Mercury 3 targets latency-bound work such as voice agents, search pipelines and support bots, where a single request fans out into many model calls. It offers 4 reasoning-effort levels and replaces Mercury 2.5, but it takes text only and had no independent quality score at launch.
What changed
Mercury 3 is a diffusion language model from Inception, released on 19 October 2026, that accepts a 1M-token context window. It refines blocks of text in parallel instead of writing token by token, which Inception reports at 1,068 tokens per second on NVIDIA GPUs, and it takes text input only.
Where it sits
- $0.163/M$ per 1M tokensBlended price (3:1)Lower is better#9 / 88peer median $1.63/Mvendor price, checked by HokAI
Cheaper than 91% of the 88 GA models with a published price, and one of 5 whose vendor states it may train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Inception · Family: Mercury
Context window: 1,000,000 tokens · Max output: 131,072
Input modalities: text, tool-calls · Output: text, tool-calls
About Mercury 3
Mercury 3 is the flagship of the Mercury family from Inception, released on 19 October 2026 as the successor to Mercury 2.5. It is a diffusion language model (dLLM): rather than predicting one token after another, it drafts and revises blocks of text in parallel. Inception says it is more than 40x larger than Mercury 1, that intelligence across standard benchmarks has risen 3x since Mercury 1 (mini), and that the release shows text diffusion scales. The weights are closed and the model is sold as an API.
Inception's quality claim is that Mercury 3 matches or beats cost-optimized models such as GPT-6 Luna and Claude Haiku 5.5 while running 4 to 8 times faster. The comparison is a chart image in the launch post with no score table in text, so HokAI lists no benchmark scores for the model yet. Artificial Analysis had no Mercury 3 entry on 22 October 2026, and OpenRouter's benchmark tab showed only Design Arena Elo ratings for design tasks, which are not comparable with standard reasoning tests. Treat the quality claim as vendor-reported until an independent index publishes a score.
Speed is the selling point. Inception reports 1,068 tokens per second on NVIDIA GPUs. On 22 October 2026 OpenRouter showed Inception serving the model at about 257 tokens per second with 0.90 seconds of latency, and the launch post quotes Octen, a search customer, at a median of 0.9 seconds for generation and 2.2 seconds end to end. The gap between those figures comes from prompt length, reasoning effort and load, so time your own prompts before committing.
The context window grows from 260K tokens on Mercury 2.5 to 1M, and a single reply can run to 131,072 tokens. Input is text only. The chat endpoint follows the OpenAI format, so frameworks such as LangChain connect by changing the base URL. Among early-access partners, TinyFish reports that Mercury 3 completed half of 50 live Mind2Web navigation tasks in its production web agent, ahead of GPT-6.1 Sol on the same run.
Mercury 3 is available on the Inception API, OpenRouter, Baseten and Vercel. Mercury 2 and Mercury 2.5 are closed to new users, and existing customers keep access until 15 November 2026, so teams still on them have a deadline. One default changes in that move: Mercury 3 answers at low reasoning effort unless you ask otherwise, where the older models defaulted to medium. Choose another model if you need image input, repeatable seeded runs or an independent quality score; DiffusionGemma is the open-weights diffusion alternative. On HokAI it sits with the proprietary, text-input and 400K-plus context models, next to the other Inception models.
Pricing
HokAI stores the 50% launch rates ($0.10 input, $0.35 output, $0.01 cached) because the API bills them today, and list is double that, so expect the stored figures to rise when the discount ends. Accounts on pay-as-you-go get 3,000 requests, 3 million input tokens and 300,000 output tokens per minute, while accounts still on the credit tier get 1,000, 1 million and 100,000. Enterprise accounts negotiate custom limits and an SLA with Inception sales. For comparison, HokAI lists [Claude Haiku 5.5](/hub/models/claude-haiku-5.5) and [GPT-6 Luna](/hub/models/gpt-6-luna) at the same $0.10 input and $0.50 output, and [GPT-6.1 Sol](/hub/models/gpt-6.1-sol) at $2 input and $10 output.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0030 | $0.0003 | $0.0034 |
| Support reply | $0.0002 | $0.0001 | $0.0003 |
| One coding agent run | $0.020 | $0.0070 | $0.027 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Diffusion decoding: Generates blocks of tokens in parallel and refines them over several passes instead of writing left to right, and the API can stream blocks or show the denoising steps in diffusion mode.
