Perspective · Jun 2026 · 8 min read

Open Source Stopped Waiting

For three years the open-source question was "can it catch up?" This month it changed. An open model beat a flagship frontier model on the work that pays the bills, at a fortieth of the cost. A systemically important bank started publishing its AI tooling in the open. The biggest buyer of frontier capacity began building its own. The gap didn't just narrow — the direction of travel reversed. Here's why open source is now the side accelerating, and what it means for where you build.

Teana Baker-Taylor
Teana Baker-Taylor
Chief Executive Officer

The gap closed to months, institutions started publishing, and the biggest buyers began building to get off frontier labs.

Most companies are building on the assumption that open source AI is the fast follower — useful someday, a year or so behind the frontier, safe to revisit later. That assumption was reasonable for a long time, and this month it became the risky one. The smartest-sounding position in AI just quietly turned into the wrong default, and a lot of roadmaps are still resting on it.

I held a version of that position myself: rent the best closed model, treat open weights as a someday option, revisit when it's ready. The reason I'm walking it back isn't that open source quietly closed the distance — it's done that gradually for a while. It's that in the space of two weeks the direction of travel became impossible to misread. The question stopped being "can open catch up?" and became "why are you still betting against the side that's accelerating?"

At BasedAI, we've spent two years building the unglamorous infrastructure that turns open models into dependable products, so I'll admit we're not neutral observers. But the receipts this month aren't ours, and they aren't close.

What just landed

Take the past two weeks on their own terms.

An open model beat a flagship frontier model on the work that actually pays. On June 12, Kimi K2.7 Code scored 81.1% on a standard tool-use benchmark, ahead of Claude Opus 4.8's 76.4% on the same test — at roughly $0.95 per million input tokens, a fraction of frontier pricing (devFlokers, June 2026). Tool use is not a parlor trick. It's the exact capability that decides whether an AI can read a system, take an action, and hand back finished work. Open won that round outright, at something like a fortieth of the cost.

The releases didn't stop there. The day after, Zhipu shipped GLM-5.2 under a permissive MIT license with a usable million-token context window, topping the open rankings (TechTimes, June 21). MiniMax M3 became the first open-weight model to combine frontier-tier software-engineering ability with a million-token window and native multimodal computer use. And NVIDIA — the company selling the shovels to everyone in this gold rush — released the Nemotron 3 family of open models, data, and libraries, betting its own developer story on open rather than closed (NVIDIA, June 2026).

Institutions stopped just consuming AI and started publishing it. Banco Santander — a global systemically important bank, the kind of institution that historically keeps everything behind its own walls — now runs a public open source organization on GitHub, releasing a vendor-neutral model client library, governance tooling for high-stakes model decisions, and fraud-detection benchmarks for the whole industry to use (SantanderAI on GitHub). When a bank's lawyers sign off on giving its AI tooling away, the calculus on open has changed.

And the largest buyers began building to depend less on the labs. In early June, Microsoft — OpenAI's closest partner and largest backer — launched its own in-house models explicitly to reduce its reliance on OpenAI and lower costs for developers (CNBC, June 2026). When the company with the deepest seat at the frontier table decides it needs an exit from single-vendor dependence, that is not a cost optimization. It's a tell.

Any one of these would be a data point. Four of them, in the same fortnight, pointing the same direction, is a trend line. The open side is the one accelerating.

The reframe: it was never "can open compete?"

The instinct is to score this as a benchmark race — open is up three points here, down a few there, and we tally it like a league table. That framing is comforting and it misses the story.

The real shift is structural. A closed frontier model has to earn back a training run that cost billions, and the only levers it has are price, terms, and access. An open model that's already trained carries none of that burden into its price, and once its weights are public, no one can reprice it, revoke it, or rewrite the terms underneath you. So when an open model reaches parity on a task, it doesn't reach a tie. It reaches a better deal — same finished work, a fraction of the cost, and an asset you can hold instead of a service you rent. Parity on capability plus an open license isn't an even match. It's a rout on everything except the headline benchmark.

That's the contrast worth sitting with: the frontier is racing to justify its price, and open source is racing to erase the reasons that price exists. Only one of those races compounds in the customer's favor.

Be precise about where closed still leads

It would be easy, and wrong, to claim open has won everything. It hasn't, and we should be precise, because precision is the whole brand.

On the hardest reasoning benchmarks — graduate-level science, frontier math, the genuinely difficult problems — the top closed models still hold a lead of roughly three to four months over the best open ones (Epoch AI, 2026). And in the enterprise, plenty of buyers are consolidating on a single closed vendor right now for reasons that have nothing to do with raw capability: governance, auditability, one throat to choke. Those are real advantages and I won't wave them away.

But notice what that lead is now measured in. Months, not years. And notice where it applies: the frontier of reasoning, not the floor of useful work. For the work most businesses actually run every day — reading, drafting, summarizing, using tools, carrying a workflow from start to finish — the gap isn't three months. It's gone. The decision about where to build shouldn't be governed by who wins the hardest math problem. It should be governed by who does your real work, reliably, at a price that doesn't assume a venture subsidy underneath it.

What this means if you're deciding where to build

If you're choosing a foundation for a product or a team this quarter, the month just rewrote your defaults.

The old default was to rent the best closed model and revisit open "later." The cost of that default used to be low, because open really was behind. It isn't low anymore. Build your critical path on a single closed model and you inherit a price you don't set, terms you don't write, and an access agreement a third party can end — a risk that stopped being hypothetical on June 12, when a government order forced one lab to switch off two flagship models for every customer overnight (TechTimes, June 21). Open weights, served on infrastructure you control, carry none of that exposure and now match the capability on the work that matters.

The catch — and it's the reason most teams rented in the first place — is that running open weights well is hard. Choosing the right model for each task, serving it reliably, keeping it current as a better one lands every few weeks: that's an infrastructure discipline, not a weekend project. The reason "rent closed" won for so long is that it was the only fast option.

What we're building for this moment

That's the gap BasedAI exists to close. BasedAI is the acceleration and commercialization layer for open source AI. We build products that turn the best of open source into reliable, useful work for real people and real businesses.

That work shows up in two products. Hirebase is live in closed beta today: AI coworkers you hire for a role, running on the best open weights, served on our infrastructure, with each task routed to the model that handles it best — so a small team gets open-weight economics and portability without ever managing a model. And BasedAPIs, coming soon, is the same capability for developers and enterprises that want to build directly: open models served reliably for production workloads, at open-weight cost. The accelerating supply of open models this month is exactly the raw material both products are built to turn into dependable work.

The direction of travel is the decision

There's a version of this argument that waits for open source to win outright before acting on it, and it will wait too long. The companies that took customer data seriously in 2014, before regulation forced everyone's hand, didn't have proof it would matter — they read the direction of travel and moved early. This is the same kind of moment. The benchmarks will keep trading points. The trend underneath them won't reverse: more capable open models, arriving faster, from more serious institutions, adopted by the biggest buyers in the market.

Open source stopped waiting. The only real question left is whether your roadmap is still built as if it didn't.

If you're working through what it would take to move real work onto a stack you actually control, that's the conversation we want to be having. Come find us.

---

BasedAI is the parent company of Hirebase and BasedAPIs. Written by Teana Baker-Taylor, CEO of BasedAI.

Sources