{ "updated": "2026-06-03", "verifiedDate": "2026-06-03", "note": "FACTUAL fields (pricing, context window, max output, release date) are sourced from each provider's own official docs and reconciled to assets/data/model-figures.json (the single source of truth); where the two disagreed, the officially-sourced figure won. The benchmarks block: swe_bench_verified and gpqa_diamond are reconciled to the official published figure where the provider published one (see model-figures.json) and are otherwise a benchr editorial ESTIMATE; gpqa_diamond is null where no official figure exists. All capability 0–100 scores and latency numbers are benchr EDITORIAL estimates, not lab measurements — see the methodology page. Re-verified 2026-06-03: Claude Opus 4.8/4.7 + Sonnet 4.6 context windows (now 1M), DeepSeek V4-Pro/Flash pricing, Mistral Medium 3.5 license/price, Qwen3.6 open-weight specs.", "models": [ { "id": "claude-opus-4-8", "name": "Claude Opus 4.8", "company": "Anthropic", "type": "frontier", "license": "proprietary", "released": "2026-05-28", "api_name": "claude-opus-4-8", "review_url": "articles/claude-opus-4-8-review", "deprecated": false, "pricing": { "input_per_million": 5.00, "output_per_million": 25.00, "cache_input_per_million": 0.50, "batch_discount": 0.50, "fast_mode_input": 10.00, "fast_mode_output": 50.00 }, "context": { "max_tokens": 1000000, "effective_tokens": 700000, "max_output_tokens": 128000 }, "benchmarks": { "swe_bench_verified": 88.6, "lmsys_arena": 1425, "mmlu": 93.5, "humaneval": 96.0, "math": 93.0, "gpqa_diamond": 93.6, "arc_agi_2": null }, "latency": { "first_token_ms": 700, "tokens_per_second": 68 }, "capabilities": { "coding": 97, "reasoning": 96, "writing": 91, "vision": 86, "long_context": 94, "multilingual": 90 }, "best_for": ["Highest-stakes coding tasks", "Complex multi-step agents", "Architecture decisions", "Production SWE-bench-level work"], "skip_if": ["You need cheap volume", "Sonnet handles your workload", "Speed is the priority — use Fast Mode instead"] }, { "id": "claude-opus-4-7", "name": "Claude Opus 4.7", "company": "Anthropic", "type": "frontier", "license": "proprietary", "released": "2026-04", "api_name": "claude-opus-4-7", "review_url": "articles/claude-opus-4-7-review", "deprecated": false, "pricing": { "input_per_million": 5.00, "output_per_million": 25.00, "cache_input_per_million": 0.50, "batch_discount": 0.50 }, "context": { "max_tokens": 1000000, "effective_tokens": 700000, "max_output_tokens": 128000 }, "benchmarks": { "swe_bench_verified": 87.6, "lmsys_arena": 1410, "mmlu": 92.5, "humaneval": 95.2, "math": 91.8, "gpqa_diamond": 94.2, "arc_agi_2": 24.0 }, "latency": { "first_token_ms": 720, "tokens_per_second": 65 }, "capabilities": { "coding": 96, "reasoning": 96, "writing": 90, "vision": 85, "long_context": 94, "multilingual": 89 }, "best_for": ["Complex coding tasks", "Long-document analysis", "Production agent loops", "Architecture decisions"], "skip_if": ["You need cheap volume", "Simple summarization", "Sonnet covers your workload"] }, { "id": "claude-sonnet-4-6", "name": "Claude Sonnet 4.6", "company": "Anthropic", "type": "mid", "license": "proprietary", "released": "2025-09", "api_name": "claude-sonnet-4-6", "review_url": "articles/claude-sonnet-4-6-review", "deprecated": false, "pricing": { "input_per_million": 3.00, "output_per_million": 15.00, "cache_input_per_million": 0.30, "batch_discount": 0.50 }, "context": { "max_tokens": 1000000, "effective_tokens": 700000, "max_output_tokens": 64000 }, "benchmarks": { "swe_bench_verified": 79.6, "lmsys_arena": 1365, "mmlu": 89.6, "humaneval": 91.5, "math": 86.0, "gpqa_diamond": 89.9, "arc_agi_2": 15.0 }, "latency": { "first_token_ms": 480, "tokens_per_second": 95 }, "capabilities": { "coding": 88, "reasoning": 87, "writing": 89, "vision": 80, "long_context": 91, "multilingual": 