Mercury 3: Price, Speed, 1M Context and How to Use Inception's Diffusion LLM
Mercury 3 launched on October 19, 2026. It is a text-only diffusion LLM with a 1M-token context window, a claimed 1,068 output tokens per second on NVIDIA GPUs, and launch pricing of $0.10 per 1M input tokens and $0.35 per 1M output tokens. You can use it today through Inception's API, OpenRouter, Baseten or Vercel. aiwith and Router do not list Mercury 3, so an existing subscription or API key there will not unlock it.
Last checked: October 20, 2026. This is an access and pricing guide based on Inception's launch post, its model page and OpenRouter's model listing. It is not a hands-on benchmark review.
What is Mercury 3?
Mercury 3 is the flagship model from Inception Labs, built on a diffusion architecture instead of the autoregressive token-by-token approach used by GPT, Claude and most other LLMs. A diffusion LLM generates and refines many tokens in parallel before settling on an answer, which is where the speed advantage comes from.
Inception says Mercury 3 is more than 40x larger than Mercury 1 and that intelligence on its benchmark suite has risen about 3x from Mercury 1 (mini) to Mercury 3. The model takes text in and returns text out; there is no image input.
The company positions it for latency-sensitive work: browser-use agents, real-time search, interactive tutoring, voice and chat support. Several early-access partners quoted in the launch post describe using it as a drop-in replacement for slower models in those workloads.
How fast is Mercury 3?
The headline number is 1,068 tokens per second on NVIDIA GPUs, which Inception describes as 4-8x faster than comparable autoregressive models. Those are vendor figures measured on Inception's own serving stack. Throughput will vary by provider, batch size, prompt shape and how many tokens the model spends reasoning.
Two data points from partners in the launch post give a more concrete feel:
- Octen AI reports a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds for its search workload.
- TinyFish AI says Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in its web-agent harness, significantly more than GPT-6.1-Sol, using the same token volume and number of agent steps as a frontier model.
Both are customer testimonials, not independent benchmarks. OpenRouter lists the model with a 1M-token context window and a maximum output of 131,072 tokens.
What does Inception claim on benchmarks?
Inception says Mercury 3 is on par with or better than GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4-8x faster. The launch post presents the comparison as a chart; the company has not published a separate independent evaluation.
Treat the quality claim as vendor-reported. Benchmark definitions and harnesses change, and a model that wins on a chart can still lose on your specific prompts. If quality is the deciding factor, run your own eval set before switching production traffic.
Mercury 3 pricing
Inception lists these per-1M-token prices:
| Input | Cached input | Output | |
|---|---|---|---|
| Standard price | $0.20 | $0.02 | $0.70 |
| Launch price (50% off) | $0.10 | $0.01 | $0.35 |
At launch pricing, a simple mix of 1M input tokens and 1M output tokens costs about $0.45. At the standard price, the same mix costs about $0.90. Cached input is 90% off the input price, which matters for agent loops that resend long tool results and transcripts.
The 50% discount is a launch promotion, not the permanent list price. Inception has not stated an end date on the model page, so check the current price before you build a budget around it. OpenRouter's listing shows the same $0.10/$0.35 launch price and $0.01 cached-input price at the time of writing.
How do I use Mercury 3?
Pick the route that matches your stack:
- Inception API: create a key at platform.inceptionlabs.ai. The API is OpenAI-compatible, the base URL is
https://api.inceptionlabs.ai/v1, and the model ID ismercury-3. New API keys come with 100 million free tokens, which is enough to test a real workload rather than a few demo prompts. - OpenRouter: use the model ID
inception/mercury-3. OpenRouter lists support for tool calling, structured outputs and reasoning effort. - Baseten and Vercel: Inception says the model is also available through both platforms on launch day.
- Inception's own chat: try it at chat.inceptionlabs.ai before writing any code.
A minimal direct API call looks like this:
curl https://api.inceptionlabs.ai/v1/chat/completions \
-H "Authorization: Bearer $INCEPTION_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mercury-3",
"messages": [{"role": "user", "content": "Summarize this changelog."}],
"max_tokens": 1000
}'
One migration note: Inception says Mercury 2 and Mercury 2.5 are no longer available to new users. Existing customers with usage before October 19, 2026 keep access until November 15, 2026. If you were calling those model IDs, plan the move to mercury-3 before that cutoff.
Can I use Mercury 3 on aiwith or Router?
Not yet. Neither aiwith nor Router lists Mercury 3, so a subscription or key on either service does not give you access to it. If you want a fast, cost-optimized model inside those products today, they already carry close alternatives.
On aiwith, the two models Inception compares Mercury 3 against are available in the browser:
- GPT-6 on aiwith includes GPT-6 Luna, the cost-optimized member of OpenAI's GPT-6 family.
- Claude on aiwith includes Claude Haiku 5.5 for fast coding, long writing and file-heavy work.
On Router, these low-cost models are listed with a 1M-token context window and OpenAI-compatible endpoints:
| Router model ID | Input $/1M | Output $/1M | Cached input $/1M |
|---|---|---|---|
qwen/qwen3.8-flash | $0.80 | $2.70 | - |
glm/glm-5.3-flash | $0.80 | $2.80 | $0.23 |
xiaomi/mimo-v2.6-flash | $1.00 | $2.00 | - |
deepseek/deepseek-flash | $2.00 | $8.00 | $0.04 |
These are different models, not Mercury 3 under another name. Mercury 3's launch price is lower than all of them, and even its standard $0.20/$0.70 price is lower on both input and output. The trade-off is availability, provider maturity and the ecosystem around the model, not token price. If Mercury 3 is added to Router later, the model list is the authoritative place to check.
Should you switch to Mercury 3?
Mercury 3 is worth testing if your product is bottlenecked by generation latency: chat that feels sluggish, agents that wait on every step, voice interfaces where time to first answer matters, or high-volume classification and support replies. The 1M context and low launch price make it easy to run a realistic pilot.
Before you move production traffic, keep three caveats in mind:
- The benchmark comparison is vendor-reported. Test on your own prompts, tools and failure cases.
- Launch pricing is temporary. Budget from the standard $0.20/$0.70 rate, not the discounted one.
- It is a new model from a smaller lab. Check rate limits, data-retention terms and provider coverage for the region you serve.
If you need Mercury 3 specifically, start with Inception's API and its 100 million free tokens. If you want a fast model with an established product around it, the GPT-6 Luna and Claude Haiku 5.5 pages on aiwith and the low-cost models on Router are available now.