Inception's Mercury 2.5 shipped Monday and drops into any OpenAI compatible client, which is exactly why two of its behaviors get missed. It accepts a reasoning_effort parameter, so its speed is a dial rather than a fixed property, and its optional diffusing stream returns blocks of noisy tokens that refine in place instead of appending left to right. Set the first deliberately. Keep the second away from anything that parses.
Inception launched Mercury 2.5 on Monday, a diffusion language model it says reaches 1,107 tokens per second on widely available NVIDIA GPUs, with a 260K token context window. Standard pricing is 20 cents per million input tokens and 75 cents per million output tokens, with an 80 percent launch discount at 4 cents input and 15 cents output. The API lives at https://api.inceptionlabs.ai/v1, presents an OpenAI compatible interface, and the model identifier is mercury-2.5. That compatibility is the trap. Because it drops into existing code with a base URL change, most teams never read the documentation, and two behaviors that are genuinely different from an autoregressive model go unnoticed.
The first is reasoning_effort. Inception's quickstart shows it as a request parameter on mercury-2.5, with medium as the example value, alongside max_completion_tokens rather than the older max_tokens spelling. This matters because it means Mercury's headline throughput is not one number. It is a setting. If you benchmark whatever your client library sends by default and conclude the model is either fast or slow, you have measured a default, not the model. Set reasoning_effort explicitly on every call and record which value produced the latency you are quoting internally.
The second is the diffusing option. Inception documents that when it is enabled, the API streams blocks of noisy tokens that are steadily refined into the correct output. This is not a variation on token by token streaming. A diffusion model refines an entire sequence in parallel, so intermediate output is genuinely wrong text that becomes right text, in place. Any consumer written on the assumption that a stream is append only will produce garbage on screen or, worse, will parse an intermediate state as a final answer. Inception also notes that the noisy tokens returned when this option is enabled are not counted for billing purposes, which makes the refinement effect free to display but does not make it safe to parse.
The practical division is clean. Use diffusing only where a human is watching text form and the visible refinement is the point. Never use it on a path that extracts JSON, triggers a tool call, or feeds another model. For everything structured, leave it off and treat the response as an ordinary completion.
Inception gives new accounts 100 million free tokens without requiring payment details, and Mercury is also reachable through Baseten and OpenRouter. That is enough headroom to benchmark against your own traffic rather than a leaderboard, which is the only comparison worth making. Point it at the boring high volume work first: classification, extraction, routing, summarization, tool selection and context compaction. Those are the calls where latency and price dominate and where nobody can perceive the difference between an adequate answer and an excellent one. Then price the result at the standard 20 and 75 cent rates, not the launch discount, because the standard rate is what renewal looks like.
Try it today
- Point your existing OpenAI compatible client at https://api.inceptionlabs.ai/v1 and set model to mercury-2.5.
- Set reasoning_effort explicitly on every request instead of accepting a client library default, and log the value alongside your latency measurements.
- Use max_completion_tokens, which is the spelling in the current Mercury 2.5 quickstart, and set it rather than relying on an implicit cap.
- Leave the diffusing option off on any path that parses output, extracts JSON, or triggers a tool call. Enable it only in a human facing view where visible refinement is the intended effect.
- Benchmark on the 100 million free tokens using your real traffic mix, starting with classification, extraction, routing, summarization, tool selection and context compaction.
- Model the economics at the standard 20 cents input and 75 cents output rates rather than the 80 percent launch discount, then compare against what those same calls cost on your current model.