86 }, "best_for": ["Production default", "Cost-effective coding", "Bulk content tasks", "Daily-driver API workloads"], "skip_if": ["You need frontier-grade reasoning"] }, { "id": "claude-haiku-4-5", "name": "Claude Haiku 4.5", "company": "Anthropic", "type": "small", "license": "proprietary", "released": "2025-11", "api_name": "claude-haiku-4-5", "review_url": "articles/claude-haiku-4-5-review", "deprecated": false, "pricing": { "input_per_million": 1.00, "output_per_million": 5.00, "cache_input_per_million": 0.10, "batch_discount": 0.50 }, "context": { "max_tokens": 200000, "effective_tokens": 150000, "max_output_tokens": 64000 }, "benchmarks": { "swe_bench_verified": 73.3, "lmsys_arena": 1285, "mmlu": 82.5, "humaneval": 85.0, "math": 77.0, "gpqa_diamond": null, "arc_agi_2": 8.5 }, "latency": { "first_token_ms": 240, "tokens_per_second": 145 }, "capabilities": { "coding": 75, "reasoning": 76, "writing": 80, "vision": 72, "long_context": 82, "multilingual": 78 }, "best_for": ["High-volume simple tasks", "Real-time chat", "Classification, routing, extraction"], "skip_if": ["Complex reasoning needed", "Long-document analysis"] }, { "id": "gpt-5-5", "benchmarks_estimated": ["swe_bench_verified"], "name": "GPT-5.5", "company": "OpenAI", "type": "frontier", "license": "proprietary", "released": "2026-04-23", "api_name": "gpt-5.5", "review_url": "articles/gpt-5-5-review", "deprecated": false, "pricing": { "input_per_million": 5.00, "output_per_million": 30.00, "cache_input_per_million": 0.50, "batch_discount": 0.50, "batch_input": 2.50, "batch_output": 15.00 }, "context": { "max_tokens": 1050000, "effective_tokens": 700000, "max_output_tokens": 128000 }, "benchmarks": { "swe_bench_verified": 84.0, "lmsys_arena": 1418, "mmlu": 93.0, "humaneval": 94.5, "math": 95.0, "gpqa_diamond": null, "arc_agi_2": 22.0 }, "latency": { "first_token_ms": 520, "tokens_per_second": 82 }, "capabilities": { "coding": 93, "reasoning": 95, "writing": 88, "vision": 89, "long_context": 86, "multilingual": 90 }, "best_for": ["Frontier math and reasoning", "Computer use", "Multi-step agents", "Vision + reasoning tasks"], "skip_if": ["Cost-sensitive workloads — use GPT-5", "Quick chat — the price is hard to justify"] }, { "id": "gpt-5", "name": "GPT-5", "company": "OpenAI", "type": "frontier", "license": "proprietary", "released": "2025-08-07", "api_name": "gpt-5", "review_url": "articles/gpt-5-review", "deprecated": false, "pricing": { "input_per_million": 1.25, "output_per_million": 10.00, "cache_input_per_million": null, "batch_discount": 0.50 }, "context": { "max_tokens": 400000, "effective_tokens": 300000, "max_output_tokens": 128000 }, "benchmarks": { "swe_bench_verified": 74.9, "lmsys_arena": 1375, "mmlu": 90.0, "humaneval": 91.0, "math": 87.0, "gpqa_diamond": null, "arc_agi_2": 16.0 }, "latency": { "first_token_ms": 520, "tokens_per_second": 90 }, "capabilities": { "coding": 88, "reasoning": 88, "writing": 86, "vision": 85, "long_context": 78, "multilingual": 87 }, "best_for": ["Production workhorse at a rational price", "Breadth tasks", "Coding agents", "The everyday OpenAI pick"], "skip_if": ["You need the absolute deepest reasoning — use 5.5", "Very large context windows"] }, { "id": "gpt-5-mini", "benchmarks_estimated": ["swe_bench_verified"], "name": "GPT-5 Mini", "company": "OpenAI", "type": "small", "license": "proprietary", "released": "2025-08", "api_name": "gpt-5-mini", "review_url": null, "deprecated": false, "pricing": { "input_per_million": 0.25, "output_per_million": 2.00, "cache_input_per_million": 0.025, "batch_discount": 0.50 }, "context": { "max_tokens": 400000, "effective_tokens": 200000, "max_output_tokens": 128000 }, "benchmarks": { "swe_bench_verified": 48.0, "lmsys_arena": 1275, "mmlu": 82.0, "humaneval": 78.0, "math": 74.0, "gpqa_diamond": null, "arc_agi_2": 6.0 }, "latency": { "first_token_ms": 220, "tokens_per_second": 160 }, "capabilities": { "coding": 72, "reasoning": 74, "writing": 75, "vision": 72, "long_context": 70, "multilingual": 80 }, "best_for": ["Cheap chat at scale", "Simple extraction and routing", "High-volume classification"], "skip_if": ["Production-quality code review", "Complex reasoning", "Long documents"] }, { "id": "gemini-3-5-flash", "benchmarks_estimated": ["swe_bench_verified"], "name": "Gemini 3.5 Flash", "company": "Google", "type": "mid", "license": "proprietary", "released": "2026-05-19", "api_name": "gemini-3.5-flash", "review_url": "articles/gemini-3-5-flash-review", "deprecated": false, "pricing": { "input_per_million": 1.50, "output_per_million": 9.00, "cache_input_per_million": 0.15, "batch_discount": null, "free_tier": true }, "context": { "max_tokens": 1048576, "effective_tokens": 700000, "max_output_tokens": 65536 }, "benchmarks": { "swe_bench_verified": 80.6, "lmsys_arena": 1385, "mmlu": 89.5, "humaneval": 89.0, "math": 87.5, "gpqa_diamond": null, "arc_agi_2": 14.5 }, "latency": { "first_token_ms": 195, "tokens_per_second": 289 }, "capabilities": { "coding": 88, "reasoning": 86, "writing": 84, "vision": 92, "long_context": 90, "multilingual": 91 }, "best_for": ["Coding agents at speed", "Parallel agent execution", "Multimodal tasks", "Default frontier-quality model"], "skip_if": ["You need the deepest single-call reasoning — use Gemini 3.1 Pro"] }, { "id": "gemini-3-1-pro", "name": "Gemini 3.1 Pro", "company": "Google", "type": "frontier", "license": "proprietary", "released": "2026-02", "api_name": "gemini-3.1-pro", "review_url": "articles/gemini-3-1-pro-review", "deprecated": false, "pricing": { "input_per_million": 2.00, "output_per_million": 12.00, "input_per_million_over_200k": 4.00, "output_per_million_over_200k": 18.00, "cache_input_per_million": 0.20, "batch_discount": null }, "context": { "max_tokens": 1000000, "effective_tokens": 800000, "max_output_tokens": 64000 }, "benchmarks": { "swe_bench_verified": 80.6, "lmsys_arena": 1380, "mmlu": 91.0, "humaneval": 90.0, "math": 90.0, "gpqa_diamond": 94.3, "arc_agi_2": 17.0 }, "latency": { "first_token_ms": 380, "tokens_per_second": 108 }, "capabilities": { "coding": 84, "reasoning": 90, "writing": 84, "vision": 95, "long_context": 92, "multilingual": 91 }, "best_for": ["Deep reasoning in the Gemini family", "Long-context vision work", "Workspace integration"], "skip_if": ["Coding agents — Flash is faster and cheaper", "Cost-sensitive workloads — note the over-200K price bump"] }, { "id": "grok-4-3", "benchmarks_estimated": ["swe_bench_verified"], "name": "Grok 4.3", "company": "xAI", "type": "frontier", "license": "proprietary", "released": "2026-04", "api_name": "grok-4-3", "review_url": "articles/grok-4-3-review", "deprecated": false, "pricing": { "input_per_million": 1.25, "output_per_million": 2.50, "cache_input_per_million": 0.20, "batch_discount": null }, "context": { "max_tokens": 1000000, "effective_tokens": 800000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 68.0, "lmsys_arena": 1370, "mmlu": 87.0, "humaneval": 87.0, "math": 84.0, "gpqa_diamond": null, "arc_agi_2": 10.0 }, "latency": { "first_token_ms": 300, "tokens_per_second": 120 }, "capabilities": { "coding": 82, "reasoning": 84, "writing": 80, "vision": 78, "long_context": 85, "multilingual": 75 }, "best_for": ["Real-time data access via X", "Very cheap output tokens", "Long-context tasks on a budget"], "skip_if": ["Coding-first agent work — others edge it out", "Non-English-heavy workloads"] }, { "id": "deepseek-v4-pro", "name": "DeepSeek V4-Pro", "company": "DeepSeek", "type": "frontier-open", "license": "MIT", "released": "2026-04-24", "api_name": "deepseek-v4-pro", "review_url": "articles/deepseek-review", "deprecated": false, "pricing": { "input_per_million": 