Our research team spent the first half of 2026 evaluating 42 large language models using data compiled from the Artificial Analysis Intelligence Index, LM Council’s independently run benchmark leaderboard, vals.ai’s standardized SWE-bench harness, GPQA Diamond, Humanity’s Last Exam, and official developer documentation published by each model’s creator. We scored each model on eight weighted criteria:
- AA Intelligence Index (25%): Artificial Analysis composite score, aggregating performance across coding, reasoning, math, and knowledge tasks into a single 0–100 scale.
- SWE-bench Verified (20%): Standardized resolution rate on 500 real-world GitHub issues, evaluated on the vals.ai Mini-SWE-agent bash-only harness for cross-model comparability.
- GPQA Diamond (15%): Accuracy on graduate-level science questions spanning physics, chemistry, and biology, scored by Artificial Analysis independently.
- Context Window (10%): Maximum tokens the model can process in a single inference call. Relevant for long documents, large codebases, and extended agent sessions.
- Output Speed (10%): Median tokens per second across providers, measured by Artificial Analysis under standardized conditions.
- Blended API Cost (10%): Weighted cost per 1M tokens at a standard input/output ratio, sourced from Artificial Analysis; open-weight models scored on the lowest confirmed third-party hosting cost.
- Modalities (5%): Breadth of supported input and output types, including text, vision, audio, video, and image generation.
- Open-Weight Availability (5%): Whether model weights are publicly released, enabling self-hosting, fine-tuning, and deployment without vendor dependency.
Where benchmark data was not publicly available, a conservative below-average penalty score was applied to that factor. All blended prices and Intelligence Index scores are sourced from Artificial Analysis (June 2026) unless otherwise noted. SWE-bench Verified scores reflect vals.ai’s standardized Mini-SWE-agent harness, which provides the same evaluation environment for all models and may differ from developer-reported scores that use proprietary harnesses. Scores above approximately 80% on SWE-bench Verified should be interpreted with caution, given active community debate about benchmark saturation and varied utility.
The Top AI Models of 2026
| # | Model | Developer | AA Intel. | SWE-bench Verified | GPQA Diamond | Context | Speed (tok/s) | Blended $/1M | Modalities | Open-Weight |
| 1 | Claude Fable 5 | Anthropic | 60 | 95.0%ᵃ | 94.1%ᵇ | 1M | N/Aᶜ | $7.70 | Text, Vision | No |
| 2 | Claude Opus 4.8 | Anthropic | 56 | 88.6%ᵃ | 93.6%ᵈ | 1M | 61 | $3.85 | Text, Vision | No |
| 3 | GPT-5.5 | OpenAI | 55 | 82.6%ᵃ | 93.5%ᵉ | 922K | 67 | $4.35 | Text, Vision, Audio, Images | No |
| 4 | GLM-5.2 | Z AI | 51 | 82.8%ᵃ | 89.5%ᵉ | 1M | 106 | $0.90 | Text, Vision | Yes (MIT) |
| 5 | Gemini 3.5 Flash | 50 | 78.8%ᵃ | 92.2%ᵉ | 1M | 167 | $1.31 | Text, Vision, Audio | No | |
| 6 | Gemini 3.1 Pro | 46 | 78.8%ᵃ | 94.1%ᵉ | 1M | 129 | $1.74 | Text, Vision, Audio, Video | No | |
| 7 | Qwen 3.7 Max | Alibaba | 46 | 80.4% | 92.3%ᵉ | 1M | 198 | $1.43 | Text, Vision | No |
| 8 | Claude Sonnet 4.6 | Anthropic | 47 | 79.6%ʰ | 89.9%ⁱ | 1M | 48 | $2.31 | Text, Vision | No |
| 9 | DeepSeek V4 Pro | DeepSeek | 44 | 80.6% | 90.5%ᵉ | 1M | 86 | ~$2.18ʲ | Text, Code | Yes (MIT) |
| 10 | MiniMax-M3 | MiniMax | 44 | 80.5%ᵐ | 92.9%ᵉ | 1M | 85 | $0.22 | Text, Vision | Yes (MiniMax Community License) |
| 11 | Kimi K2.6 | Moonshot AI | 43 | 80.2% | 91.1%ᵉ | 256K | 74 | $0.70 | Text, Vision | Yes (Mod. MIT) |
| 12 | Grok 4 | xAI | ~47ᶠ | 69.1%–72%ᵒ | 87.7%ᵉ | 1M | N/A | Sub.ᵍ | Text, Vision, Audio, Video | No |
| 13 | Llama 4 Maverick | Meta | 49 | N/Aᵏ | N/Aᵏ | 1M | Varies | Open-weightˡ | Text, Vision, Audio | Yes (Meta Lic.) |
| 14 | GPT-5.3 Codex | OpenAI | 44* | N/A | 91.5%ᵉ | 400K | 91 | $1.87 | Text, Code | No |
| 15 | DeepSeek V4 Flash | DeepSeek | N/A | 79.0% | 89.4%ᵉ | 1M | 108.9 | $0.15 est.ⁿ | Text, Code | Yes (MIT) |
ᵃ Vals.ai standardized Mini-SWE-agent harness (vals.ai/benchmarks/swebench), June 17, 2026. Scores may differ from developer-reported figures that use proprietary harnesses.
