If you build agents for a living, you’ve probably noticed two things this year. First, the frontier closed models keep getting better, at least on paper. Second, more and more devs are saying that the models they pay for feel worse in practice, and that the small open model they run themselves handles the boring 80% of their workload just fine.Open-source models are getting close to GPT-class models’ performance. But, this “close” depends on the task. For instance, when it comes to broad reasoning and coding, the difference is often just a few points. As for the hardest reasoning, multimodal breadth, and long-running agent work, the proprietary frontier models still win.What "GPT-5-Class" Means in 2026"GPT-5-class" is a moving target.Let’s go back for a second to illustrate everything properly. GPT-5 launched on August 7, 2025. Since then, OpenAI alone has shipped 5.1, 5.2, 5.4, 5.5 (codename "Spud", April 2026), and 5.6. On Artificial Analysis, the original GPT-5 (high) is now marked deprecated and scores below the current median, not because it got worse, but because the benchmark got harder and the field advanced.So, when someone asks whether open models have "caught up to GPT-5”, the better question would actually be “have they caught up to the current top proprietary reasoning tier?” That tier is actually a small cluster of models.Keep that in mind whenever you read a headline with a version number in it since that number will likely become stale by the next quarter.How Big Is The Trench?An image with the table below?But how do you measure a gap when every lab reports its own favorite benchmark? Well, you can pick one neutral composite index.The Artificial Analysis Intelligence Index combines 9 rigorous evaluations, including Humanity's Last Exam (HLE), GPQA Diamond, Terminal-Bench, and long-context reasoning:ModelTypeAA Intelligence IndexClaude Fable 5.1Proprietary66Kimi K3 (max)Open weights60GLM-5.3 (max)Open weights60Qwen3.8 (2.4T)Open weights58GPT-5 (high), originalProprietary (deprecated)35Source: Artificial Analysis leaderboard, September 2026. Scores change frequently; check the live page before quoting.The best open-weight model trails the best proprietary model by about six points. But this average glosses over some significant differences:Task AreaAre Open Models Competitive?NotesRetrieval, embeddings, rerankingYes, and often betterOpen models lead public retrieval leaderboards outright.Structured extraction, classification, OCRYesNarrow tasks favor small, cheap, fine-tunable models.General coding and reasoningMostly, within a few pointsThey are fine for the bulk of agent steps.Hardest math and science reasoningNo, the gap persistsFrontier still leads on FrontierMath, HLE-tier problems.Long-horizon agentic and multimodal breadthNo, the gap persistsFrontier leads on tool use, vision, and long tasks.⚠️ Something to keep in mind before you put your trust in a single number: The large majority of open-model benchmark scores are vendor self-reported, not independently verified, as the August 2026 Morph analysis has found. In fact, none of the SWE-bench Verified entries tracked at the time were listed as independently verified. Therefore, make sure to treat "we beat GPT-5.x" claims in a model release as a hypothesis that you should test using your own data, not an established fact.Where Open Models Already WinAn image with the 3-4 model costs breakdown?So, which parts of your stack are no longer worth paying per-token for? Right now, the obvious candidates are the retrieval layer and most narrow tasks. Embeddings and reranking are the clearest starting point.The Qwen3 embedding family shipped under Apache 2.0 with a real 32,768-token context and its 8B model entered the MTEB (Massive Text Embedding Benchmark) multilingual leaderboard at number one. Qwen3-Embedding-0.6B costs about $0.011 per million tokens to serve, a rounding error compared to hosted alternatives. If your agent uses retrieval-augmented generation (RAG), there’s little reason to outsource retrieval.Generation on narrow tasks is the second obvious case. Alibaba's Qwen3.8-27B (Apache 2.0) scores 52 on the Artificial Analysis Intelligence Index, quantizes to roughly 14 to 17 GB, and runs on a single 24 GB GPU. You can run Qwen3.8-27B on one card doing classification, extraction, routing, and routine drafting without a monthly API cost.When the frontier is a better choice: If the task needs