On October 19, 2026, Inception announced Mercury 3, the latest version of its diffusion-based large language model (LLM). The company says that despite having more than 40 times the total parameters of the original Mercury, it reaches a generation speed of 1,068 tokens per second on NVIDIA GPUs. The context length it can process at once has also grown from the previous generation's 260,000 tokens to 1 million tokens, and the model is available at half the regular price during the launch period. For AI agents that call a model repeatedly, faster generation could shorten processing time. However, overall application response speed and operating costs vary with usage conditions, so how far the stated performance carries over into real processing time and cost will be a key factor when deciding whether to adopt it.
Model size up more than 40x, with generation speed maintained
The defining feature of Mercury 3 is that it keeps text generation fast while greatly expanding the size of a diffusion language model. According to Inception's official announcement, total parameters have grown to more than 40 times those of the original Mercury, yet generation speed stays at roughly the same level as the original. That said, the specific parameter count has not been disclosed, and the 40x figure does not indicate how much the model's intelligence or accuracy has improved.
The autoregressive approach used by typical large language models generates tokens one at a time, in order, based on the tokens produced so far. The diffusion approach Inception has used from the start, by contrast, generates chunks of text in parallel and completes the final response by repeatedly refining it from an initial state. The basic idea behind the speedup is to remove the constraint of waiting for the previous token to be finalized before generating the next.
In this announcement, Inception explains that Mercury 3 shows text diffusion models can be scaled up. In other words, it claims that the speed advantage of parallel generation can be retained even as the model grows larger. However, this is a product announcement from the company, not independent testing that demonstrates large scale and high speed together for diffusion language models in general.
Context length has also expanded, from the 260,000 tokens supported by the previous Mercury 2.5 to 1 million tokens in Mercury 3. This opens up the possibility of feeding in long documents or multiple sources at once and processing them together. Note, however, that 1 million tokens is the upper limit of context the model can handle at one time; it does not mean the model can generate an answer of the same length. Whether the model can accurately retrieve the necessary information from a long input also needs to be verified for each actual use case.
Ahead of rivals on some benchmarks, but not all
Inception claims that Mercury 3 delivers quality exceeding GPT-6 Luna and Claude Haiku 5.5 at 4 to 8 times the generation speed. In the benchmark comparison table the company published, all three models were evaluated with reasoning set to "medium." Looking at individual items, though, Mercury 3 does not always post the top score, and the models differ in what they handle best.
| Evaluation item published by Inception | Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 |
|---|---|---|---|
| GPQA Diamond | 89.23 | 86.36 | 84.51 |
| IFBench | 76.84 | 61.82 | 54.65 |
| DeepSWE v1 | 46.31 | 37.76 | 51.62 |
| MRCR v2 | 75.50 | 62.72 | 78.40 |
Mercury 3 outscores the other two models on GPQA Diamond and IFBench, while Claude Haiku 5.5 records higher scores on DeepSWE v1 and MRCR v2. The results also show that fast generation and the ability to handle specific tasks correctly need to be evaluated separately. Moreover, these are figures published by Inception, not the results of verification by an independent organization. The comparison table does not give enough detail on conditions needed to reproduce the results, such as the number of trials or limits on tool use, so these numbers alone cannot establish superiority under different settings or in real-world work generally.
How much does 1,068 tokens per second shorten real response times?
A case study Inception presented from Octen AI offers a clue to how Mercury 3's generation speed affects real applications. According to the company's CEO, in a test using the Sense API, the median text generation time with Mercury 3 was 0.9 seconds, while the median end-to-end response time was 2.2 seconds. The search itself finished in about 80 milliseconds, and he explained that text generation by the model was the main bottleneck.
It is worth noting that the stated 1,068 tokens per second refers to the output token generation rate. It is not the time from sending a request to receiving the first token, nor the time for the whole process to complete. In real applications, search, external tool execution, and network communication are all part of processing time, so faster model generation does not necessarily shorten overall response time by the same proportion. The 2.2-second median at Octen AI is likewise a customer case under that company's test conditions and does not guarantee the same response speed in every application.
Still, for uses like AI agents that call a model repeatedly, faster generation could matter a great deal. An agent that reads the contents of a web page, decides on the next action, checks the result, and continues, for example, has to wait for the model's response at every step. If the time for each call can be reduced, the whole sequence of work can finish sooner, and more steps may be executed in the same amount of time.
Inception's announcement also presents TinyFish AI's evaluation of a web agent. According to the company's co-founder, when running 50 live Mind2Web navigation tasks with a production web agent, Mercury 3 completed 50% of them. He explained that token usage and the number of agent actions were matched with the frontier model used for comparison, and that Mercury 3 completed more tasks than GPT-6.1-Sol. However, the specific completion rate of the comparison model has not been disclosed, so the size of the improvement cannot be calculated. This result is also an evaluation on a specific set of 50 tasks and does not show that half of general browser operations can be automated.
Half price at launch, and a migration deadline for older models
Mercury 3 comes with a special launch price of 50% off the regular price. Inception's official models page lists the regular and discounted prices as follows, all per 1 million tokens.
| Type of processing | Regular price | Launch discount price |
|---|---|---|
| Input | $0.20 | $0.10 |
| Cached input | $0.02 | $0.01 |
| Output | $0.70 | $0.35 |
At launch pricing, processing 1 million input tokens and 1 million output tokens without using the cache costs $0.45 in total. At regular prices the same workload costs $0.90, so it is literally half price. This is a simple comparison assuming equal input and output usage, however; actual costs will vary with the respective token counts and cache usage. The calculation also excludes taxes and any fees set independently by external API distribution services.
Neither the announcement text nor the official models page states when the launch discount will end. For that reason, anyone assessing long-term operating costs should base estimates on the regular price. In particular, supporting a 1-million-token context does not mean you need to fill it to the limit every time. When processing long documents or large volumes of material, adjusting how much information is passed to the model appropriately is important for keeping input costs down.
Existing users should also check the migration deadline for older models. According to the official models page, Mercury 2 and Mercury 2.5 are no longer offered to new users. Existing customers with usage records before October 19, 2026, can continue to access them until November 15. Enterprise customers will continue to receive support according to their contracts, so November 15 should not be taken as the end-of-service date for every contract.
Mercury 3 is available through the Inception API Platform as well as OpenRouter, Baseten, and Vercel. It is compatible with the OpenAI API, and new API keys come with a free allowance of 100 million tokens. It is also listed in BenchLM's October 2026 release log under October 19, with links to the official announcement and models page. However, this is a record of model release information and does not mean the model's performance has been independently verified.
API compatibility makes it easier to build the model into existing applications, but switching models does not necessarily produce the same outputs or results. Developers using Mercury 2 or Mercury 2.5 in particular should, with the November 15 migration deadline in mind, check task completion rates in their actual AI agents, overall response times, and operating costs at regular prices. What matters in judging Mercury 3's value is not just the stated generation speed, but how quickly, reliably, and cheaply it can complete the same work.