ᵇ Anthropic Fable 5 and Mythos 5 system card — https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf (Mythos 5 shares the same underlying model as Fable 5).
ᶜ AA provider speed data for Fable 5 not publicly confirmed at time of publication.
ᵈ Anthropic Claude Opus 4.8 official system card — https://www-cdn.anthropic.com/0b4915911bb0d19eca5b5ee635c80fef830a37ea.pdf
ᵉ Artificial Analysis GPQA Diamond benchmark, mirrored on benchlm.ai/benchmarks/aaGpqaDiamond, June 18, 2026.
ᶠ Estimated from GPQA Diamond (87.7%) and HLE (~45%) performance. Official AA Intelligence Index not publicly confirmed for Grok 4.
ᵍ Subscription-based: SuperGrok $30/mo, SuperGrok Heavy $300/mo. Per-token API pricing: $3.00/$15.00 per 1M input/output (Grok 4); $0.20/$0.50 per 1M input/output (Grok 4 Fast, up to 120K context).
ʰ Anthropic official announcement and Sonnet 4.6 system card, confirmed by Mashable.
ⁱ Anthropic Claude Sonnet 4.6 official system card — https://anthropic.com/claude-sonnet-4-6-system-card
ʲ Confirmed from FriendliAI (friendli.ai/blog/deepseek-v4-pro-flash) and Lushbinary (April 2026): V4 Pro API is $1.74/$3.48 input/output per 1M tokens; ~$2.18 blended at 75/25. Note: the Artificial Analysis leaderboard labels an entry “DeepSeek V4 Pro (Max)” at $0.18 blended, which, based on pricing, appears to reflect V4 Flash ($0.14/$0.28) rather than V4 Pro; confirmed V4 Pro pricing is used here.
ᵏ Meta did not publish SWE-bench Verified or GPQA Diamond scores for Llama 4 Maverick in official launch materials (ai.meta.com/blog/llama-4-multimodal-intelligence/, April 2025). AA Intelligence Index of 49 is confirmed from the Artificial Analysis leaderboard.
ˡ Open-weight via Meta Llama 4 Community License. API pricing is approximately $0.20 to $0.45/1M via cloud providers.
ᵐ morphllm.com, June 2026. Flagged for potential training data contamination on SWE-bench Verified; treat as indicative.
ⁿ Estimated from $0.14 input / $0.28 output (Lushbinary, April 2026) at AA’s standard 95/5 input/output ratio.
ᵒ Grok 4 Heavy class range per tech-insider.org (May 24, 2026), cross-referenced against xAI developer documentation. Standard Grok 4 may score lower; Grok 4 Fast variant not separately benchmarked on SWE-bench at publication. *Preliminary per Artificial Analysis notation.
All benchmark data sourced from Artificial Analysis, vals.ai, benchlm.ai, and official developer documentation as of June 2026. N/A = not publicly confirmed at time of publication.
Here is what each model on this list is genuinely best for, and the one thing that sets it apart from every other option. Each model is expanded upon, including their biggest pros and cons and best use cases.
