I Spent $18 Testing Vercel's Bleeding Edge Features announced at Ship 2025

Five features announced. One actually ships.

Key Takeaways

• Timeouts are teachers Every failure reveals a constraint. Every constraint forces better design - but failure can be costly

• Models have personalities Under pressure, their true nature emerges. Use it.

• Architecture beats implementation Pseudocode delivers 10x more value at 1/10th the cost.

• Motion communicates state Async operations need visual grammar. Users understand waiting when they see progress.

• Ship the 20% One working feature beats five promises. Always.

• Public code pushes platforms Every workaround is a feature request with working proof.

• Consensus doesn't need consensus AI teams work best with constraints, not complete features.

I spent $18 to discover how constraints create clarity.

Following my consensus article about AI teams working together, I imagined the perfect orchestration. Three models building in harmony. Queues managing workflows. Sandboxes executing code. Microfrontends giving each AI its own runtime.

The final test: These AIs would collaborate to build the consensus product itself.

Vercel's AI Gateway announcements caught my eye, I discovered by shipping though that:

Beta now means 'we thought about it.'"

• AI Gateway: ✓ See below...

• Queues: × "Limited beta" (no access)

• Sandbox: × "Public beta" (broken links)

• Microfrontends: × Not available

• BotID: × Invite only

One working feature out of five.

But that one feature taught me everything.

"Life's meta: asking the models to build the medium in which they were operating."

Asked three models to build the very application they were running in. The AI showdown platform itself. Complete with model switching, streaming responses, cost tracking, judge evaluation.

Grok died at 28 seconds. Claude flatlined at 64. Gemini was still streaming at 179 when Vercel killed it.

Cost of failure: $0.73 for incomplete responses.

The models weren't failing. They were drowning in their own reflection.

// From full implementation to pseudocode
"Build the complete application with all features"
↓
"Generate pseudocode architecture only"

Changed one constraint. Changed everything.

Response times dropped from 180 seconds to 80. Token usage fell 85%. Success rate went from 0% to 100%.

"We've trained AI to be comprehensive when we need clarity."

Each model revealed its nature through constraint:

Grok streams fast, hallucinates confidently. Claude writes carefully, thinks deeply. Gemini considers everything, loves databases.

Token economics emerged:

• Complete applications: $0.50-2.00 per attempt

• Pseudocode architecture: $0.02-0.08

• Failed timeouts: Pure waste

So I built what I could with what worked.

Three models racing in parallel. Same prompt. Different approaches. A judge evaluating responses. Animations showing data in flight.

Not because it's useful. Because it reveals how AI thinks under pressure.

"Animations aren't decoration. They're grammar for async operations."

Paper planes show requests traveling. Pulsing cards indicate processing. Streaming text creates perceived speed.

Users understand waiting when they see motion.

The dream was models building together in sandboxes. The reality was models timing out alone.

But even this taught something valuable.

"Platform limitations reveal architectural truths."

• OpenRouter has 200+ models but no deployment story.

• Cloudflare or AWS Bedrock add further inspection and other capabilities including evals

Providers have other tricks under their sleeves like caching and smart routing.

• Requesty supports tools but adds complexity.

• Invariant Labs guards data flows that Vercel ignores completely.

The competition is far ahead in critical areas.

"Vercel has no guardrails. None."

While Invariant Labs offers contextual security:

// Invariant Labs security features
- PII detection and redaction
- Data flow monitoring
- Tool permission management
- Content filtering policies

Vercel offers... nothing. You're on your own.

No PII detection. No data flow control. No tool restrictions. No content filtering.

Just raw model access and a prayer.

For $10 experiments? Fine. For production with user data? Terrifying.

Yet Vercel won my usecase $18 later for one reason:

// Vercel AI Gateway simplicity
import { openai } from '@ai-sdk/openai';
// Just works. No setup.

Zero config. Zero new services. It just works.

For $10 experiments? Perfect. For production with PII? Add Invariant's guardrails.

Consensus doesn't require complete features. It requires clear constraints.

AI teams can work together. Just not with features that don't exist.

"Your timeout is someone else's breakthrough."

• Try it: https://vercelship25.vercel.app/ If credits exhausted: https://vercelship25.vercel.app/showdown-anime

• Fork it: https://github.com/ramakay/vercelship25

• Read consensus backstory: https://www.linkedin.com/feed/update/urn:li:activity:7343306764064284672/

Clone it. Improve it. My remaining budget is yours to experiment with.

The gap between roadmap and reality isn't empty space. It's where builders live.

I spent $18 to learn what no keynote would teach:

Ship beats promise. Constraint beats completeness. Motion beats waiting. Working beats perfect.

Every time.