the top reasoning tier and open alternatives fail your evaluation, use the frontier model. Swapping for a weaker one to save money only creates a false sense of “being economical” if task success drops. Measure first.For a task-by-task walkthrough, Superlinked‘s guide on what small open source models can handle is a good starting point.Free to Download Doesn’t Mean Free to RunThe open models now rivaling the frontier ones are ENORMOUS, and unfortunately, free weights don’t mean free GPUs. We wish…Most of the top open models today are mixture-of-experts (MoE) models, and their weights sit in VRAM (the memory on your GPUs). Running them yourself often requires a cluster, not a laptop:ModelRough Footprint to Self-HostRealitygpt-oss-20b16 GBRuns on consumer hardwareQwen3.8-27B14 to 17 GB (4-bit)Single 24 GB GPUgpt-oss-120b~80 GBSingle H100-class GPUDeepSeek V4 (Pro tier)Multi-GPU nodeReal cluster infrastructureKimi K3 (2.8T params)~1.4 TB VRAM8x B200 node, roughly $32K per monthFootprints are approximate 4-bit figures from public deployment notes. You’ll also want some headroom for the KV cache (the key-value cache), which grows with context length. The Kimi K3 cost is based on a 2026 self-host breakdown.And there you have it. The choice isn’t a free open model vs. an expensive API, but a per-token bill vs. a GPU bill + the operational overhead of running them yourself.The math can flip based on only one variable: utilization. Superlinked's own analysis on whether you should self-host inference is a great starting point if you’re struggling to make this call.Route by Task, Not One ModelA production agent can do loads of different things in a single run:Embed queriesRetrieve and rerank documentsExtract fieldsDrafts responsesHandle hard reasoning sometimesThese tasks vary in difficulty. Forcing all of them through the same frontier API can result in a large bill for the work that a smaller embedding model could have done.The solution is to route by task:Agent StepSensible DefaultEmbedding and rerankingSmall open model, self-hostedExtraction, classification, OCRSmall open model, self-hostedRoutine generation and draftingMid-size open modelHardest reasoning, rare and high-stakesFrontier APIYou keep the high-volume, low-difficulty work on cheap infrastructure you control, and you spend frontier tokens only where they can actually change the outcome. Superlinked's practical guide to choosing models for agents goes deeper into the topic.YOLOing It YourselfYou have done the evaluation and concluded that open models clear your bar for the retrieval, extraction, and routine-generation parts of your stack.To help you with encoders, rerankers, extractors, generators, and all the other puzzle pieces of a full agentic stack, Superlinked has built SIE, the Superlinked Inference Engine.SIE is an open-source inference engine that gives you a single endpoint for more than 100 open models covering embeddings, sparse retrieval, reranking, extraction, OCR, and open-model generation. You can run it on your own hardware, or if you prefer, on Superlinked cloud.Two things SIE doesn’t do, though:It does not make an open model smarter. It serves the model you’ve chosen. The capability gap measured in this article can’t be closed this way. So, if a task needs the frontier, SIE is not the answer.It’s not automatically cheaper. Below the utilization threshold, a hosted API is still a better option. SIE makes sense when GPUs stay busy.If all this sounds like your situation, the quick start is the fastest way to see whether SIE is the right fit for you. The SIE vs. OpenAI breakdown will walk you through the hybrid design (self-host the routine work, and keep the frontier API for the hard 10%).FAQHave open-source models caught up to GPT-5-class performance?Almost. On broad tasks, at least. As for the hardest reasoning, still perform better.What is the best open-source model in 2026?There’s no single best. Kimi K3, GLM-5.3, and Qwen3.8 rank highest on the neutral index, but the right pick depends on your own task, hardware, and license needs.Can I run a GPT-5-class open model on my own machine?If it’s a capable smaller model, yes. Qwen3.8-27B and gpt-oss-20b can be hosted on a single GPU, for example.Are self-hosting open models cheaper than a hosted API?Only at high, sustained utilization. Below that, a hosted API is usually cheaper, especially after you consider the cost of engineering and operations.