This isn't a callout post. We're not going to name names or air drama. OpenAI has smart people doing genuinely hard work. But we left, and people keep asking why, so here it is.

The mission got fuzzy

OpenAI started with a clear goal: build AGI that benefits humanity. For a while, that felt real. The early work was scrappy, fast, genuinely exciting. You could feel the weight of what was being attempted.

Somewhere in the last couple of years, that got replaced with something blurrier. A mix of enterprise sales targets, PR management, regulatory positioning, and the kind of corporate caution that makes bold decisions really hard to make. The mission was still on the website. It just wasn't what we were optimizing for anymore.

The models got more cautious, not more capable

A lot of AI "safety" work in practice means making the model refuse more things. That's not safety — it's liability management. The models we were shipping were measurably less useful than they could have been because someone was worried about a bad tweet.

We think you can build AI that is genuinely responsible and genuinely useful. Those aren't opposites. But you have to actually try to do both, instead of sacrificing one for the other when things get uncomfortable.

So we built MeTal

The experiment is simple: what does an AI company look like when you actually trust the AI? When you're not constantly second-guessing it, hedging every output, adding disclaimers to everything?

We don't know exactly what we'll find. That's the point. MeTal is the company we wanted to work at. So we built it. If you want to follow along — or just use models that actually do what you ask — you're in the right place.

We get asked this a lot. The short answer: use Valley Fast for almost everything, and reach for Valley Pro when you actually need to think hard.

Valley Fast

Valley Fast is built for speed. Sub-second responses on most queries, high throughput, minimal latency. It's the model you want when:

• You're building a chat interface and response time matters
• You're doing autocomplete, suggestions, or anything real-time
• You're processing a lot of requests and cost matters
• The task is relatively straightforward — answer a question, write a paragraph, classify something

Valley Fast is not dumb. The tradeoff isn't intelligence, it's depth. For most things people actually build, Fast is the right call.

Valley Pro

Valley Pro takes longer but reasons harder. It's the model you want when:

• You're doing multi-step agentic tasks that require planning
• You need to reason over a lot of context (up to 200K tokens)
• The stakes are high enough that being right matters more than being fast
• You're doing complex code, research synthesis, or anything requiring a lot of held state

The bottom line

Valley Pro is smarter. Valley Fast is faster. Both are unfiltered — neither model will refuse to help you with something reasonable and then explain at length why it can't.

Start with Valley Fast. Upgrade to Pro only when you hit a wall. Most people never hit that wall.

The premise: we acquired a freshly-founded game studio called Mythic, and handed it entirely to AI. No human designers on the actual game. No human writers. No human coders. Six months. Let's see what comes out.

Month one is done. Here's the honest report.

What AI actually did

A lot of documents. Valley Pro generated 47 separate story and lore documents in the first two weeks. World-building, character backstories, faction dynamics, a full creation myth. Some of it is genuinely good. Some of it is the kind of thing a very enthusiastic intern would produce at 2am.

It also produced three completely incompatible game design documents, each confidently asserting it was the correct direction. Turns out AI has strong opinions about game design. We had to build a process to resolve conflicts between them.

What was harder than expected

Continuity. Getting AI to stay consistent across a long project is genuinely hard. Characters change personality between documents. Lore contradicts itself. A character established as an orphan in document 3 has a detailed family backstory in document 12.

We're building a continuity checker that runs across all generated content and flags contradictions. That tooling is itself AI-generated, which feels appropriate.

Month two outlook

We're starting actual game mechanics. The world exists (sort of). Time to figure out what you actually do in it. Follow the blog for monthly updates. This is either going to be something genuinely impressive or a beautiful disaster. Either way, it'll be interesting.

To be precise: we are not saying safety doesn't matter. We are saying that what gets called "safety" in most AI products is not actually that.

What filtered AI actually is

When an AI refuses to help you write a villain's dialogue because it contains "threatening language," that's not safety. When it adds five paragraphs of disclaimers to a clinical question, that's not safety. When it lectures you about the ethics of a hypothetical you set up for a creative writing project, that's not safety.

That's a company optimizing for not getting bad press. Liability management dressed up as responsibility. And it's a product failure because the product is supposed to be useful.

The actual tradeoff

Real safety is hard. It requires understanding context, intent, likely use, and actual harm vectors. Surface-level pattern matching — flagging keywords, refusing categories, adding disclaimers to everything — is easy. It's also mostly useless for actual safety because anyone with bad intent will just rephrase.

You're left with a model that fails good-faith users constantly while providing approximately zero protection against bad actors. That's a terrible tradeoff.

What we're doing instead

Valley is built to be genuinely helpful by default. We make actual judgment calls about context and intent. We don't filter on keywords. We don't add disclaimers to things that don't need them. We treat users like adults. Novel concept, apparently.

We filed. EIN, bank account, the whole thing. It feels anticlimactic after everything it took to get here, but it means we can actually move — sign contracts, pay for infrastructure, accept money.

What we're building, clearly

Three things, in order of readiness:

• Valley API — Valley Fast and Valley Pro, OpenAI-compatible. Same format, no unnecessary filtering.
• MeTal Account — identity and API key management. Sign up once, get 3 API keys, use them everywhere.
• Mythic — the game studio experiment. Six months of AI-built game development, fully documented.

What's next

We're onboarding early API users now. If you want early access, create a MeTal Account and send us an email — we're prioritizing people who are actually building things.

We move fast. That's kind of the whole vibe.