For most of the GPT era, open models were a respectable second place. Capable, community-built, often faster, but trailing on the hardest benchmarks and on the long tail of real-world tasks that decide whether a product works.
That era is over.
The shift did not happen with a single release. It happened across roughly 18 months, with five compounding forces arriving in the same window. Once they did, the parity question stopped being interesting and the production question took over. This post is our attempt to be specific about what changed, what the evidence actually shows, and what it means for the teams building on top.
What the benchmarks actually say
A year ago, the standard line was "open is 6–12 months behind closed, sometimes a bit closer on narrow tasks." That line is no longer accurate.
The MMLU gap between the leading open model and the leading closed model narrowed from 17.5 points in 2024 to 0.3 points in 2025. On reasoning suites like GPQA and MATH, open weight families now trade the lead with closed peers on a monthly cadence. Coding benchmarks like SWE-bench and HumanEval show the same pattern: Llama 4, Qwen 3.5, GLM-5, and DeepSeek R1 each lead or match GPT-class performance on subsets of the eval, with the closed frontier still ahead on a handful of frontier reasoning tasks.
For the 80–90% of real-world enterprise use cases, the meaningful answer is now: pick the model that fits the workload. Open models deliver equivalent results at a fraction of the API cost, and the cost gap has widened, not narrowed, as the quality gap closed.
What changed: five forces, in order
1. Training recipes leaked, in the good sense.
The single most underrated story of the last two years is the slow diffusion of the recipes that produce frontier behaviour. Better data mixtures, better synthetic data pipelines, better RLHF, better preference modelling, better tool-use traces. None of this was secret in the academic sense. It just took time for the patterns from closed lab papers and leaked notes to be reimplemented, tuned, and proven on open infrastructure.
By mid-2025, the methodology gap was effectively gone. Open labs were not "catching up to" closed labs. They were running variants of the same techniques.
2. Mixture-of-experts collapsed the deployment economics.
The architectural shift from dense transformers to mixture-of-experts changed who could run a frontier model.
Llama 4 Scout activates 17B parameters per token from a 109B total. Llama 4 Maverick activates 17B from 400B. Mistral Large 3 follows the same pattern. The practical consequence is that a single modern GPU can serve a model that would have required a small cluster two years ago. Inference latency goes down, throughput goes up, and the cost-per-token at the edge of the curve goes from "API-only" to "self-host viable."
3. Licences stopped being a research project for legal.
Apache 2.0 and MIT licences now cover the majority of leading open weight models. The bespoke "research only" and "non-commercial" terms that used to make enterprise procurement nervous have largely disappeared from the top tier.
This is small in the technical story and enormous in the buying story. Procurement, security, and compliance reviews now treat the leading open models the way they treat any other Apache-licensed dependency. No carve-outs, no usage caveats, no separate negotiation.
4. Six labs is enough to be a market, not a project.
The open weight frontier is no longer one company plus a long tail.
Six major labs now ship competitive open weight families: Meta with Llama 4, Mistral with Small 4 and Large 3, Alibaba with Qwen 3.6 Plus, Google with Gemma 4, Zhipu AI with GLM-5, and OpenAI itself with gpt-oss-120b. Each is shipping on a quarterly or faster cadence. Each has at least one tier in the leading benchmarks. Each is supported by an inference ecosystem of providers, fine-tuners, and hardware partners.
Markets behave differently from projects. They have price discovery. They have compatibility pressure. They have a steady supply of newer, better releases. The closed labs are now competing against a market, not against a community.
5. Hardware caught up to the workload.
Hardware accelerated alongside the models. Newer-generation accelerators, better serving stacks, faster KV-cache management, speculative decoding, and continuous batching combined to drop per-token cost across the open inference ecosystem. The same generation that made closed inference cheaper made open inference cheaper too, and the open side had more room to absorb those gains because its margins were not propping up a research lab's revenue line.
What "parity" does and does not mean
Two clarifications, because the word does a lot of work.
Parity is not "open is better at everything." It is not. On the absolute frontier of multi-step reasoning, agentic chain-of-thought, and certain vision-language compositions, the leading closed models still have a measurable edge. When that edge translates to a real-world product difference, the closed option is the right call.
Parity also is not "open is better on every cost dimension." Self-hosting is not free. Open weight inference at production quality requires the same operational surface that any other tier-1 service requires: autoscaling, observability, eval harnesses, structured outputs, tool calling, safety filters, regional failover. Build all of that yourself and you may not save money. Buy it from a provider that runs open weights at scale and you do.
What parity does mean is that for the vast majority of production workloads, the choice between an open weight model and a closed API is now a commercial and strategic decision, not a quality compromise. That is the meaningful change.
What this means for the products being built
If you are building on AI today, the consequence of parity is that you should design your stack as if the model layer is a commodity, not a choice.
Choose providers that give you a single API across multiple open weight models. Build your evals against your own workload, not against generic benchmarks, and rerun them against new releases as they ship. Treat the model the way you treat a database: a swappable, benchmarked component that lives behind a clean interface. The teams that win the next cycle will be the ones that can switch backends in a week, not a quarter.
For enterprises specifically, this also means the conversation with security, legal, and compliance is no longer about whether to use open weight models. It is about which inference partner can offer the contractual and operational surface that closed APIs have spent four years polishing.
This is the gap we are building BasedAI to close. BasedAPIs gives you a drop-in OpenAI-compatible API across the leading open weight models, at margin pricing, with the production reliability and operational surface enterprises actually need. The commercial layer on top of those models is where the next decade of software gets built, and we want to be the layer that makes it boring to ship.
The era of open weight as second-best is over. The era of open weight as substrate has begun.