 does not write left to right one token at a time. It generates tokens in parallel and refines them over a small number of steps.

The model is strictly text-in and text-out. Inception's docs list tool calling and structured outputs as features.

• Context window of 1M tokens in chat, with a maximum completion size of 131,072 output tokens
• Reported speed of 1,068 tokens per second on NVIDIA GPUs
• Base pricing of $0.20 per million input tokens and $0.70 per million output tokens, discounted 50% at launch to $0.10 and $0.35
• API compatibility with OpenAI client libraries via the v1/chat/completions route

How Mercury 3 Differs From a Typical Language Model

Standard autoregressive models generate language sequentially.

Inception says Mercury 3 is more than 40x larger than Mercury 1 and "vastly more intelligent, but just as fast."

Inception states that the scaling laws behind autoregressive models also hold for text diffusion. Inception says intelligence across a range of standard benchmarks has risen 3x from Mercury 1 (mini) to Mercury 3. Inception has not published an exact parameter count.

Speed and Quality Claims, and What They Rest On

Inception claims that Mercury 3 matches or exceeds the quality of speed-optimized frontier models while generating text four to eight times faster. In its launch post, Inception specifically names GPT-6 Luna and Claude Haiku 5.5 as comparison targets, claiming Mercury 3 delivers equal or better output while producing 1,068 tokens per second on NVIDIA GPUs.

These claims are vendor-reported. Inception did not link a technical report or publish a detailed benchmark score table comparing model scores across standard academic evaluations at launch.

Inception did publish results from a 50-task web-agent evaluation using the Mind2Web benchmark inside a production harness from TinyFish AI. In that test, Inception reports that Mercury 3 completed 50% of the navigation tasks, which Inception says was significantly more than GPT-6.1-Sol, using the same token volume and number of agent steps as a frontier model at a fraction of the per-token cost.

Customer quotes on Inception's launch post describe speed in the customers' own tests. Octen AI founder Kuan Zou stated that Mercury 3 achieved a median generation time of 0.9 seconds and an end-to-end response time of 2.2 seconds in search workflows where retrieval takes 80 milliseconds. Michael Tong of RevisionDojo reported that hints, marked answers, and definitions return in approximately one second. Independent evaluations across public benchmark datasets will be required to confirm whether these quality claims hold outside vendor harnesses.

Target Workflows and Reasoning Controls

Inception's launch post gives four uses: browser-use agents, interactive education, real-time search agents and customer support. In Octen's search test, model generation rather than retrieval (about 80 ms) was the bottleneck, which is the case Inception's speed claim targets.

Lorikeet, which builds customer-support concierges for voice, chat and email, says in Inception's launch post that in early testing Mercury 3 beat other models on latency while also performing higher on evals and in production.

Setting reasoning_effort to instant is intended for voice, realtime turns, and high-volume automations. Setting reasoning_effort to high is intended for complex multi-step reasoning. If the parameter is omitted, Mercury 3 defaults to low, unlike older Mercury models that default to medium. Temperature settings are adjustable between 0.5 and 1.0.

How to Start Using Mercury 3

Mercury 3 is available on the Inception API Platform and through OpenRouter, Baseten and Vercel. Inception says the API is OpenAI compatible and "drops into your existing stack with no rewrites".

Every new account on the Inception platform receives a one-time grant of 100 million free tokens. This credit is shared across all models in the account rather than split per model. You can also test the model without code by visiting Inception's hosted chat interface.

• 1. Create an account on the Inception API Platform. Every new account includes a one-time credit of 100 million free tokens.
• 2. Point an OpenAI-compatible client at the base URL https://api.inceptionlabs.ai/v1.
• 3. Set the model parameter to mercury-3 on calls made to the v1/chat/completions endpoint.
• 4. Set reasoning_effort to instant, low, medium, or high based on whether your workload prioritizes absolute latency or complex planning.
• 5. Or use Mercury 3 through OpenRouter, Baseten or Vercel; each channel may charge different rates.

Pricing, Token Caching, and API Limits

Prompts using cached input tokens receive a 90% discount, lowering the cached input rate to $0.02 per million tokens.

At launch, Inception lists a 50% discount: $0.10 per million input tokens, $0.35 per million output tokens and $0.01 per million cached input tokens. Inception has not published an expiration date for this launch pricing. See the Mercury 3 pricing guide for the price breakdown.

As a concrete arithmetic example, processing 1,000,000 fresh input tokens and generating 1,000,000 output tokens costs $0.90 at standard rates ($0.20 plus $0.70) and $0.45 under the launch discount ($0.10 plus $0.35).

On the Free tier, accounts are capped at 1,000 requests, 1,000,000 input tokens, and 100,000 output tokens per minute. Pay As You Go accounts receive 3,000 requests, 3,000,000 input tokens, and 300,000 output tokens per minute. Requests exceeding these caps return HTTP 429 status codes, which Inception advises handling with exponential backoff.

Mercury 3 in the Inception Model Catalog

Mercury 2.5, released on September 8, 2026, serves as the previous-generation reasoning model. Inception reported Mercury 2.5 running at 1,107 tokens per second with a 260K-token context window at $0.20 input and $0.75 output per million tokens. The reported speed of 1,107 tokens per second for Mercury 2.5 is numerically higher than the 1,068 figure reported for Mercury 3, though Inception's posts do not indicate whether benchmark prompts or hardware test harnesses were identical.

The earlier Mercury 2 model, launched in February 2026, offers a 128K-token context window and costs $0.25 input and $0.75 output per million tokens. For deep dives on the earlier generation, see the Mercury 2 review, Mercury 2 benchmarks, and Mercury 2 API pricing guide.

Inception also offers specialized models. Mercury Voice is a diffusion model tuned for voice agents with a 128K-token context, priced at $0.40 input ($0.20 at launch) and $1.50 output ($0.75 at launch) per million tokens for enterprise accounts. Mercury Decide is a high-volume classification model billed exclusively on input at $0.02 per million tokens during launch ($0.04 regular), which Inception reports outperforms Jev on LegalBench and JevBench. Mercury Edit 2 provides code autocomplete over a 32K-token window.

For the previous generation side by side, see the Mercury 3 vs Mercury 2 comparison.

Who Should Evaluate Mercury 3

Engineering teams building low-latency agent loops, interactive tutoring products, conversational search pipelines, or high-volume customer service systems should evaluate Mercury 3 today. Inception reports 1,068 tokens per second and a 1M-token context for the model.

Teams should hold off on adopting Mercury 3 if their pipelines require vision inputs, as the model cannot process images. Organizations that require formal enterprise compliance documentation such as HIPAA or SOC 2 status should also pause until Inception publishes formal compliance status for the model.

Adoption decisions may shift once independent benchmarks are published, or if Inception changes the regular token price after the launch discount. To test the system immediately, register on the Inception API Platform, issue a test call to v1/chat/completions using your 100-million token credit, and time the responses against the model you would replace.