AI NewsInfrastructureAnnouncement
Isoquant launches on YC with 1.5 trillion tokens and 1,000 customers
Isoquant launched on Y Combinator as a self-optimizing inference cloud for open models. The company says it has served more than 1.5 trillion tokens and signed over 1,000 customers one month after launch.

Why it mattersTeams running agents on open models pay for every token. A provider that tunes the stack to each workload changes cost and speed with no code change.
A team running an agent on an open model pays for every token the agent reads and writes. Isoquant launched on Y Combinator with a product that aims at that bill. The company calls itself a self-optimizing inference cloud. It runs open models behind an API. Isoquant says auto-research agents tune the serving stack to each customer's work.
The launch page has 22 votes at the time of writing. Isoquant says it has served more than 1.5 trillion tokens since launch. It says it has signed more than 1,000 customers. It says it is growing at 40 percent week over week. The company says it launched about one month before the YC post. All four numbers are Isoquant's own.
What the product does
Isoquant offers an inference API for open-source models. The company names DeepSeek V4.1 Flash and GLM-5.3-Flash on its landing page. For GLM-5.3-Flash, Isoquant says the model returns the first token in 452 milliseconds. It says the model runs at 158.9 tokens a second. For DeepSeek V4.1 Flash, the page says the model reaches more than 600 tokens a second. The YC launch page says the service crosses 1,200 tokens a second. It says latency stays under 500 milliseconds. It says the service runs four times faster than other providers. Each figure is from Isoquant.
The company says it uses auto-research agents to tune the stack. Isoquant's YC page does not say which parts the agents rewrite. It does not say how the agents measure a win. It does not say how often the stack changes.
Four customer case studies
Isoquant's landing page lists four case studies. One names a banking customer. For that customer, Isoquant says accuracy rose from 74.7 percent to 91.8 percent. It says cost fell by 91.7 percent. Another customer reached 98.7 percent task success on an agent workflow. The landing page does not name the four customers. It does not state the task or the model they ran. It does not say who measured the numbers. Those are Isoquant's own figures.
The team
The YC launch page names two founders. Aniket Deshpande is the chief executive. The page says he worked on post-training, inference and agent orchestration at Meta Superintelligence Labs. Before that he worked at Instagram. Mugdhaa is the chief technology officer. The page says she led AI for developer productivity at Snowflake. Before that she worked on video playback at Twitch. The page says that system handled more than ten million concurrent requests. The page gives no last name for Mugdhaa.
Pricing starts at $0.035 per million input tokens, by the landing page. The company identifies as Agent FM Inc.
The third choice most teams never change
A team that runs an agent against an open model has three choices to make: the model, the prompt and the serving stack. Most teams tune the first two and leave the third one at the provider's default. Isoquant says the third one can be tuned by a different kind of agent. Isoquant says that change needs no edit to the team's application code.
Every figure above is Isoquant's own account. A team that cares should try the API with its own workload against its current provider. Measure the first token latency, the tokens per second and the dollar cost at the same accuracy. That is the only test that matters.
Source
- Isoquant: Self-optimizing inference cloud for 4x faster open models, Y Combinator Launches
- Isoquant, Isoquant
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


