Experiential Labs AI: Is GPT-6 Astra Still Free? What the Models Page Shows Now

Experiential Labs AI went from unknown to breakout search term in four days, and the reason is simple. On September 4, 2026, co-founder Silen Naihin posted that GPT-6 Astra is “insanely good, but expensive,” so his company was giving it away for free. The post drew thousands of likes. Meanwhile, paying ChatGPT users were still locked out of OpenAI’s staged rollout, so the startup’s platform page became their first stop. Here is what the company actually is, what the free offer covered, and what the platform shows now.

Update, September 8, 2026: the promo is over. The public list at platform.experientiallabs.ai/models now prices GPT-6 Astra at OpenAI’s full rate, $10 in and $50 out per million tokens, and the promo row is down to three free models. Details in the next section, plus a table of what the platform still lists.

Experiential Labs is a San Francisco startup in Y Combinator’s Summer 2026 batch. It has two employees and one product: an open source AI gateway. You point any OpenAI-compatible client at a single endpoint, swap in one key, and reach every model the company routes to. The tagline on its homepage says it plainly: every AI model, one key, zero markup.

The gateway sits in front of hosted models, your own provider keys and local models alike. According to its Y Combinator profile, the bigger idea is what happens after the traffic passes through. The company records those requests, builds task simulations from them, and trains a smaller model that the customer owns outright. Hence the name, not experimental labs AI as many people type it. The model learns from experience, meaning your traffic, rather than from a generic dataset.

In addition, the code is public. The repository carries an Apache-2.0 license, roughly 2,400 GitHub stars, and installs with a single pip command. In other words, anyone can run the gateway without paying the company a cent.

The search spike is about one page: the models list on the platform. OpenAI released GPT-6 Astra on September 3 at $10 per million input tokens and $50 per million output tokens, 2.5 times the rate of GPT-5.6 Sol. One day later, Experiential Labs put the same model on a promo row at $0.

On Product Hunt, the founders listed what was included. GPT-6 Astra, DeepSeek V4 Flash, Qwen 3.8 27B, GPT-5.6 Luna and Claude Fable 5.1 were all free through the end of launch week. The promo card Naihin posted says only “free for a limited time.” Notably, when a commenter asked about token limits, nobody answered.

As of September 7, the models page still showed Astra with a $0 output price. By the morning of September 8, it did not. The public list at platform.experientiallabs.ai/models now carries GPT-6 Astra at $10 per million input and $50 per million output, the same as OpenAI, with no promo tag. The Promotional row has shrunk to three models: Qwen3.8 27B, DeepSeek V4 Flash and GPT-5.6 Luna, all at $0 in and $0 out. Claude Fable 5.1 is back at $10 and $50 as well. So the Experiential Labs AI GPT-6 Astra giveaway lasted four days, with no published end date, cap or closing note.

One more search that keeps showing up is platform.experientiallabs.ai/overview. That is the dashboard you land on after signing in, where the API key, credits and usage live. Logged out, the URL simply redirects to the sign-in page, which is why so many people end up back on Google after clicking it.

The Experiential Labs AI models list is bigger than the promo row suggests. On September 8 the public page counts 761 models, sorted into a Promotional row, a Recommended set of twelve, and a full catalog by maker: 116 GPT variants, 111 Qwen, 90 Llama, 64 Gemini, 58 Microsoft, 52 Mistral, 36 DeepSeek and 33 Claude among them. Prices are provider list prices per million tokens, and the site shows the cheapest route for each model. Here is the part of that list people are actually checking.

Two details matter. The three free models are real, current releases rather than leftovers, and DeepSeek V4 Flash at $0 is the cheapest reasoning model on the page even at list price. In addition, GPT-6 Astra Pro is routed through OpenRouter rather than OpenAI directly, so its speed and uptime numbers come from a different pipe.

