Foundra
Product8 min readNov 20, 2026
ByFoundra Editorial Team

Fireworks Says Most Startups Can Skip Pricey AI Models. When?

Fireworks AI cofounder Chenyu Zhao says open models trail the frontier by three to six months, and that the switch makes sense around $10,000 to $100,000 a month in token spend. Here is how to use that rule, how to split work across two models, and what the Uber breach says about picking the vendor that runs them.

Fireworks Says Most Startups Can Skip Pricey AI Models. When?

On the day Anthropic went public, one of the companies that competes for the same AI budgets gave founders some advice that sounds almost rebellious: you probably do not need the most expensive model for most of what your product does.

The advice came from Chenyu Zhao, cofounder of Fireworks AI, in an on-stage interview at the Slush conference in Helsinki that TechCrunch published on November 19. Fireworks runs open-weight models for other businesses and helps them tune those models on their own data. Its named customers include Uber, Shopify and Notion. Investors valued it at $17.5 billion this summer.

Zhao has an obvious interest in this view. He sells the alternative. But his rule of thumb is specific enough to test against your own numbers, and the same week brought a sharp lesson about the other half of the decision: the company you hand your data to.

Here is how to think about both.

What Zhao actually said

Three claims from the interview matter for an early-stage team.

  • Open models are close behind. Zhao said open models are at most six months behind frontier models, and probably closer to three.
  • Scale decides the switch. His rule of thumb: once a company passes roughly 10,000 to 100,000 users and spends $10,000 to $100,000 a month on tokens, it is time to consider open source.
  • The future is two models, not one. He expects a "two-model world," where a frontier model handles high-level planning and cheaper models do the routine work.

Notice what he did not say. He did not tell a pre-product-market-fit startup to spend weeks tuning open models. Before you know what users want, the frontier model is the cheap option, because your scarcest resource is engineering time, not tokens.

Why "three to six months behind" is good news and a trap

If open models trail the frontier by a few months, then the task your product needed a frontier model for last spring may run fine on an open model today. That is the good news. Most product features do not need the newest reasoning ability. Classifying a support ticket, pulling fields out of an invoice, rewriting a sentence in a brand voice: these are routine jobs.

The trap is assuming the gap is the same for every task. A model that scores close on public tests can still fail on the narrow, messy inputs your users send. The only benchmark that counts is your own:

  1. Pull 200 to 500 real inputs from your logs (with personal data removed).
  2. Save the answers your current model gave, and mark which ones were right.
  3. Run the same inputs through the cheaper candidate.
  4. Compare accuracy, speed and cost per correct answer, not cost per token.

If the cheaper model matches on 95% of cases and the misses are easy to catch, you have a candidate for the routine lane. If it misses on the cases that make customers angry, you do not.

How to split work in a two-model setup

Zhao's two-model idea maps to a simple routing pattern most small teams can build.

Lane one: the planner. A frontier model handles the hard, rare, high-stakes calls. Deciding which steps an agent should take, writing a first draft of something complex, handling an input your system has never seen.

Lane two: the worker. A cheaper open model handles the high-volume, well-defined steps. Formatting, extraction, tagging, summarizing a known document type, checking an output against a rule.

The router. A small piece of logic, sometimes just a few if-statements, decides which lane a request goes to. Start with rules you can read ("anything with a file over 20 pages goes to the planner"). Add a fallback: if the worker's answer fails a check, send the request to the planner.

The point is not to be clever. The point is that 80% of your calls may be cheap work wearing an expensive model.

Stop reading. Start building.

Your AI co-founder is ready when you are.

Foundra turns everything in this article into an actual plan. Validation, customers, pricing, launch. In one place, in your voice, in an afternoon.

Get started→

$39/month. Cancel anytime.

Run the numbers on your own bill

Here is an illustration with made-up round numbers, not a quote from any provider.

Say your product makes 2 million model calls a month. Today all of them go to a frontier model and your bill is $40,000 a month. After testing, you find 75% of calls are routine and a smaller open model handles them at one tenth of the cost per call.

  • Planner lane: 25% of calls at the old cost, about $10,000.
  • Worker lane: 75% of calls at one tenth the cost, about $3,000.
  • New total: about $13,000 a month, a saving near $27,000 a month.

Now subtract what it costs to get there: engineering weeks to build the router and test set, plus ongoing time to re-test when models change. If that work costs two engineers a month, the switch pays back fast at $40,000 a month. At $2,000 a month, it never would. That is why Zhao's threshold sits where it does.

Before you move anything, put the before-and-after into your financial model so the savings show up in your runway math, not just in a Slack message. A plain spreadsheet works, and so do planning tools like Foundra that tie cost assumptions to your burn and runway.

The same week: a vendor breach reached 149,346 Uber drivers

Cost is only half the choice. The other half is who runs the model for you.

TechCrunch reported on November 16 that Uber notified close to 150,000 drivers that hackers stole their names, addresses, birth dates and driver's license numbers. Uber told Oregon's attorney general it notified 149,346 people. According to BleepingComputer, Uber said a service provider told it on October 9 that an unauthorized party had compromised the provider's internal credentials, and that the attacker used them to reach Uber data connected to that provider.

TechCrunch identified the provider as Fireworks AI. Uber's spokesperson described the notices as relating to an incident at a third-party vendor, Fireworks.ai. Uber said it had alerts turned on, but because the attackers used the vendor's own credentials, no alerts were triggered.