0.435, "output_per_million": 0.87, "cache_input_per_million": 0.003625, "batch_discount": null, "self_hosted": true }, "context": { "max_tokens": 1000000, "effective_tokens": 700000, "max_output_tokens": 384000 }, "benchmarks": { "swe_bench_verified": 80.6, "lmsys_arena": 1390, "mmlu": 92.8, "humaneval": 93.5, "math": 96.1, "gpqa_diamond": 90.1, "arc_agi_2": 17.5 }, "latency": { "first_token_ms": 380, "tokens_per_second": 95 }, "capabilities": { "coding": 93, "reasoning": 92, "writing": 84, "vision": null, "long_context": 86, "multilingual": 87 }, "best_for": ["Frontier-grade open-weight coding", "Math-heavy work", "Self-hosted production at near-zero API cost"], "skip_if": ["You need vision/multimodal", "You can't manage GPU hosting"] }, { "id": "deepseek-v4-flash", "name": "DeepSeek V4-Flash", "company": "DeepSeek", "type": "open", "license": "MIT", "released": "2026-04-24", "api_name": "deepseek-v4-flash", "review_url": "articles/deepseek-review", "deprecated": false, "pricing": { "input_per_million": 0.14, "output_per_million": 0.28, "cache_input_per_million": 0.0028, "batch_discount": null, "self_hosted": true }, "context": { "max_tokens": 1000000, "effective_tokens": 600000, "max_output_tokens": 384000 }, "benchmarks": { "swe_bench_verified": 79.0, "lmsys_arena": 1350, "mmlu": 88.0, "humaneval": 87.5, "math": 89.0, "gpqa_diamond": 88.1, "arc_agi_2": 11.0 }, "latency": { "first_token_ms": 290, "tokens_per_second": 135 }, "capabilities": { "coding": 86, "reasoning": 86, "writing": 80, "vision": null, "long_context": 88, "multilingual": 86 }, "best_for": ["Cheapest frontier-grade inference on the market", "Self-hosted at low cost", "Volume coding tasks"], "skip_if": ["You need frontier-tier reasoning depth", "Vision-first workflows"] }, { "id": "kimi-k2-6", "name": "Kimi K2.6", "company": "Moonshot AI", "type": "frontier-open", "license": "Modified MIT", "released": "2026-04-20", "api_name": "kimi-k2-6", "review_url": "articles/kimi-review", "deprecated": false, "pricing": { "input_per_million": 0.95, "output_per_million": 4.00, "cache_input_per_million": 0.16, "batch_discount": null, "self_hosted": true }, "context": { "max_tokens": 262144, "effective_tokens": 200000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 80.2, "lmsys_arena": 1355, "mmlu": 85.0, "humaneval": 84.0, "math": 82.0, "gpqa_diamond": 90.5, "arc_agi_2": 9.0 }, "latency": { "first_token_ms": 350, "tokens_per_second": 100 }, "capabilities": { "coding": 80, "reasoning": 78, "writing": 74, "vision": null, "long_context": 78, "multilingual": 82 }, "best_for": ["Mid-range open-weight option", "Multilingual tasks", "Cost-efficient API with self-hosting option"], "skip_if": ["You need the deepest reasoning or best coding", "Vision workloads"] }, { "id": "mistral-large-3", "benchmarks_estimated": ["swe_bench_verified"], "name": "Mistral Large 3", "company": "Mistral", "type": "frontier-open", "license": "Apache 2.0", "released": "2025-12-02", "api_name": "mistral-large-2512", "review_url": "articles/mistral-review", "deprecated": false, "pricing": { "input_per_million": 0.50, "output_per_million": 1.50, "cache_input_per_million": null, "batch_discount": null, "self_hosted": true }, "context": { "max_tokens": 256000, "effective_tokens": 200000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 62.0, "lmsys_arena": 1340, "mmlu": 84.0, "humaneval": 83.0, "math": 79.0, "gpqa_diamond": null, "arc_agi_2": 7.5 }, "latency": { "first_token_ms": 280, "tokens_per_second": 115 }, "capabilities": { "coding": 78, "reasoning": 79, "writing": 78, "vision": null, "long_context": 76, "multilingual": 88 }, "best_for": ["Apache-licensed production workloads", "European data residency", "Very cheap inference with decent reasoning"], "skip_if": ["Coding at frontier quality", "Vision or multimodal