The LLM landscape in 2026 looks nothing like it did two years ago. Back then the assumption was simple: if you wanted the best model, you paid OpenAI or Anthropic, and that was that. Open source models were a respectable second tier, good for experimentation, fine-tuning, and budget workloads, but not quite there for serious production use.

That assumption is now outdated. Models like DeepSeek V3, Kimi K2, and GLM-4.6 have forced a real reckoning. The gap in raw intelligence has closed dramatically, the price gap has widened in open source’s favor, and the speed story depends heavily on who is doing the inference.

This article breaks it down: what closed source models actually give you, where open source has caught up, how the pricing compares across tiers, and what all of this looks like when you run these models on DeepInfra.

What Open and Closed Source Actually Mean Today

Closed source means the weights are proprietary and inaccessible. You call a hosted API and pay per token. GPT-5, Claude 4, and Gemini 2.5 Pro all fall into this category. You never see the weights and interact purely through an endpoint the provider controls entirely.

Open source (or more accurately, open weight) means the model weights are publicly released and can be downloaded, self-hosted, fine-tuned, and quantized. Llama 4, DeepSeek V3, Qwen3, Kimi K2, and GLM-4.6 all fall here. Open weight is the more precise term since licenses vary and not every model ships full training data or code, but the key point is that the weights exist and you can run them yourself.

The practical consequence is that open weight models can be hosted by third-party inference providers, which is exactly what DeepInfra does. You get the flexibility of open source without needing to manage GPU infrastructure yourself.

Intelligence: How Close Is the Gap?

For most of 2023 and 2024, the intelligence gap was real and meaningful. GPT-4 sat noticeably above open source alternatives across reasoning, coding, and instruction-following benchmarks. Today that lead has substantially narrowed and in some dimensions it has disappeared entirely.

The closed source frontier

The three labs with clear frontier-level closed source models are OpenAI, Anthropic, and Google.

OpenAI’s GPT-5.2 is currently the strongest all-around reasoning model. It excels at complex multi-step problems, long-horizon agentic tasks, and code generation that requires architectural thinking rather than syntax completion. The flagship tier comes at a significant price premium, which we will get to shortly.

Anthropic’s Claude 4 family brings a different profile. Claude 4 Sonnet is widely regarded as one of the best models for nuanced instruction-following, long-document analysis, and tasks where tone and precision matter. Think legal review, writing assistance, and structured content extraction. Claude 4 Opus is the top of the range with a price tag to match.

Google’s Gemini 2.5 Pro stands out for multimodal reasoning and long-context performance. The 976K context window is genuinely useful for codebases, long documents, and large data dumps. It sits comfortably in the frontier tier on most benchmarks and has made meaningful improvements in reasoning since the 1.x generation.

The open source challengers

The more interesting story is what has happened on the open source side.

DeepSeek V3 (and its variants V3.1 and V3.2) established that open weight models could compete on reasoning benchmarks that were once considered closed source territory. The V3 architecture combined with MoE efficiency delivers frontier-adjacent quality at a fraction of the cost.

Kimi K2 (moonshotai/Kimi-K2-Instruct-0905) is purpose-built for long-context agentic work. With a 256K token context window and strong coding performance, it handles repo-scale and multi-document tasks that used to require a closed source model. On ArtificialAnalysis benchmarks it ranks competitively against mid-tier closed source options.

GLM-4.6 from Zhipu AI is a reasoning-tuned model that has emerged as one of the better options for coding copilots, long-context RAG, and multi-tool agent loops. It has shown particularly strong first-token latency on DeepInfra, sub-second TTFT, which matters a lot for interactive use cases.

Qwen3 from Alibaba has become one of the most actively deployed open weight families. Qwen3-235B-A22B, a large MoE model, delivers near-frontier quality across reasoning and instruction-following. The smaller Qwen3-32B and Qwen3-30B-A3B variants offer strong performance at very competitive price points.

The honest summary: for tasks requiring peak intelligence on the hardest reasoning problems, closed source models, especially GPT-5.2 and Claude 4 Opus, still hold an edge. For the broad middle of production workloads including coding assistance, document analysis, structured output generation, RAG, and agent loops, open source models are now credible alternatives and in some cases the better choice once cost is factored in.

Pricing: The Real Divide

This is where the open vs. closed source choice becomes most concrete. Closed source models are billed at source, meaning you pay OpenAI, Anthropic, or Google directly at their posted rates. Open source models can be served by inference providers like DeepInfra at rates that reflect actual infrastructure cost rather than a platform premium.

Closed source pricing at source

The leading closed source models are priced at a significant premium:

Model	Input ($/1M tokens)	Output ($/1M tokens)
GPT-5.2 (OpenAI)	$1.75	$14.00
Claude 4 Sonnet (Anthropic)	~$3.00	~$15.00
Claude 4 Opus (Anthropic)	~$15.00	~$75.00
Gemini 2.5 Pro (Google)	$1.25	$10.00
Gemini 2.5 Flash (Google)	$0.30	$2.50

Pricing reflects public API rates as of April 2026.

The output token cost is where these models can surprise you. If your workload generates long completions, which is common in agent loops, document generation, and code synthesis, output tokens dominate the bill. At $14 to $75 per million output tokens the math gets steep quickly.

Open source on DeepInfra

Open source models hosted on DeepInfra run at a completely different price level. Here is the current pricing for key models across three tiers.