The giveaway landed at the exact moment OpenAI’s own rollout stumbled. Sam Altman apologized on September 4 for a “messy” launch that reached a handful of enterprise customers before any Pro or Plus subscriber. As a result, people went looking for side doors. Here is every route that outlets and creators verified in the first week, updated for the promo that has since closed.

Two things stand out. First, a $20 ChatGPT Plus plan is now the cheapest reliable route, because the rollout to Plus finished over the weekend. Second, with the Experiential Labs row closed, every remaining free path runs through a third party on a timer, so you are trusting that company with your prompts for as long as it lasts. If the point was coding, the tools in our roundup of the best vibe coding tools bundle Astra into a plan you already pay for, so there is no promo row to chase.

Of course, the obvious comparison is OpenRouter, the incumbent gateway. One Product Hunt commenter summed the pitch up as OpenRouter without the 5% markup. OpenRouter adds a fee when you buy credits. Experiential Labs says routed tokens stay at provider list price, with 0% on top. You can also connect your own OpenAI, Anthropic, Gemini, Azure, Bedrock or even OpenRouter keys for free pass-through.

The same one-bill idea exists for people who never open a terminal. Latenode wires 335+ models into a visual workflow builder with no API keys of your own, and its free plan gives you 10,000 CPU seconds and five live workflows a month without a card. It is a different product, an automation tool rather than a gateway, but it answers the same question: how do I use every model without paying every vendor.

So where does Experiential Labs AI make money? The homepage answers directly: the gateway is free, and the company earns on hosted inference and on Pro. A Dev Community breakdown published on September 6 calls it a telemetry flywheel. Every request flows through one trace format. Because of that, the platform can spot cache misses, flag batch-friendly work, and recommend a cheaper model per agent.

The end game is training. The company claims up to 97% cost reduction on specific tasks and backs a guaranteed 50% saving at equal or better quality with an SLA. However, those are vendor numbers, and the case studies behind them are the company’s own. Even so, the incentive is clear. Four days of free Astra was the cheapest customer acquisition a two-person team could buy.

Outside the promo row, pricing is simple and published. Credits cost the same across every provider, and the company bills one invoice for all of them.

Note, however, the catch on the free plan. It gives you 500 credits a month, but bring-your-own-key and local models only arrive on Pro. Therefore, anyone who wants the gateway purely as a router for keys they already pay for either upgrades or self-hosts. Self-hosting with local models also means a card that can run them, which is what our guide to the best GPU for AI is for.

Kion Fallah is the chief executive. He earned a machine learning PhD at Georgia Tech, then built simulation systems for autonomous vehicles as a staff research scientist at Waabi. Silen Naihin is the chief technology officer. He grew AutoGPT, the early autonomous agent project, to about 160,000 GitHub stars, and worked on AI for science at the Department of Energy.

That background explains the product. Both founders spent years turning raw logs into training signal, which is exactly what the gateway does with API traffic. Meanwhile, the Product Hunt launch, hunted by Y Combinator president Garry Tan, finished sixth for the day with 129 points. Modest numbers, until the Astra post turned the company into a trending query.

Google Trends shows the confusion in real time. Alongside the company name, the queries “experimental ai,” “kie ai” and “kei ai” all broke out in the same week. In fact, they are three different things.

Experiential means learning from experience, which is the founders’ point about traffic. It is not experimental AI, and it has nothing to do with Northeastern University’s Institute for Experiential AI, an academic center that shares the phrase. Kie AI, meanwhile, is a separate API reseller that also sells GPT-6 Astra access with credits for new users. If you followed a link to platform.xplabs.ai, that is the startup’s own short domain, not a copycat.

A gateway only matters once agents push traffic through it. So the natural next read is our guide to the top agentic AI frameworks that would sit on top of it. If the free Astra row is for coding, our roundup of the best AI coding tools shows where the model already lives.

Sources

Y Combinator, Product Hunt, Experiential Labs pricing, Experiential Labs models, GitHub, Silen Naihin on X, ZEFI on X, Unite.AI, Dataconomy, GitHub Changelog, Vercel, OpenRouter, Dev Community.