| Model | Best For | What Sets It Apart |
| Claude Fable 5 | Research teams and enterprises that need the most technically capable AI available | The highest benchmark scores of any model tested, but comes at a price befitting a frontier model, and is the most expensive on the list |
| Claude Opus 4.8 | Engineering teams building AI that works autonomously over hours, not seconds | The best model outside of Fable for complex software projects, autonomous debugging, and multi-step tasks that can’t fail partway through at a much lower price than Fable |
| GPT-5.5 | Teams that want one model to handle everything: writing, research, images, voice, and code | The only model here that natively combines text, vision, audio, and image generation; the most versatile all-in-one option |
| GLM-5.2 | Developers who need near-frontier coding performance but can’t justify $4+ per million tokens, or need to self-host | Outscores GPT-5.5 on the standardized coding benchmark at one-fifth the price; fully open-weight with no regional restrictions |
| Gemini 3.5 Flash | Product and engineering teams running high-volume APIs: chatbots, document processing, real-time user-facing features | The best confirmed performance-to-cost ratio among frontier-class closed models in this dataset; outputs 167 tokens per second at $1.31 per million |
| Gemini 3.1 Pro | Scientists, medical researchers, academics, and anyone doing serious analytical work across text, images, audio, and video | Tied for the highest science reasoning score in this dataset and leads HLE among currently available models; the only model on this list with native Google Workspace integration across Docs, Sheets, Meet, and Drive |
| Qwen 3.7 Max | Live user-facing products where slow responses lose customers: real-time coding assistants, streaming interfaces, interactive chatbots | The fastest model in this dataset at 198 tokens per second, making it the only frontier option where users genuinely won’t notice a lag |
| Claude Sonnet 4.6 | Content teams, technical writers, and product teams that want Anthropic quality without Anthropic’s top-tier prices | 40% cheaper than Claude Opus 4.8 with strong long-form and document performance; the highest-rated model for long-form content and technical documentation in T-Minus AI’s May 2026 comparison |
| DeepSeek V4 Pro | Quantitative analysts, competitive programmers, and engineering teams working on math-heavy problems | Leads every math and competitive coding benchmark in this dataset, including Codeforces (3,206 rating) and LiveCodeBench (93.5); open-weight at approximately $2.18 per million tokens |
| MiniMax-M3 | Teams processing large volumes of simple, repetitive tasks: QA, bug triage, and basic summarization, where cost is the primary concern | $0.22 per million tokens with a GPQA Diamond score competitive with models costing 6 to 17 times as much; open-weight and free to self-host |
| Kimi K2.6 | Teams building multi-agent AI systems where dozens or hundreds of AI instances need to coordinate simultaneously | The highest confirmed HLE (with tools) score in this dataset at 54.0%; supports 300 parallel sub-agents across 4,000 coordinated steps; open-weight at $0.70 per million tokens, though limited to a 256K context window |
| Grok 4 | Social analysts, PR teams, journalists, and anyone whose work requires understanding what is happening online right now | The only model here with DeepSearch, which synthesizes X (Twitter) and live web data in real time within a single response; priced by subscription rather than per token |
| Llama 4 Maverick | Enterprise IT and infrastructure teams that need full control over a capable, multimodal AI model running on their own servers | The most widely deployed open-weight model in the enterprise; compatible with every major inference framework, including vLLM, Ollama, and Hugging Face Transformers, with no API vendor dependency |
| GPT-5.3 Codex | DevOps and platform engineers automating terminal commands, CI/CD pipelines, and command-line workflows | A specialist model built for automated code execution rather than general chat; leads Terminal-Bench 2.0 at 77.3% and is the most cost-efficient option for isolated pipeline automation inside the OpenAI stack |
| DeepSeek V4 Flash | Startups and AI-native product teams building at consumer scale, where cost is the hard constraint | The cheapest confirmed price in this dataset is roughly $0.15 per million tokens; processing 10 million output tokens costs approximately $2.80; open-weight under MIT license |
In-Depth Model Reviews
1. Claude Fable 5, for Peak Reasoning and Multi-Step Inference
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Claude Fable 5 | The most capable AI model available by every major independent test | Also the most expensive; running it at scale costs significantly more than any other model on this list | Complex research, multi-step analysis, and high-stakes projects where getting the best possible answer justifies a premium price |
Claude Fable 5 is Anthropic’s current flagship and the highest-scoring model on the Artificial Analysis Intelligence Index as of June 2026, with a composite score of 60. It leads SWE-bench Verified at 95.0% on vals.ai’s standardized harness, a 6.4-point gap over the next-best model, and posts 94.1% on GPQA Diamond per the Anthropic system card, tied with Gemini 3.1 Pro for the highest confirmed score in this dataset. On SWE-bench Pro, Fable 5 leads all models at 80.3%, more than 11 points ahead of Opus 4.8 (69.2%). It also tops SimpleBench at 81.9%, a benchmark designed to resist pattern memorization, and HLE (no tools) at 53.3% (shared with Claude Mythos).
At $7.70 blended per 1M tokens, it is the most expensive model in this dataset. It is cost-justified for multi-step research synthesis, long-horizon planning, and any workflow where the intelligence gap over cheaper models produces measurable downstream value.
Developer: Anthropic
AA Intelligence Index: 60
SWE-bench Verified: 95.0%
GPQA Diamond: 94.1%
Context Window: 1,000,000 tokens
Output Speed: N/A
Blended API $/1M: $7.70
Modalities: Text, Vision
Open-Weight: No
2. Claude Opus 4.8, for Agentic Coding and Long-Horizon Tasks
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Claude Opus 4.8 | The best AI available for writing, debugging, and managing code autonomously over long periods without human check-ins | One of the pricier options; not the right fit for simple or high-volume repetitive tasks | Software engineering teams that need an AI capable of working through complex coding projects on its own, from start to finish |
Claude Opus 4.8 holds the second-highest SWE-bench Verified score in this dataset at 88.6% and ranks second on the AA Intelligence Index at 56. Released in May 2026, it leads FrontierSWE, tops PostTrainBench (autonomous model improvement via post-training), and scores 85.0 on Terminal-Bench 2.1, the highest CLI execution score of any model except Fable 5.