In the November 19 interview, TechCrunch noted Fireworks disclosed two incidents this fall: one where hackers reached API keys and tokens stored as secrets for a limited number of accounts, and one where attackers got into a development system through its cloud provider and took customer data, including billing records and some datasets. Notion said chat transcripts from fewer than 1% of its workspaces were exposed and that it disconnected from Fireworks.

Zhao's explanation for Notion is worth studying. Fireworks promises zero data retention for live inference traffic. Notion, he said, used batch prediction, where uploaded datasets are stored at rest. Same vendor, two features, two different data promises.

A vendor checklist before you move a single workload

None of this means open models are unsafe or that hosted providers are worse than doing it yourself. It means the cost switch comes with homework. Before you route customer data to any AI vendor, get written answers to these:

  1. Which features keep data at rest? Ask about live inference, batch jobs, fine-tuning uploads and logs separately. "Zero retention" may apply to only one of them.
  2. How long is uploaded data kept, and can you delete it on demand?
  3. How are your API keys stored, and can you scope them? Use separate keys per environment and per feature, with spending caps. The OWASP Top 10 for LLM applications lists unbounded consumption as a risk category, and stolen keys are a direct path to it. Fireworks attributed its attack to a group that sells stolen model access.
  4. Who gets told, and how fast, if something breaks? Uber learned on October 9 and confirmed access to its data by October 13, per its statement.
  5. Can you cut the connection in minutes? Uber said it immediately severed data connections. Know where your own switch is.

Then do the boring parts on your side. Rotate keys every quarter. Strip personal data you do not need before it leaves your system. Keep a list of every AI vendor that touches customer data, and put it in your security page so customers can see it.

What to do this month

If you are pre-launch or pre-product-market fit, keep using the best model you can afford and spend your time on users. Log every call with its cost and task type so the switch is easy later.

If your token bill has crossed five figures a month, run the 200-input test on your single highest-volume task. Pick one routine job, try one open model, and measure cost per correct answer.

Whichever stage you are at, send the five vendor questions above to every AI provider you use, including the frontier labs. The answers will tell you as much as the price sheet.

Frequently asked questions

What is an open-weight model? It is an AI model whose trained weights are published, so companies can run it on their own servers or through a hosting provider and tune it on their own data.

When should a startup consider switching from a frontier model? Fireworks cofounder Chenyu Zhao suggests considering it once you pass roughly 10,000 to 100,000 users and $10,000 to $100,000 a month in token spend. Below that, engineering time usually costs more than the savings.

How far behind are open models? Zhao said at most six months behind frontier models, probably closer to three. Test on your own data, because the gap varies by task.

What is a two-model setup? A frontier model handles planning and hard cases, and a cheaper model handles routine, high-volume steps. A simple router decides which model gets each request.

What happened with Uber and Fireworks? Uber notified 149,346 people that names, addresses, birth dates and license numbers were stolen after a vendor's internal credentials were compromised. TechCrunch identified the vendor as Fireworks AI.

Does a vendor breach mean I should avoid hosted AI providers? No. It means you should ask how each feature stores data, scope and rotate your keys, and know how to disconnect quickly.

#product#AI costs#open models#open-weight models#Fireworks AI#unit economics#vendor risk#security#LLMjacking#first-time founders
The shortcut that 1,000+ founders took

You just read the theory. Ready to build the thing?

Foundra is your AI co-founder. It turns an idea into a validated business plan, a go-to-market, and your first 10 customers. In an afternoon, not a semester.

$39/month. Cancel anytime. Works in 20 languages.

Related reads

Product8 min

The AI-Code Cleanup Bill: How First-Time Founders Avoid the $50K-$500K Rebuild in 2026

Around 8,000 startups that shipped AI-built products now face rebuilds costing $50K to $500K each, and 40% to 62% of AI-generated code carries security flaws. The speed that got you a demo can become a liability. Here is how a non-technical first-time founder uses AI coding without building a trap.

Product11 min

Vibe Coding Is Over: What Karpathy's Agentic Engineering Pivot Means for First-Time Founders Shipping AI Products in May 2026

Andrej Karpathy, who coined the term vibe coding in early 2025, declared it passé in February 2026 and rebranded the future as agentic engineering. With 25 percent of YC Winter 2025 startups shipping codebases that were 95 percent AI-generated and AI co-authored code carrying 1.7 times more major issues than human-written code, the first-time founder building an AI product in May 2026 has to choose between speed and shippability. Here is how to do both.

Product11 min

The Big Tech Coding Pivot: What Microsoft's MAI-Code-1-Flash and Google's Gemini 3.5 Flash Mean for First-Time Founders Building Dev Tools in June 2026

On June 2, 2026, Microsoft shipped MAI-Code-1-Flash inside GitHub Copilot, two weeks after Google released Gemini 3.5 Flash at I/O. Both hyperscalers now have first-party coding models in the same price tier as Anthropic's Claude Haiku and OpenAI's GPT-5 mini. The math on building a dev tool startup in 2026 changed inside a 14-day window.

Product8 min

When Your Vibe-Coded MVP Hits the Wall: A Founder's Rescue Playbook

Vibe coding helped thousands of non-technical founders ship fast. Now a lot of those products are breaking under real usage. Here is how to tell whether you need a refactor, a rewrite, or a rescue.

Key terms

Related guides