Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Founders choose multiple models instead of one at TechCrunch Disrupt 2026 - OpenSmartRoute
TechCrunch Disrupt 2026 announces a shift from single-model to multi-model strategies. Startups are moving away from picking just one AI model for everything. They are now building applications that use several different models together. This change happens because technology is evolving faster than before. Founders face new decisions about cost, performance, and flexibility.
Previously, companies often chose a single large model to run all their tasks. That approach worked when tools were simpler and less flexible. Today, AI products can call on different models for specific jobs. Some workloads need speed, while others need deep reasoning or low cost. A multi-model strategy lets teams match the right tool to the right task.
This shift means founders must manage more complex architectures. They need to decide which model handles image generation versus text analysis. They also must balance how much they pay for different services. The technology landscape is changing rapidly, offering new options every day. Companies that adapt quickly will have an advantage in the market.
The session "The Real Tokenmaxxing: How the Best AI Companies Navigate a Multi-Model World" brings together experts to discuss these choices. Speakers include Mo Jomaa from CapitalG and Vipul Ved Prakash from Together AI. Zuzanna Stamirowska, CEO of Pathway, also joins this conversation on the Builder's Stage. They explore why companies use multiple models in their products.
They will explain how teams balance cost against performance and flexibility. The discussion covers when open models can outperform proprietary alternatives. Open models are software that anyone can download and modify freely. Proprietary models are closed systems controlled by a single company like Nvidia or Google.
For founders, this flexibility affects more than just model performance metrics. It influences operating costs and how product teams make strategic decisions. Companies can take advantage of better models as soon as they emerge from the market. Waiting for one perfect model might cost too much in time and money.
Choosing between open and proprietary models assumes a company needs to pick one initially. Increasingly, AI products can call on different models for different jobs simultaneously. A single API endpoint cannot always meet all the diverse needs of a modern application. Teams need a strategy that allows them to switch tools as requirements change.
Mirror Particle raises capital to create an AI engine that simulates changing human motivations. The company rejects large language models in favor of a foundation model trained on longitudinal data.
Mistral launched Mistral Large 4, nicknamed Le Chonk. It is a 1 trillion-parameter model available for free.
How much of your AI stack should you own? Model choice raises a bigger question about ownership. Manos Koukoumidis, CEO and co-founder of Oumi, addresses this on the Real World AI Stage. His session is titled "Which AI Should Your Company Actually Deploy: Rent, Customize, or Build." He tackles whether startups should build their own models from scratch.
He also discusses when customization can become a competitive advantage for a business. The conversation helps attendees evaluate frontier APIs and customized open weights. Open weights are model parameters that are publicly available for modification by users. Frontier APIs refer to the most advanced and cutting-edge AI services currently offered by vendors.
Through audience polls and startup scenarios, Koukoumidis provides a practical framework for decision-making. Attendees will leave with three decision principles they can use in future architecture conversations. Building more of the AI stack offers greater control and product differentiation. However, it also requires more time, talent, and significant financial resources.
Relying on existing models might be faster but offers less control over the underlying code. The choice founders make shapes both the final product and the business behind it. Companies that own their infrastructure can differentiate themselves from competitors who rent services. But they must invest heavily in engineering and maintenance teams.
For some startups, the decision still comes down to trade-offs between open and proprietary models. Nader Khalil, Director of Developer Tech at Nvidia, joins Sydney Sykes from Nvidia for a discussion on this topic. Their session is called "Building AI Startups Worth Betting On" and takes place on the Builders Stage.
They examine what founders are choosing today regarding model architecture. The trade-offs between frontier APIs and open-weight models have clear business consequences. Model choice shapes costs, infrastructure requirements, and product control levels. It determines where a startup can create differentiation that competitors cannot easily replicate.
Those trade-offs affect everything from initial development costs to long-term scalability plans. Some proprietary models offer superior performance but come with high licensing fees. Open models are cheaper but may require more engineering effort to optimize for specific tasks. Founders must weigh these factors carefully when planning their product roadmap.
Model architecture is only part of the equation in building an AI system. AI performance ultimately depends on the hardware underneath it, and advances in AI are changing how that hardware gets designed. Anna Goldie, founder and CEO of Ricursive Intelligence, leads a session on this critical link. Her co-speaker is Azalia Mirhoseini, founder and CTO of the same company.
Their session "When AI Starts Designing Its Own Hardware" explores how AI optimizes chips and hardware systems. They discuss why model architecture and hardware are becoming more closely connected in modern systems. An increasingly open AI ecosystem could mean significant changes for the infrastructure underneath it all.
For founders, the connection between AI and hardware affects how quickly new capabilities reach the market. Faster chip development could change the infrastructure available to startups building next-generation products. Traditional semiconductor design cycles are taking years to complete now. AI-driven design promises to shorten these timelines dramatically.
Get your ticket to Disrupt to hear what happens when AI starts helping design the hardware that powers it. Learn how to keep your AI options open at Disrupt 2026. These conversations are part of over 200 sessions across six industry stages. The event includes roundtables, breakouts, and networking opportunities for attendees.
More than 10,000 founders, investors, operators, and tech leaders are expected to attend this year. There will be over 250 speakers and 300 exhibiting startups presenting their work. Matchmaking and dealmaking sessions give attendees chances to connect with potential partners. These connections help builders face many of the same technology and business decisions daily.
