If you search for whether open source AI is catching up to GPT and Claude, the honest answer changed twice in 2026 alone. Independent testing from Artificial Analysis found the gap between the best open weight models and the best closed models on its Intelligence Index shrank from about 13 points to roughly 6 points over the past year. A Moonshot AI model called Kimi K3 landed inside the top five of all 189 models the index tracks, ahead of everything except two closed models from OpenAI and one from Google. That has not happened before. But the story is not a clean win for open source, and the price side of the story is even bigger than the benchmark side.
Is open source AI catching up to GPT and Claude on benchmarks?
On raw intelligence scores, yes, and the pace is accelerating. Research firm SemiAnalysis tracked three separate eras of this catch-up. In 2024, it took Meta about eight months for Llama 3.1 to match GPT-4o level capability after starting far behind. In the reasoning-model era of 2024 to 2025, DeepSeek closed a 12-point gap with OpenAI o1-preview in roughly 8.5 months. By the most recent stretch, Moonshot AI's Kimi K2.6 passed Anthropic's Claude Opus 4.5 on the same tracked benchmarks in under five months. Each cycle has closed faster than the last.
The catch is that this progress is not even across every skill. A retrospective from Digital Applied comparing open and closed models through mid-2026 found open weight models actually beating or matching closed frontier models on coding benchmarks like LiveCodeBench and Codeforces, and DeepSeek's V4 Pro model scored a perfect result on a formal math reasoning test called Putnam-2025. But on general knowledge tests like MMLU-Pro and GPQA Diamond, the best open models still trailed Google's Gemini 3.1 Pro by 4 to 8 points, and on long-context retrieval tasks the gap against Anthropic's Claude Opus was closer to 9 points. So the gap is closing, but unevenly, and which model wins depends heavily on the task.
The real disruption is price, not benchmarks
Even people who do not follow AI benchmarks closely have felt this part. In May 2026, DeepSeek announced a permanent 75 percent price cut on its V4 Pro model, pricing it roughly 7 times cheaper than Anthropic's Claude Sonnet on input tokens and about 17 times cheaper on output tokens, according to reporting from VentureBeat. Its lighter V4 Flash model undercuts entry-level competitors by 10 to 25 times. This is not a temporary promotion, DeepSeek built the savings into the model's architecture itself, using memory-compression techniques that let V4 Pro process a million tokens of context using a fraction of the GPU memory that comparable closed models need.
That price collapse is changing how companies choose AI vendors in a very measurable way. Survey data cited by VentureBeat showed that cost per token jumped from being a top criterion for 25 percent of buyers in January 2026 to 37 percent by March, in just two months. Enterprises running their own open model deployments, or using specialized hosting providers built around open weights, are reporting operating cost cuts of 85 to 95 percent compared to relying on closed APIs alone. Some of that is offset by the engineering work needed to host and maintain your own models, which closed APIs handle for you, but for high-volume use cases the math is hard to ignore.
Adoption numbers tell a mixed story
Adoption data is where you have to be careful, because the numbers vary wildly depending on who is measuring and what counts as adoption. One estimate puts DeepSeek's enterprise API adoption at only about 1 percent at the end of 2025, rising to a median of roughly 4 percent of enterprise API calls among frontier models by the end of 2026, a real but modest slice. A separate figure often cited says over 26,000 companies have integrated DeepSeek's API into their products, and that 58 percent of new AI startups launched in 2025 included it somewhere in their stack, largely because it is cheap enough to use as a fallback or bulk-processing option even alongside a primary closed model.
Download and usage data from model-hosting platforms tells a clearer story about which open labs are actually winning developer attention. By March 2026, Alibaba's Qwen family had been downloaded roughly 942 million times cumulatively, more than double Meta's Llama at about 476 million. On OpenRouter, an API marketplace that routes requests across many models, DeepSeek alone served an estimated 14.37 trillion tokens between late 2024 and late 2025, well ahead of Qwen's 5.59 trillion and Llama's 3.96 trillion. Whatever the exact enterprise percentage, developers are clearly choosing Chinese open weight models over Meta's at a large and growing scale.
Meta started this race, and now it is not leading it
This is the part of the story that gets missed most often. Meta is the company that made open weight AI mainstream with the original Llama releases, and its "Behemoth" model, previewed in April 2025 as its most ambitious open release yet, has still not shipped publicly as of this year. Reports point to internal technical setbacks, including a training routing change that hurt the model's specialization and an attention mechanism that created blind spots in long documents.
Meta's response has been inconsistent. It pivoted toward a closed, proprietary model line called Muse Spark in April 2026, then reversed course again in August by open-sourcing a smaller 30-billion-parameter model called Muse Glimmer, while reportedly still considering keeping its next major model, code-named Avocado, closed with API-only access. Into the gap left by Meta's hesitation, Alibaba's Qwen team, Moonshot AI, Xiaomi's MiMo team, DeepSeek, and France's Mistral have all been shipping new open weight models on a near-monthly cadence. The company that started the open source AI movement is now arguably the most uncertain about whether to keep participating in it.
What this actually means if you are choosing a model
For most everyday and small business use cases, the practical question is no longer whether open models are "good enough," it is which specific open model fits which specific task, and whether you have the technical capacity to self-host or need a managed API. If your work is code generation, mathematical reasoning, or high-volume text processing where cost per token matters most, open weight models from DeepSeek or Qwen are now genuinely competitive with, and sometimes better than, closed alternatives. If your work depends on the most reliable general knowledge accuracy or very long document retrieval, closed models from OpenAI, Anthropic, and Google still hold a measurable, if shrinking, edge.
If you are trying to estimate what any of this will actually cost you before committing to a model or provider, our free AI Token Tracker tool can help you compare token pricing across providers so the decision is based on real numbers rather than marketing claims from either side.
The takeaway
The open versus closed AI debate used to be framed as a philosophical argument about transparency and control. In 2026 it has become a straightforward cost and capability calculation, and the labs driving open weight progress forward are increasingly not the ones people expected. Meta built the on-ramp and then hesitated at the merge. DeepSeek, Qwen, Kimi, and Mistral did not wait for permission to take the lane. The gap with GPT and Claude has not closed everywhere, but where it has closed, it closed fast, and the price gap closed even faster.
