Best AI for Summarization
in 2026

Top 25 AI models benchmarked on summarization with key facts retention — passage summarization, argument extraction, and bullet-point condensation. One model retained 90% of required facts.

What This Benchmark Tests (and Does Not Test)

Task Summarization with Key Facts Retention — can the model produce concise summaries that preserve required information? Categories Passage summarization (1 test), main argument extraction (1 test), bullet-point condensation (1 test) Scoring Deterministic — contains_all with partial credit, case-insensitive (no LLM-as-judge) Points 3 tests × 1 point each = 3.0 max score, with partial credit for partially retained facts Models tested 25 models from 11 providers (24 completed all tests) Stability 2 runs per model Config Default API configurations, recommended temperature where available Date March 2026 Does not test Long-document summarization, abstractive quality, coherence, readability, or multi-document synthesis

All models tested with default API configurations. This benchmark measures fact retention in short summaries, not subjective summary quality.

Benchmark Results

Bar chart showing AI model accuracy scores for summarization benchmark, March 2026. Minimax M2.5 Lightning leads at 90%, followed by Grok-4 at 87%.
Accuracy by model — Summarization with Key Facts Retention March 2026
Table showing full benchmark results for 24 AI models on summarization tasks, including accuracy, cost per run, speed, and stability metrics
Full results — scores, cost, speed, stability, and token usage March 2026

Key Findings

🏆 Accuracy: Minimax Dominates at 90%

Minimax M2.5 Lightning leads at 90% (2.7/3.0) with perfect ±0.000 stability — retaining nearly all required key facts across every run. Grok-4 follows at 87%, then Grok-4-1 Fast Reasoning and GPT-5-mini tied at 83%. The top scorer comes from the least expected provider, while XAI and OpenAI fill the podium.

💰 Cost Efficiency: DeepSeek Chat — 70% at a Fraction of a Cent

DeepSeek Chat scored 70% at $0.000108/run — the highest accuracy-per-dollar ratio in the test (19,430 Acc/$). Llama4 Maverick scored 77% at $0.000161/run (14,279 Acc/$). For top accuracy on a budget, Grok-4-1 Fast Reasoning scored 83% at just $0.000474/run — 127x cheaper than GPT-5.4-Pro (73%).

⚡ Speed: Claude Haiku Is the Fast Fact-Retainer

Claude Haiku 4.5 responded in 4.87s with 73% accuracy — the best speed-to-accuracy ratio in the test. Mistral Medium was fastest overall at 3.62s but scored only 53%. Llama4 Maverick hit 77% in 6.32s. The winner Minimax took 13.61s, and GPT-5.4-Pro was slowest at 66.50s for a lower 73% score.

📉 Flagships Beaten by Their Own Cheaper Siblings

GPT-5-mini (83%) outperformed GPT-5.4-Pro (73%) at 1/60th the cost. Claude Haiku 4.5 (73%) beat both Claude Sonnet 4.6 (60%) and Claude Opus 4.6 (60%) — the cheapest Claude model outperformed the most expensive by 13 points. In summarization, bigger and pricier does not mean better at retaining facts.

Why These Results May Surprise You

Minimax scoring 90% on summarization while Claude Opus manages only 60% seems wrong — until you understand what's being measured:

🎯 Fact retention ≠ summary quality: This benchmark checks whether specific key facts appear in the output. A model that writes an elegant, readable summary but paraphrases a required keyword ("city officials" instead of "city council") loses points. Precision matters more than prose.
💡 Over-abstraction is penalized: Premium models tend to abstract and generalize. When a passage says "$2 million" and "eight months," models that summarize as "a multi-million dollar project over several months" lose credit for missing the exact figures.
📐 Compact models excel at following instructions: Smaller, faster models like GPT-5-mini and Llama4 Maverick tend to stick closer to the source material, reproducing key terms rather than rephrasing them. This is exactly what contains_all scoring rewards.
🔄 Partial credit separates the field: With 7-9 required key facts per test, even small differences in retention compound. A model retaining 7/9 facts scores much higher than one retaining 5/9, even if both summaries read well.

A model that excels at retaining exact facts in short summaries might not be the best at producing executive summaries of 50-page documents. Task design determines rankings.

⚖️ Generic Benchmarks vs. Custom Benchmarks

These results are a directional signal, not a universal summarization ranking. This benchmark tests short-form fact retention across three specific task types — passage summarization, argument extraction, and bullet-point condensation. Your summarization needs might involve legal documents, medical records, financial reports, or meeting transcripts.

The fact that Minimax — a model most teams wouldn't consider for production summarization — scored 90% while Claude Opus scored 60% shows how sensitive rankings are to task design and scoring method. A contains_all scorer rewards literal fact retention; a human evaluator might rank these models very differently.

These results are valid for this task design and scoring setup. Change the task, constraints, or scoring, and rankings can change — which is exactly why custom benchmarking matters.

Frequently Asked Questions

Which AI model is best for summarization in 2026?

On our key facts retention benchmark, Minimax M2.5 Lightning scored highest at 90% with perfect stability, followed by Grok-4 (87%) and GPT-5-mini (83%). But the best model depends on your document types and what you need from summaries. Run your own benchmark to find out.

Is Claude or GPT better for summarization?

On this benchmark, GPT significantly outperformed Claude. GPT-5-mini scored 83% while Claude Haiku 4.5 scored 73%, and both Claude Sonnet 4.6 and Opus 4.6 scored just 60%. Notably, the cheapest Claude model (Haiku) outperformed both premium Claude models on fact retention.

What is the cheapest AI for summarization?

DeepSeek Chat scored 70% at just $0.000108/run — the cheapest option tested. Llama4 Maverick scored 77% at $0.000161/run. For the best balance of accuracy and cost, Grok-4-1 Fast Reasoning scored 83% at $0.000474/run (Low tier).

How do you benchmark AI summarization?

OpenMark uses deterministic scoring — no LLM-as-judge. Summarization tasks are scored using contains_all with partial credit, checking whether each summary retains specific key facts from the source text. This measures fact retention, not subjective quality. Results are 100% reproducible. Try it yourself for free.

Can AI accurately summarize documents?

Yes, but accuracy varies widely. Scores on our benchmark ranged from 47% to 90%. Even premium models missed key facts — Claude Opus retained just 60%. For production use, always test on your actual document types to understand fact loss rates.

Benchmark AI Models on Your Summarization Tasks

Test which model retains the facts YOU need in summaries.
100 free credits — no credit card required.

Run a Summarization Benchmark — Free →

More from OpenMark

Best AI for Math Best AI for Writing Best AI for Translation Best AI for Classification Best AI for Coding Best AI for Agents GPT vs Claude 2026 AI Pricing Comparison Compare AI Models LLM Leaderboard Why Benchmark AI Models? Launch OpenMark App