A startup may use a frontier API today, customize an open model tomorrow, or move workloads among several models over time. Preserving flexibility could be as important as choosing the right model now. Companies need strategies that allow them to adapt as the technology evolves.
Secure your pass to TechCrunch Disrupt 2026 and hear how AI leaders are approaching these choices. Save up to $100 on your pass and get a second pass at 50% off. If you have been affected by a layoff, grab your Expo+ Pass for just $75.
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence or the quality of the content provided. The Disrupt experience is meant to be shared with colleagues and peers. Bring a partner at 50% off to cover more ground during the event.
Covering more ground means building momentum and discovering what's next in the startup ecosystem. Federal judge calls Flock 'indiscriminate mass surveillance' while Amazon responds to data center backlash. Meta wants your next gadget to be Muse-infused with new features. Google thinks SpaceX's Starship has to launch 1,800 times before space data centers get off the ground.
World's first enhanced geothermal power plant completed in just 23 months recently. Google releases Gemini 4 Argon, called its most powerful model yet by the company. The Pentagon taps Elon Musk and Palmer Luckey to help decide what the military should do next. These headlines show how fast the technology landscape is moving right now.
Why it matters
Multi-model strategies reduce risk and lower costs for AI startups compared to single-model approaches. Companies can optimize spending by using cheaper models for simple tasks and expensive ones for complex work. Flexibility allows teams to adopt new tools quickly without rewriting entire product architectures. Owning parts of the stack provides control over data privacy and model behavior. Hardware advances driven by AI could accelerate innovation cycles for all builders.
What to do
Start evaluating which models fit your specific workload needs before committing to a single vendor. Compare open weights against proprietary APIs based on your budget and customization requirements. Attend sessions at TechCrunch Disrupt 2026 to hear direct advice from industry leaders like Manos Koukoumidis. Research how AI is currently designing hardware to understand future infrastructure constraints. Build a flexible architecture that can switch between models as technology improves over time.
TechCrunch Disrupt 2026 announces a shift from single-model to multi-model strategies.
The event moves past the idea of picking just one AI model for everything.
Founders now see value in using different models for different jobs.
This approach replaces the old rule of committing to a single architecture.
Companies can mix open weights with proprietary APIs in one product.
Some startups are building tools that call several models at once.
This gives teams more ways to build new software features.
It also creates harder choices about where to spend money and resources.
Speakers discuss why companies use multiple models for different jobs.
Mo Jomaa, Vipul Ved Prakash, and Zuzanna Stamirowska lead this conversation.
They explain that no single model fits every possible task perfectly.
A cheap model might handle simple questions without costing much money.
An expensive frontier model could solve a complex problem with high accuracy.
Using both lets teams balance cost against performance needs daily.
Open models can outperform proprietary ones in specific niche areas sometimes.
Flexibility becomes more important than raw speed or size alone.
Teams avoid locking themselves into a technology that might become obsolete soon.
They keep the option to switch vendors if prices rise or quality drops.
Manos Koukoumidis explains how much of the AI stack companies should own.
He argues that owning everything is not always the best path forward.
Startups must decide whether to rent models, customize them, or build from scratch.
Building a model requires significant time and specialized talent for training.
Renting via an API offers speed but less control over data privacy.
Customizing open weights provides a middle ground with some flexibility.
Koukoumidis will share a framework for making these ownership decisions.
Attendees can learn to evaluate frontier APIs against customized open models easily.
The choice shapes both the final product and the long-term business model.
More control means differentiation, but it also demands more internal resources.
Nader Khalil and Sydney Sykes compare frontier APIs against open-weight models.
They represent Nvidia's perspective on what builders are choosing today.
Khalil leads Developer Tech while Sykes manages VC partnerships at the company.
They discuss the trade-offs between relying on closed systems versus open code.
Frontier APIs offer stability but can be expensive for large-scale deployments.
Open-weight models allow full customization but require significant engineering effort.
These choices directly affect product strategy and how competitors view your startup.
Differentiation comes from unique capabilities that others cannot easily replicate.
A proprietary model might lock you into a specific vendor's roadmap.
Open weights let you integrate the model with any existing software stack.
Ricursive Intelligence founders explore how AI designs its own hardware.
Anna Goldie and Azalia Mirhoseini lead a session on this emerging trend.
They explain that AI is now optimizing chip architecture for better performance.
Model requirements are driving changes in how silicon gets designed today.
Hardware and software are becoming more closely connected than before.
This connection could change the infrastructure available to new startups quickly.
Faster chip development might shorten the time to market for new features.
The open AI ecosystem could mean more diverse hardware options soon.
Founders need to understand these physical constraints when planning their products.
Why it matters
Multi-model strategies reduce risk and lower costs for AI startups compared to single-model approaches. Companies can optimize spending by using cheaper models for simple tasks and expensive ones for complex work. Flexibility allows teams to adopt new tools quickly without rewriting entire product architectures. Owning parts of the stack provides control over data privacy and model behavior. Hardware advances driven by AI could accelerate innovation cycles for all builders.
What to do
Start evaluating which models fit your specific workload needs before committing to a single vendor. Compare open weights against proprietary APIs based on your budget and customization requirements. Attend sessions at TechCrunch Disrupt 2026 to hear direct advice from industry leaders like Manos Koukoumidis. Research how AI is currently designing hardware to understand future infrastructure constraints. Build a flexible architecture that can switch between models as technology improves over time.
