① Conversation/reasoning model comparison
| Model | Context | Capability tier | Scientific reasoning (GPQA) | Coding / Agent | Speed positioning | Multimodal Input | Price (USD / M tokens) |
|---|---|---|---|---|---|---|---|
| lala-super | 200K | International top-tier flagship (frontier): hardest reasoning, coding and agent tasks | Top tier (~93%+) | Flagship (Top Tier on SWE-bench Pro Leaderboard) | Deep reasoning tier | Text + Image + Audio | $4.35 / $21.73 |
| lala-6-pro | 1M | Open-source frontier tier: multimodal conversation flagship, ideal for ultra-long documents / large codebase | Frontier tier | Frontier-class agent capabilities (terminal/OS automation with proven leading-edge performance) | Balanced | Text + Image + Audio | $0.446 / $0.893 |
| lala-6 | 1M | Open-source frontier tier: cost-effective choice for daily conversations, customer service, content generation | Frontier tier | Strong Agent Capability (Leading in Same-tier Testing) | Fast | Text + Image + Audio | $0.149 / $0.298 |
| lala-ultrafast | 1M | Ultra-fast inference tier: high-volume, low-latency scenarios (classification, extraction, simple Q&A) | Comparable first tier (~88%) | Strong (Leading in Same-tier Testing) | Ultra-fast | Text | $0.149 / $0.595 |
| GPT-5.5 (OpenAI) | ~1M | International flagship | 93.6% | SWE-bench Verified 88.7% (official) | Deep reasoning tier | Text + Image + Audio | $5.00 / $30.00 |
| Claude Opus 5.5 (Anthropic) | 200K | International flagship | ~93%+ | SWE-bench Pro 89.9% | Deep reasoning tier | Text + Image | $5.00 / $25.00 |
| Gemini 3.1 Pro (Google) | 1M | International flagship | 94.3% | SWE-bench Verified 80.6% | Balanced | Text + Image + Audio + Video | $2.00 / $12.00 |
| DeepSeek V4 Flash | 1M | Open-source frontier | 88.1% | Terminal-Bench 2.1 90.6% | Ultra-fast | Text | $0.14 / $0.28 |
| Llama 4 Maverick (Meta) | 1M | Open-source general-purpose | 69.8% | LiveCodeBench 43.4% | Fast | Text + Image | $0.15 / $0.60 |
lala series capability tiers positioned by official published benchmarks of comparable international models; exact scores vary by test configuration and are for magnitude reference only. Pricing reflects actual lala API charges (see model pricing page).
② lala full family capability overview
| Model | Type | Core Capability | Key specifications | Use cases |
|---|---|---|---|---|
| lala-6 | Conversation | Multimodal conversation, reasoning, coding | 1M context | Customer Service, Content Generation, Daily Assistant |
| lala-6-pro | Conversation | Flagship Multimodal Conversation, Complex Reasoning | 1M context | Long document analysis, large codebase, high-difficulty tasks |
| lala-ultrafast | Conversation | Ultra-fast Inference | 1M context, pay-per-minute pricing | Bulk Classification, Extraction, Real-time Q&A |
| lala-super | Conversation | International top-tier flagship reasoning | Deep reasoning, superior coding | Most Challenging Reasoning, Agent, Coding Tasks |
| lala-rag | Vector | Text Vectorization | Standard OpenAI embeddings format | RAG retrieval, semantic search, knowledge base |
| lala-decision | Research | Structured judgment (no text generation): Choice / Score / Noul three answer types with calibrated confidence; supports image judgment (up to 4 images) | 1–5 seconds; images ≤4 | Ticket Routing, Intent Recognition, Content Moderation, Document Review |
| lala-image | Image | Image Generation (Quality Priority) | Per image billing | Product images, posters, social media assets |
| lala-video | Video | Video generation (auto voice-over / sound effects / background music) | 4–30 seconds; 480p/720p/1080p | Commercials, Product Demos, Social Media Short Videos |
| lala-speech | Voice | High-definition voice synthesis | 80+ voice tones, 9 emotions | IVR, audio content, voice customer service |
