AI Model Comparison 2026: Claude vs GPT vs Gemini vs Open Source

The Artificial Analysis Intelligence Index tracks 381 AI models across ten benchmarks as of June 2026. Claude Fable 5 holds the top position with a composite score of 65, but Kimi K2.6, an open-weights model from Moonshot AI, scores 54. That 11-point gap puts the best freely available model within 17% of the proprietary ceiling. Meanwhile, API pricing for a 1M-token context window spans a 71x range, from $0.14 on DeepSeek V4 Flash to $10.00 on Claude Fable 5.

This convergence changes how teams should think about model selection. The raw capability differences between the top five models are smaller than they were at the start of the year. The real decision comes down to which combination of speed, cost, context length, and tool-calling reliability fits your workload. A team running batch document classification probably doesn't need the same model as a team building an agentic coding assistant, even if both teams could afford the most expensive option.

Context windows have also expanded dramatically. Thirteen models now ship with 1M+ token windows, and Llama 4 Scout offers a 10M-token context for self-hosted deployments. These windows are large enough to process entire codebases, legal contracts, or financial reports in a single call, though effective context (measured by retrieval accuracy at depth) still falls short of advertised maximums on every model tested. That gap between advertised and effective context is something to test against your own data before committing to a context-length strategy.

Pricing tells an equally important story. The cheapest option (DeepSeek V4 Flash at $0.14 per million input tokens) can process roughly 71 times more text for the same dollar as Claude Fable 5. For many workloads, a model that scores 10% lower on benchmarks but costs 50x less is the better engineering decision.

This guide compares models across four proprietary providers and the open-source ecosystem, with every benchmark score and price cited from primary sources updated within the last 30 days.

The Artificial Analysis Intelligence Index aggregates ten benchmarks into a single composite score. These benchmarks span mathematics (AIME), scientific reasoning (GPQA Diamond), coding (SWE-bench Verified, SWE-bench Pro, Terminal-Bench Hard), agentic tasks (GDPval-AA, tau-squared-Bench), instruction following (IFBench), knowledge breadth (AA-Omniscience), and general intelligence (Humanity's Last Exam). The platform runs eight measurements per day for individual requests and two per day for parallel requests, with performance metrics calculated from the most recent 72 hours of live data.

No single benchmark tells the full story. A model that scores 95% on SWE-bench Verified might underperform on GPQA Diamond, and vice versa. The composite approach catches this by weighting breadth of capability over peak performance in any one area.

Beyond the Intelligence Index, we tracked three additional dimensions for this comparison:

• Speed: output tokens per second, measured by Artificial Analysis across standardized hardware
• Price: per-million-token costs for both input and output, pulled from official API pricing pages
• Context window: maximum advertised context length, with notes on effective retrieval depth where available

These four dimensions (intelligence, speed, price, context) are the ones that matter when you're choosing between models for production work. Vibes and Twitter discourse are not reliable selection criteria.

Four companies are shipping frontier-class models with distinct strengths. The differences between them are less about raw intelligence scores and more about where each model excels in practice: coding accuracy, reasoning depth, tool-calling reliability, ecosystem integration, and cost at scale.

A practical way to evaluate these is to run the same prompt through each model's API and compare outputs side by side. Services like Artificial Analysis automate this with standardized benchmarks, but nothing replaces testing against your own production data. A model that scores 5% lower on a public benchmark might handle your specific document format or coding style better than the leader.

One constraint worth noting: pricing tiers shift frequently. OpenAI, Anthropic, and Google all adjusted prices at least once in Q1 2026. Pin your cost estimates to a specific date and recheck before signing annual commitments.

Open-source and open-weights models have reached a point where they compete directly with proprietary options on most benchmarks. The trade-off is infrastructure: you either self-host (with the associated GPU costs and operational overhead) or use a hosted API provider like Together, Fireworks, or Replicate.

Self-hosting gives you full control over data residency, latency, and per-token cost at scale, but it introduces operational complexity. A single A100 GPU can serve Qwen 3.5 or DeepSeek V4 Flash at reasonable throughput, but running Llama 4 Maverick's full parameter count requires multi-GPU setups. Hosted API providers remove that burden at a markup. For teams that want open-weights pricing without managing GPUs, providers like Together AI and Fireworks AI offer pay-per-token access to most of the models below, typically at 2x to 5x the self-hosted cost but with zero infrastructure overhead.

The licensing landscape also matters. Apache 2.0 (Qwen, Mistral) allows unrestricted commercial use. Meta's Llama license adds a 700M monthly active user threshold. DeepSeek uses an MIT license. Check the specific terms before building a product on any open model, especially if you plan to fine-tune and redistribute.

Test AI models against your own documents

Fast.io gives any AI model MCP access to your files, search, and workflows. 50GB free storage, 5,000 credits per month, no credit card required.

Sign Up Free

The right model depends on what you're building. Here are concrete recommendations by use case.

Coding and software engineering: Claude Fable 5 (95.0% SWE-bench Verified) is the top performer, but Claude Opus 4.8 at $5/$25 per million tokens hits 88.6% for less than half the price. DeepSeek V4 Pro (83.7%) is the open-source alternative at $0.44/$3.48. For budget-constrained coding pipelines, DeepSeek V4 Flash handles routine tasks at $0.14/$0.28.

Scientific reasoning and research: Gemini 3.1 Pro leads with 94.3% on GPQA Diamond at $2/$12 per million tokens. Qwen 3.5 (88.4% GPQA Diamond) is the open-source option for teams that need to self-host.

Long-context document processing: Llama 4 Scout's 10M-token window is unmatched for single-pass processing. For proprietary options, Gemini 3.1 Pro and Grok 4.1 Fast both offer 1M+ windows at competitive prices ($2/$12 and $0.20/$0.50 respectively).

Cost-sensitive batch processing: DeepSeek V4 Flash ($0.14/$0.28) and Grok 4.1 Fast ($0.20/$0.50) deliver usable intelligence at prices that make million-document processing feasible. Gemini 3.1 Flash-Lite ($0.10/$0.40) is even cheaper for simpler classification or extraction tasks.

Agentic workflows with file access: Agent performance depends on the model and the infrastructure it connects to. Most teams building agents need persistent file storage, tool-calling reliability, and a way to hand results off to humans. The model handles reasoning; the workspace handles state. Fast.io provides this layer through its MCP server, giving agents built on any model access to shared workspaces, file versioning, semantic search, and ownership transfer. The free agent plan includes 50GB of storage, 5,000 credits per month, and five workspaces with no credit card required.

For tool-calling specifically, Claude Opus 4.7 scores 79.1% on MCP-Atlas versus GPT-5.5's 75.3%, making the Claude family the strongest option for agents that rely heavily on external tool integrations.

General-purpose applications: GPT-5.5 at $5/$30 with an Arena Elo of 1,561 is the safest default. It scores well across every category and has the broadest third-party integration ecosystem. Claude Opus 4.8 at $5/$25 is the main competitor, with stronger coding performance and slightly lower output costs.