Mercury Decide vs Celeris-1: Speed and Benchmarks

👁 120 reads

Mercury Decide from Inception and Celeris-1 Decision from Celeris are two new decision models from companies known for diffusion language models, which generate text in parallel rather than one token at a time. In the Artificial Analysis Decision Index dated October 13, 2026, they score almost the same: 52.7 and 52.3 points, ranking fifth and sixth out of 11. They differ in almost every other respect: Celeris-1 is the fastest model in the ranking at 63 ms, while Mercury Decide is the slowest at 500 ms - but it is cheap and available through OpenRouter, including a free option. For more on the category, see my guide to AI decision models.

Artificial Analysis benchmark comparison

Mercury Decide (Inception)

Inception develops Mercury, a family of diffusion language models that generate text in parallel instead of token by token. Mercury Decide launched on September 30, 2026. Its documentation does not describe its architecture, so it is not clear whether it is also a diffusion model. The documentation does confirm that it has:

• the endpoint POST https://api.inceptionlabs.ai/v1/decisions and model name mercury-decide;
• choice questions with 1-255 options and scores with 1-10 levels;
• no chat, streaming or tool calling - decisions only;
• a regular price of $0.04, currently promotional pricing of $0.02 per million input tokens, with no stated end date.

In the AA ranking, Mercury Decide ranks third for admitting uncertainty (32.0, behind Microsoft-Decision-1 and Jev), but has the weakest WildGuardTest classification score (88.0) and is the slowest model (500 ms). Inception promotes it as “the smartest decision model on OpenRouter,” based on the JevBench v1.4 test - it ranks fifth in the AA ranking.

Celeris-1 Decision

On October 8, 2026, Tom Hamer described Celeris-1 Decision as “a diffusion language model built for typed decisions.” Its main features include:

• the endpoint POST https://inference.celeris.ai/celeris-1-decision/v1/systemone, compatible with clients written for Jev, plus a second endpoint in the OpenAI style;
• an optional one-sentence explanation of the decision, using the x_celeris.explain parameter, at no extra cost - the only model in the ranking with this option;
• a median of 67 ms for one question and 81 ms for 50 questions in a single request, according to the provider (measured from AWS us-east-1);
• no open weights or fine-tuning.

The AA measurement confirms its speed: 63 ms per task, twice as fast as Jev. Celeris-1 also ties with GLiDE for the best tool-selection score. It has one serious weakness: its “I don’t know” score is 0.0, meaning it always guesses when it does not know the answer.

My Polish-language test: Mercury Decide only

Mercury Decide is available on OpenRouter, so I included it in my test on October 13: 50 Polish customer-service requests (choose a department from 7 options and set urgency) and 10 conversations in which an assistant had to choose a tool. Celeris-1 is not on OpenRouter, and its API requires an account, so I did not test it.

• Cheapest in my test: one million tasks of this type would cost about $4.4 at the promotional price. Mercury Decide processes Polish text efficiently, averaging 219 tokens per task.
• Slowest among decision models: median latency was 619 ms from Poland through OpenRouter, but still almost three times faster than Claude Haiku 5.5 (1,689 ms).
• Tool use: one error - after the customer supplied an order number in a second message, the model still wanted to ask a follow-up question.
• Confidence: when confidence was at least 0.9, it chose the correct department for 42 of 46 requests - a weaker filter than Jev or Decider.

Which model should you choose?

• Celeris-1 Decision - choose it when response time matters, such as making a decision during a voice conversation or in an agent loop, and a short explanation is useful. Be cautious with factual questions.
• Mercury Decide - choose it when cost and easy access matter (OpenRouter offers a free option for testing) and a 0.5-second decision is fast enough.
• For the best overall score, choose Jev 1.13; for the lowest cost at a similar score, choose Decider v1.1 27B.

Frequently asked questions

How much does Mercury Decide cost?

The regular price is $0.04 per million input tokens, with promotional pricing of $0.02 (as of October 13, 2026). OpenRouter also offers a free option: inception/mercury-decide:free.

How much does Celeris-1 Decision cost?

$0.04 per million input tokens; output is free. The provider estimates a cost of about $8 per million short decisions, while AA calculated $31.6 per million tasks in its dataset.

Which decision model is the fastest?

According to Artificial Analysis, Celeris-1 Decision, with a median of 63 ms per task. Mercury Decide is the slowest in the ranking at 500 ms.

Sources

• Inception: Mercury Decide documentation (checked October 13, 2026)
• OpenRouter: inception/mercury-decide
• Inception on X: Mercury Decide launch (September 30, 2026)
• Celeris: Celeris-1 Decision (October 8, 2026)
• Artificial Analysis Decision Index v1.0 (October 13, 2026)
• My test: October 13, 2026, 60 tasks through OpenRouter from Poland.

// read morePerplexity Decider v1.1 27B: Cost, Benchmarks and APIBenchmarksDecision Models vs Small LLMs: Cost, Speed and AccuracyBenchmarksLiquid AI d1: Open Decision Model, AA Benchmark, and PerformanceNews

Kinetyra