Farewell, Fire Pass (for now) | Rafał Opiłowski

It’s over (for now). Fireworks AI’s second iteration of Fire Pass, an unlimited-token subscription for Kimi models, has been sunsetted. Version 3 is being prepared for the general public with no explicit ETA.

I was a beta tester from week one and watched it evolve from an experiment into a pivotal moment for my personal coding experience.

Here is what it felt like to use.

Phase 1: The $7/week wild west

Fireworks quietly launched a product page that made zero noise. Max Weinbach posted about it on X on March 27, 2026.

View source<br>Fireworks AI fire pass is so good

It's Kimi K2.5 Turbo right now at like 250 tok/s and idk what the limits are but it's HIGH

oh and free trial but $7 per week https://t.co/JKQbHCvS0A

Fireworks AI noticed:

View source<br>Show us what you can build. https://t.co/ah3ATdI20G

At the start, the rules were simple: $7 per week bought Kimi K2.5 Turbo with unlimited tokens and no caps. The only limitation was respecting the dynamic rate limit.

The speed was absurd; I measured 200-250 tokens per second through a router optimized for agentic coding. I plugged it into OpenCode and felt billing liberation . It felt the same as switching from a taxi to a monthly transit pass.

Phase 2: The $49/month pivot

By May 2026, Fire Pass had evolved due to demand. It became an invite-only subscription that included updated Kimi K2.6 Turbo . The price jumped from $24/month to $49/month.

Still, $49/month for unlimited usage of a 1T-parameter model was undervalued. The V1 pricing was an acquisition hook, and V2 was a stress test for sustainability.

What it felt like to use

I ran Fire Pass daily for three months, and a few things stood out.

Unlimited tokens and speed change behavior

When every response costs zero marginal dollars, you prompt differently. You explore more branches and ask more questions, and iterating becomes cheaper. The quality ceiling of the output matters less than the breadth of exploration the flat rate unlocks.

Kimi models handled real work

They handled routine work and scaffolding comfortably, and refactoring was straightforward too. They also managed complex multi-file orchestration and nuanced architectural reasoning. Their biggest strengths were atomic task splitting and clean context windows for each background task, and multimodality helped too.

It allowed me to bring the whole TypeScript monorepo from accumulated technical debt to a clean state within three months. You can read about it here.

Why is it being rebuilt from scratch?

Nobody at Fireworks has said why, but the signals were visible.

The unit economics never worked

Unlimited access to a 1.02T MoE model running on NVIDIA H100s for $49/month was generous. Agentic coding workloads are token-hungry because they use long context windows, multi-turn conversations, tool call loops, and code generation. A single power user could burn through something like 9 billion tokens per month. At that scale, the cost-to-revenue ratio was likely inverted.

Its biggest strength was also its biggest weakness

Relying only on dynamic HTTP 429 rate-limiting was a double-edged sword. There were no limits on concurrent requests. It allowed a lot of freedom, but it also made cost control and fair usage harder.

At one point, Fireworks had to email a group of beta testers to throttle their usage. They were using more concurrent requests than Fireworks considered non-commercial usage. That was affecting the experience for other users.

Fireworks never marketed Fire Pass aggressively. The docs labeled it “Early Access” and noted that features, availability, and pricing were subject to change. There were no blog posts and no launch announcements beyond the initial Kimi K2.5 blog. It felt like an internal experiment that accidentally escaped into the wild.

The addicting freedom

For three months, I coded without looking at a billing dashboard, and that freedom changed how I worked.

Flat-rate unlimited access to frontier-capable models is worth trying, but it cannot sustain a business that depends only on rate limits. Fire Pass demonstrated both.

Fireworks AI is not the only one experimenting with flat-rate unlimited access to LLMs. Umans AI and Wafer Pass are exploring similar models, and others are joining in. They are either selling out seats or converting into serverless pay-as-you-go services. The real test is whether Fire Pass V3 can sustain itself long term.

I was a Fire Pass beta subscriber from April to May 2026.