Its GPQA Diamond score of 93.6% per the official Anthropic system card places it third in this dataset on confirmed reasoning data. At $3.85 blended per 1M tokens, Opus 4.8 is the recommended model for production software engineering agents, particularly those handling multi-file GitHub issues, security-sensitive repositories, or extended autonomous debugging sessions. It serves similar use cases to Fable 5, but at a slightly lower tier of speed and accuracy, which is generally considered a worthwhile tradeoff for many adopters due to its significantly lower price.
Developer: Anthropic
AA Intelligence Index: 56
SWE-bench Verified: 88.6%
GPQA Diamond: 93.6%
Context Window: 1,000,000 tokens
Output Speed: 61 tok/s
Blended API $/1M: $3.85
Modalities: Text, Vision
Open-Weight: No
3. GPT-5.5, for Omnimodal Agentic Workflows
| Model | Biggest Pro | Biggest Con | Best Use Case |
| GPT-5.5 | The only model here that can read images, listen to audio, generate images, and write code — all in one place | More expensive than most alternatives, and not the top performer on any single task | Teams that want one AI to handle everything rather than managing multiple specialized tools |
GPT-5.5, released by OpenAI in April 2026, scores 55 on the AA Intelligence Index, posts 82.6% on SWE-bench Verified on the standardized harness, and achieves 93.5% on GPQA Diamond (AA-GPQA, June 2026). It leads Terminal-Bench 2.0 at 82.7% and tops BrowseComp (multi-source web research quality) at 84.4%. GDPval places it at 49.7% across 44 professional occupations, the highest confirmed score in this dataset on that benchmark.
With its 922K context window, natively multimodal architecture covering text, vision, audio, and image generation, and the broadest confirmed tool-use coverage of any model here, GPT-5.5 is the default recommendation for teams that need one model to handle writing, research, image analysis, and autonomous agent workflows.
Developer: OpenAI
AA Intelligence Index: 55
SWE-bench Verified: 82.6%
GPQA Diamond: 93.5%
Context Window: 922,000 tokens
Output Speed: 67 tok/s
Blended API $/1M: $4.35
Modalities: Text, Vision, Audio, Images
Open-Weight: No
4. GLM-5.2, for Frontier-Level Coding at Low Cost
| Model | Biggest Pro | Biggest Con | Best Use Case |
| GLM-5.2 | Matches or beats much pricier models on coding tasks at a fraction of the cost; can also be downloaded and run on your own servers at no ongoing fee | Can only read text and images — no audio, video, or image generation | Engineering teams that need strong coding performance without paying premium API prices, especially those that want to keep their AI infrastructure in-house |
GLM-5.2 is Z AI’s open-weight flagship, released June 13, 2026, under an MIT license with no regional restrictions. On the vals.ai standardized SWE-bench harness, it scores 82.8%, placing it third in this dataset above GPT-5.5 (82.6%), at a fraction of GPT-5.5’s cost ($0.90 vs. $4.35 blended). Its AA Intelligence Index of 51 is the highest among open-weight models in this dataset. GPQA Diamond comes in at 89.5% (AA-GPQA).
On FrontierSWE (open-ended technical projects measured in hours), it trails Claude Opus 4.8 by only 1% and edges out GPT-5.5 by 1%. At $0.90 blended per 1M tokens, GLM-5.2 delivers near-Opus long-horizon coding performance for teams that need open weights, low API cost, or on-premise deployment.
Developer: Z AI (Zhipu AI)
AA Intelligence Index: 51
SWE-bench Verified: 82.8%
GPQA Diamond: 89.5%
Context Window: 1,000,000 tokens
Output Speed: 106 tok/s
Blended API $/1M: $0.90
Modalities: Text, Vision
Open-Weight: Yes (MIT)
5. Gemini 3.5 Flash, for Speed-Optimized Frontier Performance
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Gemini 3.5 Flash | Delivers strong results faster and cheaper than almost any competing model at its quality level | Not the strongest option for deep scientific or research questions requiring expert-level reasoning | High-volume products that need fast, reliable AI responses at low cost: customer chatbots, document tools, and real-time features |
Gemini 3.5 Flash outputs at 167 tokens per second, which is 2.7x the speed of Claude Opus 4.8 and the second-fastest confirmed speed in this dataset. It holds an AA Intelligence Index of 50, scores 78.8% on SWE-bench Verified on the standardized harness, and posts 92.2% on GPQA Diamond (AA-GPQA), placing it seventh in confirmed reasoning data.
On SimpleBench (adversarial common-sense reasoning), it ranks fourth at 76.7%. At $1.31 blended per 1M tokens, it offers the best confirmed performance-to-cost ratio among frontier-class closed models in this dataset. Teams running high-volume inference pipelines, including document summarization, API-first products, and real-time user-facing applications, will find Gemini 3.5 Flash the strongest balance of speed, intelligence, and price.