- 1M context, 131,072-token output: Takes a 1M-token prompt, up from 260K on Mercury 2.5, and returns up to 131,072 tokens in one response.
- Four reasoning levels: The reasoning_effort parameter accepts instant, low, medium or high, and the model uses low when you omit it.
- Tool calls and JSON schemas: Supports parallel tool calls, a strict flag on tool schemas and JSON-schema structured output on the chat completions endpoint.
- Automatic prefix caching: Repeated prompt prefixes are cached with no request parameter, and usage.prompt_tokens_details.cached_tokens reports how many tokens came from cache.
- OpenAI-compatible API and SDKs: Accepts the OpenAI chat format and ships Python and TypeScript packages named inceptionai.
Pros
- A diffusion decoder removes the per-token wait of left-to-right models, which matters most when one task chains many sequential calls.
- Puts the Mercury line level with GPT-6 Luna and Claude Haiku 5.5 on context length, closing the gap that kept earlier Mercury models out of long-document work.
- Inception's launch post quotes four early-access teams in agents, tutoring, search and support describing a quick swap into an existing stack.
Cons
- Text input only, so image, audio and PDF workflows need a second model.
- Quality is vendor-reported: no independent index entry existed at launch, and the launch chart gives no scores in text.
- Model lifetimes are short: Mercury 2.5 launched on 8 September 2026 and is already closed to new users, so budget time for repeat migrations.
Frequently Asked Questions
What are Mercury 3's pricing plans in 2026?
Mercury 3 bills $0.10 per million input tokens, $0.35 per million output tokens and $0.01 per million cached input tokens today, a 50% launch discount on list prices of $0.20, $0.70 and $0.02. Reasoning tokens count as output, and Inception published no end date for the discount. New accounts receive a one-time credit of 100 million tokens, and Enterprise terms with custom limits go through sales.
Mercury 3 or Claude Haiku 5.5: which should you pick?
Inception says Mercury 3 outperforms both Claude Haiku 5.5 and GPT-6 Luna on several benchmarks while running 4 to 8 times faster, but its evidence is a chart in the launch post. Artificial Analysis does not list Mercury 3 yet (checked 22 October 2026), while HokAI's record gives Haiku 5.5 an Intelligence Index score of 43 and image input. Choose Haiku 5.5 when you need vision or a checkable quality score, and Mercury 3 when each call sits on the critical path of a live response.
Is Mercury 3 open source or proprietary?
Proprietary. Inception sells it as an API and publishes no downloadable weights, and the public API has no fine-tuning endpoint, so customization means contacting sales. You reach it through the Inception API, OpenRouter, Baseten or Vercel, and the terms of use let you use outputs for any purpose they do not restrict. DiffusionGemma is the nearest diffusion model with open weights.
Does Mercury 3 train on your prompts?
By default, yes: Inception's terms let it train on API submissions, and the opt-out is the 'Improve the model for everyone' switch in the API Platform's user settings. Its enterprise page says customer prompts and outputs are not used for training and that retention is configurable. No fixed retention period is published, and zero-retention setups are handled case by case through support.
Who should use Mercury 3, and who should skip it?
Use it for voice agents, search pipelines, browser agents and support bots where model latency compounds across many calls. Skip it for image or audio input, for evaluation runs that must repeat exactly, and for buyers who need an independent score before sign-off. Teams still on Mercury 2 or 2.5 should plan the move before 15 November 2026.
Top Alternatives
- GPT-6 Luna: Pick Mercury 3 for diffusion speed on text-only agents; pick GPT-6 Luna when you need image input or the lower list price.
- Claude Haiku 5.5: Pick Mercury 3 when generation speed is the constraint; pick Claude Haiku 5.5 for image input and a published independent score.
- Mercury 2.5: Pick Mercury 3 for any new build, since Mercury 2.5 is closed to new users and has the smaller context window.
- DiffusionGemma 26B-A4B: Pick Mercury 3 for a managed API with a 1M-token window; pick DiffusionGemma when you want open weights you can self-host.