Inception Labs released Mercury 2.5 on September 8, calling it the largest diffusion language model the company has trained. The model runs at 1,107 tokens per second on commonly available Nvidia GPUs and ships with a 260,000-token context window, according to the company’s own announcement.
Mercury 2.5 is a diffusion language model, a design that departs from how most chatbots you use today actually work. A standard large language model, the kind behind ChatGPT or Claude, writes one token at a time in sequence, and each new token depends on every token written before it. That chain is exactly what limits how fast the model can go, no matter how much hardware you throw at it. A diffusion approach instead starts with a rough, noisy draft of an entire passage and refines the whole span at once across several parallel passes, closer to how an image-generation model sharpens a picture than to how a typist fills a page. Removing the token-by-token bottleneck is why Inception Labs can post a throughput figure this high and why that number, not the benchmark scores, is the part of this release worth paying attention to.
On quality, Inception Labs says Mercury 2.5 performs comparably to smaller, cost-optimized frontier systems: Claude Haiku 4.5, Gemini 3.5 Flash-Lite, and GPT-5.6 Luna running in its low-compute mode. Those are the company’s own comparisons, run on its own evaluation suite, and Inception Labs has not published independent third-party benchmark results alongside the release.
For workloads where every millisecond in the response loop is visible to a user, speed can matter more than a marginal quality edge. Voice agents, live search pipelines, and coding assistants that spawn subagents to plan, search, and check their own output all live or die on latency, not on winning a leaderboard by a few points. A model that is merely comparable in quality but several times faster than the alternative is not a worse product measured on the wrong axis. It is a different product built for a different constraint, and the buyer for that product is an engineering team, not a benchmark committee.
Inception Labs backed the release with production claims from two customers. OpenCall, which builds AI phone agents, said switching to Mercury cut its P99 response time from several minutes to about one second and its median response time from roughly 0.4 seconds to under 0.2 seconds. Augment Code, a coding-assistant maker, said moving its context-compaction step to Mercury cut latency by 82 percent, from about 150 seconds to 27 seconds, while cutting cost by 90 percent. Both figures come from the vendors themselves, relayed through Inception Labs’ own blog post, not from independent measurement.
Pricing carries a catch worth separating from the launch pitch. Mercury 2.5’s list price is $0.20 per million input tokens and $0.75 per million output tokens. The number Inception Labs is promoting, $0.04 per million input tokens and $0.15 per million output tokens, is an introductory discount of 80 percent that the company frames as a launch offer rather than the model’s standing price. Teams doing cost modeling for a 2026 budget should run the math against the $0.20 and $0.75 list figures, not the launch price, since Inception Labs gave no end date or guarantee for how long the discount holds.
Alongside Mercury 2.5, Inception Labs previewed two related products: Mercury Voice, a diffusion model tuned for voice agents with a sub-170-millisecond time-to-first-token, and Mercury Router, which classifies an incoming prompt and routes it to whichever open or closed model best balances quality, speed, and cost. Both remain in preview, with no pricing or general-availability date disclosed.
Engineering teams evaluating Mercury 2.5 for a latency-bound workload should benchmark it against their own production traffic rather than against Inception Labs’ comparison set, and price the decision at the $0.20/$0.75 list rate in case the discount lapses before a contract renews.
Inception Labs announced Mercury 2.5 in a company blog post published September 8, 2026.