Premium tier — frontier-adjacent quality:

Model	Input ($/1M tokens)	Output ($/1M tokens)	Context
Kimi K2 0905	$0.40 / $0.15 cached	$2.00	128k
DeepSeek R1-0528	$0.50 / $0.35 cached	$2.15	160k

Mid tier — strong performance, budget-friendly:

Model	Input ($/1M tokens)	Output ($/1M tokens)	Context
DeepSeek V3.2	$0.26 / $0.13 cached	$0.38	160k
DeepSeek V3.1	$0.21 / $0.13 cached	$0.79	160k
Qwen3-235B-A22B	$0.071	$0.10	262k
Llama 4 Maverick	$0.15	$0.60	1048k
NVIDIA Nemotron 3 Super	$0.10 / $0.10 cached	$0.50	262k

Budget tier — fast, cheap, and surprisingly capable:

Model	Input ($/1M tokens)	Output ($/1M tokens)	Context
Llama 4 Scout	$0.08	$0.30	320k
Qwen3-32B	$0.08	$0.28	40k
Gemma 3 27B	$0.08	$0.16	131k
Mistral Small 24B	$0.05	$0.08	32k
Llama 3.1 8B	$0.02	$0.05	131k

Pricing as of April 2026.

What the numbers mean in practice

The cost difference compounds fast at scale. Take a typical production workload: a RAG pipeline sending 5,000 input tokens and receiving 1,000 output tokens per request, running 100,000 requests per month.

With GPT-5.2 at $1.75 input and $14.00 output, that comes out to around $2,275 per month.

With DeepSeek V3.2 on DeepInfra at $0.26 input and $0.38 output, the same workload costs around $168 per month. That is a 13x cost difference. For many production workloads, that delta is the difference between a profitable product and one that is not.

The Kimi K2 case is worth highlighting specifically. At $0.50 input and $2.00 output on DeepInfra, it offers performance that competes with mid-tier closed source models at a fraction of the price, especially for long-context and agentic work where its 256K window and agentic tuning shine. GLM-4.6 similarly has emerged as a strong value option for coding and reasoning workloads where you want reasoning-quality output without the reasoning-model price tag.

Speed: Where Infrastructure Shapes the Experience

Most people focus entirely on which model to use and stop thinking there. But two applications running the same model on different infrastructure can feel completely different to the end user, and in production that gap matters just as much as the model itself.

Speed has two components that matter differently depending on your use case. Time to First Token (TTFT) measures how fast the first character appears. Throughput measures how fast the full response streams in tokens per second. For interactive applications like chat UIs and IDE assistants, TTFT dominates perceived quality. For batch processing or long document generation, throughput matters more.

Closed source models route through the provider’s own infrastructure. OpenAI, Anthropic, and Google operate at massive scale, which generally means solid baseline performance. But you have no visibility into the infrastructure, no control over routing, and you share capacity with everyone else on the platform.

Open source models on DeepInfra run on dedicated H100 and A100 GPU infrastructure optimized specifically for inference. This translates to low and predictable TTFT, tight variance at p95 and p99, and throughput that does not degrade under load.

For Kimi K2, independent benchmarks from ArtificialAnalysis show DeepInfra posting a 0.33s TTFT, solidly in the instant-feel range, with tight variance that keeps IDE assistants and chat UIs responsive even under bursty traffic. For GLM-4.6, DeepInfra clocks sub-0.51s TTFT, one of the lowest in the provider cohort.

Speed also intersects with pricing in a way that is easy to miss. A faster inference stack means shorter wall-clock time per request, which matters for concurrent request capacity and user-perceived cost. Paying a few extra tenths of a cent per token for a provider that is materially slower can end up costing more in practice than the list price suggests.

Picking the Right Model for Your Workload

The open vs. closed source choice is not a single decision. It is a routing question you answer per use case.

When closed source makes sense:

For tasks requiring peak reasoning on genuinely hard problems like complex research, multi-step autonomous agents, and high-stakes synthesis, the frontier closed source models still hold an advantage. Gemini 2.5 Pro is also a strong choice for large-context multimodal tasks at Google scale. And if you are already deeply integrated with a provider’s ecosystem, the marginal migration cost may not be worth it.

When open source on DeepInfra wins:

High-volume production workloads where cost per token determines unit economics are the clearest win for open source. Long-context RAG, document analysis, and structured extraction are well-served by DeepSeek V3 and Kimi K2, which are competitive on quality and dramatically cheaper. Coding assistance and agent loops are well-handled by Kimi K2, GLM-4.6, and Qwen3. And for latency-sensitive applications, DeepInfra’s infrastructure gives you predictable, tight TTFT variance that is hard to get on shared closed source endpoints.

Cached input pricing also makes a real difference for workloads that reuse the same system prompt or large context blocks. DeepInfra supports cached input on Kimi K2, DeepSeek V3, Claude, and others, which meaningfully reduces cost on repeated or chunked context.

The approach most teams land on: use a closed source flagship for tasks that genuinely need it, route the majority of traffic to open source models on DeepInfra, and measure quality per workload rather than assuming the most expensive option is always necessary.

For the broad middle of production AI workloads in 2026, open source models are not a compromise. They are the economical choice that often performs just as well and with the right inference provider, just as fast.

Disclaimer:

All prices reflect public rates as of April 2026. Token rates change frequently. Verify current pricing at deepinfra.com/pricing and the respective provider pages before making production decisions.