Inception Labs launches Mercury 2.5, a faster and cheaper alternative to GPT-4o

Inception Labs has released Mercury 2.5, a large language model the startup says responds roughly twice as fast as OpenAI's GPT-4o while charging about a quarter as much. The model is live through the company's API at $0.25 per million input tokens and $1.25 per million output tokens, with a claimed median time-to-first-token of 0.3 seconds. Inception Labs, founded by former Google Brain researchers, is pitching the model squarely at developers building voice agents, coding assistants and live support tools where users feel every fraction of a second.

The speed claim rests on an architecture the company describes as combining sparse attention with a purpose-built inference engine. Inception Labs has not published the full technical details behind that design, nor has it released benchmark code, which leaves the performance numbers resting on the company's own measurements for now. Independent testing is the missing piece: without it, buyers are weighing a vendor's word against a track record they can check themselves.

Early developer reaction has been split. Some who ran the API reported responsive results and praised the latency in hands-on tests, while others flagged the absence of third-party verification. Several commenters pointed out that Mercury 2.5's scores on standard evaluations such as MMLU and HumanEval land close to GPT-4o rather than clearly above it, which suggests the pitch is really about cost and responsiveness, not a leap in raw accuracy. Inception Labs says a technical paper and open-source evaluation scripts are due in the coming weeks.

Even at parity on quality, the price gap is the headline. Mercury 2.5 also comes in below several open-source models once hosting costs are counted, a calculation that matters to teams that want proprietary-grade output without paying a premium. The release lands in a market where Anthropic, Meta and Google are all pushing inference costs down, and where cheaper tokens can turn previously uneconomical features — real-time translation, interactive tutoring — into shippable products.

The tradeoffs cut both ways. Inception Labs is young, with a smaller support organization and less enterprise infrastructure than OpenAI, and it has not disclosed compute capacity or uptime guarantees, both of which production teams will want before committing. Bigger rivals can also absorb price cuts to defend share, which puts pressure on a smaller lab to keep finding efficiency gains rather than compete on subsidy. What to watch next is whether the promised paper and evaluation scripts appear, and whether Mercury 2.5 holds its latency under sustained load rather than in controlled tests.