Developer: Google
AA Intelligence Index: 50
SWE-bench Verified: 78.8%
GPQA Diamond: 92.2%
Context Window: 1,000,000 tokens
Output Speed: 167 tok/s
Blended API $/1M: $1.31
Modalities: Text, Vision, Audio
Open-Weight: No
6. Gemini 3.1 Pro, for Scientific Reasoning and Multimodal Research
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Gemini 3.1 Pro | One of the sharpest reasoning models available for science and research questions; the only model here that plugs directly into Google Docs, Sheets, Meet, and Drive | Not as strong on software development tasks as the top coding-focused models | Researchers, academics, and enterprise teams already working inside Google’s ecosystem who need serious analytical depth |
Gemini 3.1 Pro is tied for the highest confirmed GPQA Diamond score in this dataset at 94.1% (AA-GPQA, June 2026). It is second among all commercially available LLMs on this list on the LM Council’s Humanity’s Last Exam leaderboard (no tools) at 46.4%. SWE-bench Verified on the standardized harness comes in at 78.8%, matching Gemini 3.5 Flash on that metric.
On METR Time Horizons, it ranks third at 384.1 minutes for sustained autonomous task execution. An output speed of 129 tok/s and a $1.74 blended price make it a compelling research-grade option at a fraction of Anthropic’s top-tier cost. Native Google Workspace integration across Docs, Sheets, Meet, and Drive gives it a deployment advantage for enterprise teams already in the Google ecosystem.
Developer: Google
AA Intelligence Index: 46
SWE-bench Verified: 78.8%
GPQA Diamond: 94.1%
Context Window: 1,000,000 tokens
Output Speed: 129 tok/s
Blended API $/1M: $1.74
Modalities: Text, Vision, Audio, Video
Open-Weight: No
7. Qwen 3.7 Max, for High-Throughput Frontier Output
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Qwen 3.7 Max | Produces responses faster than any other model on this list — critical for products where users are actively waiting on screen | No audio support, and cannot be self-hosted | Live, user-facing products where slow responses hurt the experience: real-time assistants, interactive chatbots, and streaming tools |
Qwen 3.7 Max outputs at 198 tokens per second, the fastest confirmed output speed in this dataset, while posting 80.4% on SWE-bench Verified and 92.3% on GPQA Diamond (AA-GPQA), both among the strongest figures in the mid-tier. On SWE-bench Pro, it scores 60.6%, the highest proprietary score among all models in this dataset except Claude Fable 5 (80.3%) and Opus 4.8 (69.2%). In the Text Arena Coding leaderboard, Qwen 3.7 Max ranks fourth at 1,540.8 among all confirmed models.
At $1.43 blended per 1M tokens, it is well-priced for high-throughput production workloads, including live coding assistants, interactive chatbots, and real-time summarization pipelines, where 198 tok/s output sets a meaningfully higher ceiling than any other frontier-class model in this dataset.
Developer: Alibaba
AA Intelligence Index: 46
SWE-bench Verified: 80.4%
GPQA Diamond: 92.3%
Context Window: 1,000,000 tokens
Output Speed: 198 tok/s
Blended API $/1M: $1.43
Modalities: Text, Vision
Open-Weight: No
8. Claude Sonnet 4.6, for Balanced Coding and Document Work
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Claude Sonnet 4.6 | Delivers Anthropic’s quality at a noticeably lower price — 40% cheaper than Claude Opus 4.8 | Slower response times than most models here, which can frustrate users in live or real-time applications | Content teams, technical writers, and businesses that want Anthropic-quality output for documents and long-form work without the top-tier price tag |
Claude Sonnet 4.6 carries an AA Intelligence Index of 47 and official Anthropic-confirmed benchmark scores of 79.6% on SWE-bench Verified and 89.9% on GPQA Diamond (per the official Sonnet 4.6 system card, confirmed by Mashable). T-Minus AI’s May 2026 comparison rates it as the top model for long-form content and technical documentation, citing consistent coding quality and strong long-horizon debugging.
At $2.31 blended per 1M tokens, it is approximately 40% cheaper than Claude Opus 4.8 ($3.85) and about half the blended cost of GPT-5.5 ($4.35). The output speed of 48 tok/s is the slowest among frontier models in this dataset, which is a practical constraint for real-time applications but is acceptable for batch or document-centric workloads.
