Mercury 2.5

Mercury 2.5 is Inception's third production model and the most capable diffusion LLM the company has shipped, announced 2026-09-08. The architectural bet is the same one Inception has been making since its first Mercury release: for every other model on the market, tokens arrive one at a time; a diffusion language model drafts and refines many tokens in parallel. Inception says this is the largest diffusion language model ever trained, and it targets the workloads where latency is the product rather than a detail, such as search pipelines, voice agents, and coding subagents.

Model Specifications

The 260K context is mid-pack for 2026, where the frontier sits between 500K and 1M tokens. Inception is not competing on window size; it is competing on how fast a call returns.

Where the Speed Shows Up

A single search request routinely fans out into dozens of model calls: plan the search, rewrite the query, rerank, structure facts, summarize, then check the answer. Inception's pitch is that a diffusion model keeps all of that inside one user interaction.

The company published two production case studies with the launch. OpenCall, which builds AI phone agents, reported median model response latency near 170 milliseconds. Augment Code said moving context compaction to Mercury cut latency by 82 percent, from roughly 150 seconds to 27 seconds, while reducing cost by 90 percent.

Those are vendor-reported customer numbers, so treat them as directional. The architectural claim they support is more useful than the exact figures: the models that pay off are the ones you call most often, not the ones that write the final answer.

Pricing

Inception's launch post lists these rates. Note the gap between list price and the launch promotion.

Artificial Analysis, checked on 2026-09-24, lists Mercury 2.5 at $0.25 input and $0.75 output with a 90 percent cache discount, and measured 770.4 output tokens per second. The divergence between Inception's own 1,107 tokens per second claim and the third-party 770 measurement is a good reminder to benchmark on your own traffic.

What the Independent Scores Say

Artificial Analysis ranks Mercury 2.5 second of 175 models on speed and 21st on cost, but places it 91st of 175 on its Intelligence Index with a score of 12, which it calls below average with a median of 13. The summary is blunt: below average in intelligence, well priced for its class, notably fast, fairly concise.

That is the honest shape of the trade. Mercury 2.5 is not a frontier reasoning model, and anything in your stack that needs frontier judgment should stay on a frontier model. What Mercury buys you is throughput and cost on the high-volume, low-complexity calls that surround the expensive one.

Getting Started

1. Pick the calls in your pipeline that repeat most: routing, compaction, query rewriting, tool-search summaries, reranking.
2. Route those to Mercury 2.5 and leave the hard reasoning where it is.
3. Measure end-to-end latency and cost per task before and after, because the win comes from frequency, not from single-call quality.
4. Turn reasoning effort down for simple calls. Tunable reasoning is the difference between cheap and naive here.

Limitations

• Below-average intelligence index. It is a routing and throughput model, not a replacement for a frontier model.
• Text only. No image, audio, or video input.
• 260K context. Long-context work at 1M tokens needs another model.
• Vendor-reported speed and case studies. The 1,107 tokens per second figure is Inception's own; Artificial Analysis measured 770.
• Promotional pricing. The launch discount will not last forever, so model your unit economics on the list price.

FAQ

What is a diffusion LLM?

Instead of generating tokens strictly left to right, it produces and refines several tokens in parallel. That is the source of the speed and the token efficiency advantage.

Is Mercury 2.5 a replacement for a frontier model?

No. Independent scoring puts it well below the frontier tier on intelligence. Use it for the high-frequency calls around your frontier model.

Where can I use it?

Inception's own API, plus Baseten and OpenRouter. Inception also offers 100 million free tokens to try the API.

Alternatives

• GPT-6 Luna: the low-cost frontier-adjacent option when you need intelligence more than speed.
• GPT-6 Luna: a low-cost frontier-adjacent option when you need more intelligence than speed.
• Qwen3.8-Flash-Next: high-throughput alternative with a much larger context window.

Conclusion

Mercury 2.5 makes the case that the fastest model in your stack is often the one you call forty times per request, and at 770 to 1,107 tokens per second it is the fastest reasoning model currently on the market. Independent scoring is clear that it is not a frontier model, so the value is entirely in how you place it. Route the repetitive calls here, keep the judgment calls elsewhere, and measure the end-to-end pipeline rather than the single call.