| lala-music | Music | Music generation | Per song billing | Brand jingle, background music |
| lala-lyric | Lyrics | Lyrics generation (complete song / edit and continue) | Per request billing | Song composition, marketing copy rhyming |
③ Detailed benchmark score comparison
| Benchmark (Performance Testing) | lala-super | lala-6-pro | lala-6 | lala-ultrafast | GPT-5.5 | Claude Opus 5.5 | Gemini 3.1 Pro | DeepSeek V4 Flash | Llama 4 Maverick |
|---|---|---|---|---|---|---|---|---|---|
| GPQA Diamond (Doctoral-level Scientific Reasoning) | ~91–94% | — | — | ~88–91% | 93.6% | 90.6% | 94.3% | 88.1–90.9% | 69.8% |
| SWE-bench Pro (Real repository coding) | ~85–90% | — | — | — | 58.6% | 89.9% | — | — | — |
| SWE-bench Verified (agentic coding) | Flagship | — | — | — | 88.7% (official) | — | 80.6% | — | — |
| DeepSWE (Advanced Coding) | ~70–74% | ~70–72% | ~68% | — | — | 74.2% | — | — | — |
| Terminal-Bench 2.1 (terminal agent) | ~87–88% | ~88–90% | ~87–88% | ~90% | — | 87.6% | — | 90.6% | — |
| OSWorld (Computer Operation Agent) | ~82% | ~80–82% | ~80–81% | — | 78.7% | 81.8% | — | — | — |
| Mathematics (AIME / FrontierMath Level) | ~90%+ | Strong | Strong | Strong | — | FrontierMath T1-3 91.2% | AIME 96.7%（3 Pro） | AIME 93%+ | — |
| HLE (Human Last Exam, Ultra-difficult Comprehensive) | ~61–68% | — | ~35% | — | — | 67.7% (with tools) | 51.4% | — | — |
| LiveCodeBench Pro (Competitive Coding Elo) | — | — | — | — | — | — | 2887 Elo (Top Ranked) | — | 43.4% |
| Multimodal (MMMU-Pro Visual Understanding) | ~89% (charts) | — | ~73% | — | — | 89.0%（Chartography） | 80.5% | — | — |
| Long documents (MRCR v2 / long context retrieval) | 200K full window reliable | 1M full window | 1M full window | 1M full window | — | — | 77% (128k average) | 393K output | 1M window |
| AA Intelligence Index (Composite Intelligence Index, independent assessment) | ~58 (top-tier) | ~46 | ~38 | — | 60 | 58 | #1 Tier | — | — |
| MMLU / GMMLU (Comprehensive Knowledge) | ~94% (multilingual) | — | — | — | 92.4% | GMMLU 94.3% | MMMLU 92.6% | — | 85.5% |
| SciCode (research coding) | ~66–67% | — | ~51% | — | — | 66.9% | — | — | — |
| ProgramBench (Program Synthesis) | ~90% | ~26% | ~26% | — | — | 91.2% | — | — | — |
| CyberGym (Cybersecurity Red Team) | — | ~94% | ~95% | ~88% | — | — | — | 88.1% | — |
| Toolathlon (real-world tool chain agent) | — | ~77% | ~74% | — | — | — | — | — | — |
| AutomationBench (Enterprise Process Automation) | ~40% | — | — | — | — | 40.0% | — | — | — |
| GDPval (Enterprise Knowledge Work Elo) | ~1800+ | ~1650–1700 | — | — | — | 1846 Elo (Top Ranked) | — | — | — |
| Long-form reasoning (AA-LCR long context reasoning) | — | ~75 | ~74 | — | — | — | — | — | — |
"—" = no publicly comparable data for this model. "~" range = lala series magnitude positioning by comparable independent evaluations (BenchLM, Artificial Analysis, Vals AI, llm-stats, etc.); official exact scores may vary by test configuration (effort level, scaffold, repetitions) with 5–10 point variance, cross-source comparisons for reference only.
④ Quick visual guide: performance and pricing
Scientific reasoning GPQA Diamond (%)
lala series displayed by range median
Comprehensive intelligence index (Artificial Analysis)
Independent third-party evaluation, 0–70
Input price (USD / M tokens, logarithmic scale)
lala pricing significantly below comparable international models
Price-Performance: Intelligence Index ÷ Input Price
Intelligence score per $1 (M tokens)
lala seriesInternational model
Chart Data Table ③: "~" represents Lala tier positioning median; price-performance chart uses each model's AA Intelligence Index divided by input price, DeepSeek index uses same-tier independent benchmark value.
How to Get Started?
Invocation method same as code example seeDeveloper documentation; see complete pricingModel pricing。