Developer: Anthropic
AA Intelligence Index: 47
SWE-bench Verified: 79.6%
GPQA Diamond: 89.9%
Context Window: 1,000,000 tokens
Output Speed: 48 tok/s
Blended API $/1M: $2.31
Modalities: Text, Vision
Open-Weight: No
9. DeepSeek V4 Pro, for Cost-Efficient Math and Competitive Coding
| Model | Biggest Pro | Biggest Con | Best Use Case |
| DeepSeek V4 Pro | The best AI model on this list for math, competitive coding, and quantitative problem-solving | Struggles with factual accuracy — not reliable for customer-facing content, knowledge lookup, or fact-checking tasks | Financial analysts, competitive programmers, and engineering teams whose core work is built around math and numbers |
DeepSeek V4 Pro scores 80.6% on SWE-bench Verified and 90.5% on GPQA Diamond (AA-GPQA), while leading this dataset on math benchmarks: LiveCodeBench 93.5, IMOAnswerBench 89.8, and Codeforces competitive programming rating 3,206. Its hybrid CSA+HCA attention architecture reduces the KV cache footprint to 10% of V3.2’s size at 1M context, making long-context inference dramatically cheaper per token.
Confirmed API pricing is $1.74 input and $3.48 output per 1M tokens (approximately $2.18 blended at a 75/25 ratio), per FriendliAI and Lushbinary; see footnote ʲ regarding a discrepancy with the AA leaderboard label. The primary limitation is factual recall, where SimpleQA-Verified places it at 57.9 versus Gemini 3.1 Pro’s 75.6, an 18-point gap that matters for knowledge-base and customer support applications.
Developer: DeepSeek
AA Intelligence Index: 44
SWE-bench Verified: 80.6%
GPQA Diamond: 90.5%
Context Window: 1,000,000 tokens
Output Speed: 86 tok/s
Blended API $/1M: ~$2.18
Modalities: Text, Code
Open-Weight: Yes (MIT)
10. MiniMax-M3, for Ultra-Low-Cost High-Volume Inference
| Model | Biggest Pro | Biggest Con | Best Use Case |
| MiniMax-M3 | Delivers reasoning quality close to models costing many times more, at just $0.22 per million words processed | Coding performance figures have not been independently verified and may be inflated — treat its software development scores with caution | High-volume, cost-sensitive tasks where quality still matters: content moderation, bug triage, QA workflows, and basic summarization at scale |
MiniMax-M3 is the second-cheapest model in this dataset at $0.22 blended per 1M tokens. It posts 80.5% on SWE-bench Verified and 92.9% on GPQA Diamond (AA-GPQA, June 2026), which is competitive with models priced 6 to 17x higher. Its GPQA Diamond score is the fifth-highest confirmed figure in the entire dataset, behind Claude Fable 5 (94.1%), Gemini 3.1 Pro (94.1%), Claude Opus 4.8 (93.6%), and GPT-5.5 (93.5%). Terminal-Bench 2.1 comes in at 66.0% and BrowseComp at 83.5, just behind GPT-5.5’s 84.4.
The model is open-weight (confirmed by benchlm.ai and codingfleet.com), offering self-hosting options alongside the $0.22 blended API. Its AA Intelligence Index of 44 is the only metric that limits its claim to a higher ranking. For those looking for a low-cost open-weight option for basic bug-fixing or coding assistance workflows, MiniMax-M3 can provide that value.
Note: MiniMax-M3’s SWE-bench Verified score is flagged for potential training data contamination.
Developer: MiniMax
AA Intelligence Index: 44
SWE-bench Verified: 80.5%
GPQA Diamond: 92.9%
Context Window: 1,000,000 tokens
Output Speed: 85 tok/s
Blended API $/1M: $0.22
Modalities: Text, Vision
Open-Weight: Yes (MiniMax Community License)
11. Kimi K2.6, for Open-Source Agentic Coding and Swarm Tasks
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Kimi K2.6 | The only model here that can coordinate hundreds of AI instances working simultaneously on the same problem | Hits a limit on how much text it can process in one session faster than most models here; not suited for very long documents or large codebases | Teams building multi-agent AI systems that need many AI instances working in parallel, at an open-source price point |
Kimi K2.6 is Moonshot AI’s open-source flagship, a 1-trillion-parameter Mixture-of-Experts model with only 32 billion active parameters per token, enabling frontier-level output at the inference cost of a 32B dense model. It runs on 4x H100 80GB GPUs in INT4 quantization and supports 300 parallel sub-agents across 4,000 coordinated steps, the highest confirmed agent-swarm throughput in this dataset. SWE-bench Verified comes in at 80.2%, GPQA Diamond at 91.1% (AA-GPQA, June 2026), AIME 2026 at 96.4%, and HLE with tools at 54.0%, which is the highest confirmed score in this dataset on that benchmark.
At $0.60/$2.50 per 1M input/output tokens ($0.70 blended), it is priced below both Claude Sonnet 4.6 and Gemini 3.1 Pro while matching or exceeding them on several key benchmarks. The 256K token context window is the primary constraint, limiting large-codebase traversals or extended agent trajectories that require a 1M context.