workflows"] }, { "id": "mistral-medium-3-5", "benchmarks_estimated": ["swe_bench_verified"], "name": "Mistral Medium 3.5", "company": "Mistral", "type": "frontier-open", "license": "Modified MIT", "released": "2026-04", "api_name": "mistral-medium-3-5", "review_url": null, "deprecated": false, "pricing": { "input_per_million": 1.50, "output_per_million": 7.50, "cache_input_per_million": null, "self_hosted": true }, "context": { "max_tokens": 256000, "effective_tokens": 200000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 77.6, "lmsys_arena": 1368, "mmlu": 89.0, "humaneval": 88.0, "math": 86.0, "gpqa_diamond": null, "arc_agi_2": 12.0 }, "latency": { "first_token_ms": 320, "tokens_per_second": 105 }, "capabilities": { "coding": 86, "reasoning": 87, "writing": 86, "vision": 84, "long_context": 84, "multilingual": 92 }, "best_for": ["European data residency", "Multimodal + reasoning in one model", "Self-hosted on 4 GPUs"], "skip_if": ["You need MoE efficiency for ultra-cheap inference"] }, { "id": "qwen-3-6-27b", "name": "Qwen3.6-27B", "company": "Alibaba", "type": "open", "license": "Apache 2.0", "released": "2026-04-22", "api_name": "Qwen/Qwen3.6-27B", "review_url": "articles/qwen-review", "deprecated": false, "pricing": { "input_per_million": null, "output_per_million": null, "self_hosted": true }, "context": { "max_tokens": 262144, "effective_tokens": 200000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 77.2, "lmsys_arena": 1370, "mmlu": 89.5, "humaneval": 89.0, "math": 87.0, "gpqa_diamond": 87.8, "arc_agi_2": 12.0 }, "latency": { "first_token_ms": 340, "tokens_per_second": 105 }, "capabilities": { "coding": 88, "reasoning": 86, "writing": 82, "vision": 82, "long_context": 84, "multilingual": 95 }, "best_for": ["Multilingual coding (Chinese, Japanese, Korean, Arabic)", "Local inference on consumer GPUs — dense 27B", "Tool-use agent loops at zero API cost"], "skip_if": ["You want a managed hosted API — these are open weights you self-host", "Absolute deepest single-language reasoning"] }, { "id": "llama-4-maverick", "benchmarks_estimated": ["swe_bench_verified"], "name": "Llama 4 Maverick", "company": "Meta", "type": "frontier-open", "license": "Llama 4 Community", "released": "2025-04-05", "api_name": "llama-4-maverick", "review_url": "articles/llama-4-review", "deprecated": false, "pricing": { "input_per_million": null, "output_per_million": null, "self_hosted": true }, "context": { "max_tokens": 1000000, "effective_tokens": 700000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 66.0, "lmsys_arena": 1340, "mmlu": 84.0, "humaneval": 83.0, "math": 80.0, "gpqa_diamond": 69.8, "arc_agi_2": 9.0 }, "latency": { "first_token_ms": 300, "tokens_per_second": 120 }, "capabilities": { "coding": 80, "reasoning": 81, "writing": 76, "vision": 80, "long_context": 85, "multilingual": 80 }, "best_for": ["Best open-weight frontier model", "Self-hosted production at zero licensing cost", "Multimodal at no API cost"], "skip_if": ["You need the very best reasoning or coding", "You can't manage GPU infrastructure"] }, { "id": "llama-4-scout", "benchmarks_estimated": ["swe_bench_verified"], "name": "Llama 4 Scout", "company": "Meta", "type": "open", "license": "Llama 4 Community", "released": "2025-04-05", "api_name": "llama-4-scout", "review_url": "articles/llama-4-review", "deprecated": false, "pricing": { "input_per_million": null, "output_per_million": null, "self_hosted": true }, "context": { "max_tokens": 10000000, "effective_tokens": 2000000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 56.0, "lmsys_arena": 1295, "mmlu": 80.0, "humaneval": 76.0, "math": 74.0, "gpqa_diamond": 57.2, "arc_agi_2": 6.0 }, "latency": { "first_token_ms": 180, "tokens_per_second": 180 }, "capabilities": { "coding": 73, "reasoning": 74, "writing": 70, "vision": 75, "long_context": 92, "multilingual": 78 }, "best_for": ["Ultra-long-context tasks (10M token window)", "Fast self-hosted inference", "Free multimodal at scale"], "skip_if": ["Deep reasoning tasks", "When frontier coding quality is needed"] }, { "id": "phi-4", "benchmarks_estimated": ["swe_bench_verified"], "name": "Phi-4", "company": "Microsoft", "type": "small-open", "license": "MIT", "released": "2025-01", "api_name": "phi-4", "review_url": null, "deprecated": false, "pricing": { "input_per_million": null, "output_per_million": null, "self_hosted": true }, "context": { "max_tokens": 16000, "effective_tokens": 14000, "max_output_tokens": null }, "benchmarks": { "swe_bench_verified": 30.0, "lmsys_arena": 1180, "mmlu": 84.8, "humaneval": 82.6, "math": 80.4, "gpqa_diamond": 56.1, "arc_agi_2": 5.0 }, "latency": { "first_token_ms": 95, "tokens_per_second": 220 }, "capabilities": { "coding": 70, "reasoning": 78, "writing": 74, "vision": null, "long_context": 60, "multilingual": 72 }, "best_for": ["Local inference on consumer hardware", "Edge deployment", "Reasoning at tiny scale"], "skip_if": ["Long-context tasks", "Production-quality writing or coding"] } ], "dimensions": [ { "id": "pricing", "label": "Pricing", "icon": "$", "sub_dimensions": [ { "id": "input_per_million", "label": "Input / 1M tokens", "unit": "$", "format": "currency", "lower_better": true }, { "id": "output_per_million", "label": "Output / 1M tokens", "unit": "$", "format": "currency", "lower_better": true }, { "id": "cache_input_per_million", "label": "Cached input / 1M tokens", "unit": "$", "format": "currency", "lower_better": true } ] }, { "id": "context", "label": "Context window", "icon": "↔", "sub_dimensions": [ { "id": "max_tokens", "label": "Max tokens", "unit": "tokens", "format": "number", "lower_better": false }, { "id": "effective_tokens", "label": "Effective retrieval zone", "unit": "tokens", "format": "number", "lower_better": false }, { "id": "max_output_tokens", "label": "Max output tokens", "unit": "tokens", "format": "number", "lower_better": false } ] }, { "id": "benchmarks", "label": "Benchmarks", "icon": "★", "sub_dimensions": [ { "id": "swe_bench_verified", "label": "SWE-bench Verified", "unit": "%", "format": "percent", "lower_better": false }, { "id": "lmsys_arena", "label": "LMSYS Arena", "unit": "score", "format": "number", "lower_better": false }, { "id": "mmlu", "label": "MMLU", "unit": "%", "format": "percent", "lower_better": false }, { "id": "humaneval", "label": "HumanEval", "unit": "%", "format": "percent", "lower_better": false }, { "id": "math", "label": "MATH", "unit": "%", "format": "percent", "lower_better": false }, { "id": "gpqa_diamond", "label": "GPQA Diamond", "unit": "%", "format": "percent", "lower_better": false }, { "id": "arc_agi_2", "label": "ARC-AGI 2", "unit": "%", "format": "percent", "lower_better": false } ] }, { "id": "latency", "label": "Speed", "icon": "⚡", "sub_dimensions": [ { "id": "first_token_ms", "label": "First token (ms)", "unit": "ms", "format": "number", "lower_better": true }, { "id": "tokens_per_second", "label": "Tokens / second", "unit": "tok/s", "format": "number", "lower_better": false } ] }, { "id": "capabilities", "label": "Capabilities (0–10)", "icon": "◉", "sub_dimensions": [ { "id": "coding", "label": "Coding", "unit": "/10", "format": "number", "lower_better": false }, { "id": "reasoning", "label": "Reasoning", "unit": "/10", "format": "number", "lower_better": false }, { "id": "writing", "label": "Writing", "unit": "/10", "format": "number", "lower_better": false }, { "id": "vision", "label": "Vision", "unit": "/10", "format": "number", "lower_better": false }, { "id": "long_context", "label": "Long context", "unit": "/10", "format": "number", "lower_better": false }, { "id": "multilingual", "label": "Multilingual", "unit": "/10", "format": "number", "lower_better": false } ] } ] }