Developer: Moonshot AI
AA Intelligence Index: 43
SWE-bench Verified: 80.2%
GPQA Diamond: 91.1%
Context Window: 256K tokens
Output Speed: 74 tok/s
Blended API $/1M: $0.70
Modalities: Text, Vision
Open-Weight: Yes (Modified MIT)
12. Grok 4, for Real-Time Web and Social Intelligence
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Grok 4 | The only model here built to synthesize real-time X (Twitter) posts and live web data in a single response | Sold as a monthly subscription rather than pay-per-use, and not competitive with other models here on software development tasks | PR teams, journalists, and social analysts who need to understand what is happening online right now, not just what happened last week |
Grok 4 is xAI’s current flagship. It was trained on the 200,000-GPU Colossus cluster using reinforcement learning at pretraining scale, achieving 6x the training efficiency of Grok 3. Standard Grok 4 scores 87.7% on GPQA Diamond (AA-GPQA, confirmed) and 69.1% on SWE-bench Verified, with the parallel-reasoning Heavy class reaching up to 72% on that benchmark. The Heavy variant also reaches approximately 50% on HLE, the second-highest confirmed HLE score with tools in this dataset (Kimi K2.6 scoring 54% with tools). Grok 4.3 (April 2026) extended the model with document generation, video input, and 25-language audio APIs.
Its most distinctive capability is DeepSearch mode, which generates iterative web queries and synthesizes multi-source results in a single response, making it the strongest model in this dataset for X (Twitter) data, trending social context, and real-time web synthesis. Pricing is subscription-based: $30/mo for SuperGrok and $300/mo for Heavy, so teams should compare the total cost with Gemini 3.1 Pro before committing to general agentic workloads.
Developer: xAI
AA Intelligence Index: ~47 (est.)
SWE-bench Verified: 69.1%–72%
GPQA Diamond: 87.7%
Context Window: 1,000,000 tokens
Output Speed: N/A
Blended API $/1M: Subscription
Modalities: Text, Vision, Video, Audio
Open-Weight: No
13. Llama 4 Maverick, for Open-Weight Multimodal Deployment
| Model | Biggest Pro | Biggest Con | Best Use Case |
| Llama 4 Maverick | The most widely used AI model that can be downloaded and run entirely on your own servers, with no ongoing vendor fees | Meta has not published standard independent test scores for this model, making it harder to benchmark against others on this list | Enterprise IT teams that need full control over their AI infrastructure and cannot send data to a third-party cloud provider |
Llama 4 Maverick is Meta’s open-weight flagship from the Llama 4 family, benchmarked by Artificial Analysis at an Intelligence Index of 49, the second highest among open-weight models in this dataset. Released in April 2025, Maverick remains the most widely deployed open-weight multimodal model in enterprise environments as of June 2026.
Its primary advantage is ecosystem maturity: Llama 4 Maverick supports text, vision, and audio natively, with a 1M token context window. It is compatible with every major inference framework, including vLLM, Ollama, and Hugging Face Transformers. Teams that need open-weight multimodal capability with broad infrastructure support should use Maverick as their starting point, while newer open-weight models like GLM-5.2 and Kimi K2.6 are evaluated for task-specific benchmark parity.
Developer: Meta
AA Intelligence Index: 49
SWE-bench Verified: N/A
GPQA Diamond: N/A
Context Window: 1,000,000 tokens
Output Speed: Varies
Blended API $/1M: Open-weight
Modalities: Text, Vision, Audio
Open-Weight: Yes (Meta Llama 4 Community License)
14. GPT-5.3 Codex, for Autonomous CLI and Code Execution
| Model | Biggest Pro | Biggest Con | Best Use Case |
| GPT-5.3 Codex | Specifically built to automate computer terminal tasks; the kind that normally require a developer typing commands by hand | Designed for a narrow job and not built for general writing, research, or conversation; also handles less text at once than most models here | DevOps engineers and platform teams looking to automate repetitive command-line and pipeline tasks inside the OpenAI ecosystem |
GPT-5.3 Codex is OpenAI’s coding-specialist model in the GPT-5.x generation, designed for autonomous command-line task execution rather than general-purpose use. Attainment Labs’ February 2026 report places it at 77.3% on Terminal-Bench 2.0 and 56.8% on SWE-bench Pro; GPQA Diamond is confirmed at 91.5% per the Artificial Analysis GPQA Diamond benchmark (benchlm.ai, June 2026). Its 400K context window is among the smallest in this dataset, limiting large-codebase traversal, but at $1.87 blended per 1M tokens and 91 tok/s, it is a cost-effective routing option for isolated CLI and pipeline tasks within the OpenAI API ecosystem.
Note: SWE-bench Verified score for GPT-5.3 Codex on the standardized harness was not available at publication.
Developer: OpenAI
AA Intelligence Index: 44*
SWE-bench Verified: N/A
GPQA Diamond: 91.5%
Context Window: 400,000 tokens
Output Speed: 91 tok/s
Blended API $/1M: $1.87
Modalities: Text, Code
Open-Weight: No
15. DeepSeek V4 Flash, for the Cheapest Frontier-Adjacent Inference
| Model | Biggest Pro | Biggest Con | Best Use Case |
| DeepSeek V4 Flash | The lowest cost of any model on this list by a wide margin; running it at scale costs roughly $2.80 per 10 million words generated | Has not been fully tested across all standard benchmarks, so its reliability on complex or sensitive tasks is less proven than others here | Consumer app developers and startups where keeping AI costs as close to zero as possible is a core business requirement |
DeepSeek V4 Flash is priced at $0.14/$0.28 per 1M input/output tokens (approximately $0.15 blended), making it the cheapest confirmed per-token pricing in this dataset. It delivers 79.0% on SWE-bench Verified, 89.4% on GPQA Diamond (AA-GPQA, June 2026), and 108.9 tok/s output speed, the third-fastest in this dataset after Qwen 3.7 Max (198) and Gemini 3.5 Flash (167). It shares its hybrid CSA+HCA attention architecture with V4 Pro and supports a 1M token context window.
Processing 10M output tokens costs approximately $2.80, making AI-native products at consumer-internet scale financially viable without significant subsidization. The absence of a confirmed AA Intelligence Index score is the main data gap, and teams should validate on task-specific evals before deploying V4 Flash on precision-critical workloads.
Developer: DeepSeek
AA Intelligence Index: N/A
SWE-bench Verified: 79.0%
GPQA Diamond: 89.4%
Context Window: 1,000,000 tokens
Output Speed: 108.9 tok/s
Blended API $/1M: $0.15 est.
Modalities: Text, Code
Open-Weight: Yes (MIT)
Top AI Models by Use Case
Top AI Models for Coding & Software Engineering
| Rank | Model | SWE-bench Score | Key Coding Strength |
| 1 | Claude Fable 5 | 95.0% Verified / 80.3% Pro | Leads both SWE-bench Verified and SWE-bench Pro; the absolute bleeding-edge frontier model |
| 2 | Claude Opus 4.8 | 88.6% Verified / 69.2% Pro | Leads FrontierSWE and PostTrainBench; top Terminal-Bench 2.1 (85.0) |
| 3 | GLM-5.2 | 82.8% Verified | Third on standardized harness, above GPT-5.5; open-weight at $0.90/1M |
| 4 | GPT-5.5 | 82.6% Verified | Leads Terminal-Bench 2.0 (82.7%); broadest tool-use and CLI coverage |
| 5 | DeepSeek V4 Pro | (80.6%) Verified | Leads math and competitive coding benchmarks (Codeforces 3,206, LiveCodeBench 93.5); open-weight MIT at ~$2.18/1M |
Top AI Models for Scientific Reasoning & Research
| Rank | Model | GPQA Diamond | Key Reasoning Strength |
| 1 | Gemini 3.1 Pro | 94.1% | #2 HLE at 46.4%; METR Time Horizons #3 (384.1 min) |
| 2 | Claude Fable 5 | 94.1% | Leads SimpleBench (81.9%); top SWE-bench Pro (80.3%); potentially stronger than Gemini, but the cost raises issues for non-commercial research budgets |
| 3 | Claude Opus 4.8 | 93.6% | Third-highest Anthropic-confirmed GPQA; leads PostTrainBench |
| 4 | GPT-5.5 | 93.5% | Leads BrowseComp (84.4%); strongest multi-source research synthesis |
| 5 | MiniMax-M3 | 92.9% | Fifth-highest GPQA in dataset (AA-GPQA confirmed); open-weight at $0.22/1M |
Top AI Models for Budget-Conscious Teams
| Rank | Model | Price / 1M | Performance Justification |
| 1 | DeepSeek V4 Flash | $0.15 est. | 79.0% SWE-bench, 89.4% GPQA, 108.9 tok/s at the dataset’s lowest confirmed price |
| 2 | MiniMax-M3 | $0.22 blended | 80.5% SWE-bench, 92.9% GPQA, competitive with $4+ models at $0.22 per 1M; open-weight |
| 3 | Kimi K2.6 | $0.70 blended | 80.2% SWE-bench, 91.1% GPQA, 54.0% HLE (tools); open-source with a 256K context limit |
| 4 | GLM-5.2 | $0.90 blended | 82.8% SWE-bench (standardized, 3rd in dataset), 89.5% GPQA; open-weight MIT |
| 5 | DeepSeek V4 Pro | ~$2.18 blended | 80.6% SWE-bench, 90.5% GPQA, leads math and competitive coding; open-weight |
