The current best model for each task

Frontier reasoning

The most capable model for hard multi-step problems — graduate-level science, long-horizon code agents, complex math.

Claude Fable 5

Anthropic

premium tier · highest AA Intelligence Index score

Tops Artificial Analysis Intelligence Index v4.1 at 60 with adaptive reasoning + Opus 4.8 fallback. Wins on agentic benchmarks (Terminal-Bench, GDPval) and scientific reasoning (HLE, CritPt).

Frontier general-purpose

Best balance of capability + speed + tool use for everyday production workloads at the frontier tier.

GPT-5.5

OpenAI

$5 in / $30 out per 1M

Frontier scoring (AA Index 55 at xhigh effort), 1M-context multimodal with strong agentic computer-use. The default if you want one model that does everything well.

Agentic coding

Best at autonomous code generation across a real repo — multi-file edits, terminal use, test-fixing.

GPT-5.3-Codex

OpenAI

$1.75 in / $14 out per 1M

SOTA on SWE-bench Pro at launch (Feb 2026). 400K context with strong Terminal-Bench 2.0 + OSWorld-Verified scores. The model Cursor / Devin / Codex CLI default to for hard code tasks.

Cost-efficient reasoning

Best reasoning capability per dollar — the model to use when budgets matter but you still need real thinking.

DeepSeek V4 Pro

DeepSeek

$2.10 in / $4.40 out per 1M

AA Index 44 at max effort — frontier-adjacent capability at roughly a quarter of GPT-5.5's price. Open-weight 1T-param MoE, so you can self-host if you have the GPUs.

Cheapest model meeting Budget bar (CoI)

The lowest blended $/1M-token model whose AA Intelligence Index ≥ 25 — i.e. that still does real work.

GPT-5.4 Nano

OpenAI

$0.20 in / $1.25 out per 1M ($0.46 blended)

Current CoI budget-capable champion. AA Index 38 at xhigh effort, $0.46 blended — cheaper than any other model that clears the 25 capability gate on the AA scale. Updated by the engine daily.

Best open-weight LLM

Highest-capability model whose weights are downloadable — for self-hosting, fine-tuning, on-prem deployment.

Kimi K2 Thinking

Moonshot AI

MIT license · 1T total / 32B active

Topped the AA Intelligence Index for open weights at launch (Nov 2025) with 33 on v4.1. Modified MIT license — commercial use OK. 256K context, agentic tool-use first-class.

Long-context champion

Best model for genuinely long inputs — multi-document analysis, large codebases, hours of transcripts.

Gemini 3.1 Pro

Google

2M context · $2 in / $12 out per 1M

2M-token native context — twice GPT-5.5's window. Strong recall on long-context benchmarks (AA-LCR). At-tier pricing means it's the obvious default when context > 1M tokens.

Cheapest model that's still serviceable

For high-volume / latency-sensitive uses where capability matters less — classification, routing, simple Q&A.

Amazon Nova Micro

Amazon

$0.035 in / $0.14 out per 1M ($0.06 blended)

Cheapest non-trivial model in the basket. 305 tokens/sec — among the fastest. AA Index 5 means "don't ask it to reason", but it nails high-volume routing / classification work for ~6¢ per million blended tokens.

Photorealism

Best at indistinguishable-from-photograph realism — faces, lighting, materials, fine detail.

Nano Banana Pro (Gemini 3 Pro Image)

Google DeepMind

premium per-image · 4K output

Widely cited as the strongest image model of late 2025/2026. Up to 4K, industry-leading text rendering, 14 reference images, 5-subject identity locking, Search grounding for factual realism. Won the late-2025 arena leaderboards.

Text-in-image / typography

Best at rendering legible, correctly-spelled, well-laid-out text inside generated images — posters, logos, slides.

Ideogram 3.0

Ideogram

from $15/mo · API ~$0.06/image

Best-in-class embedded text rendering since 2024 and still the default for typography / poster / logo work. Multiple style modes; speed tiers; clean API.

In-context image editing

Best at editing an existing image with text + image prompts — add/remove objects, change style, swap subjects.

FLUX.1 Kontext (Pro/Max)

Black Forest Labs

API per-image · Dev variant open weights

Dedicated in-context editing suite. Text + image prompted editing with strong consistency across edits — the standard for AI-assisted Photoshop-style workflows.

Cheap & fast generation

Best $/image for high-volume gen — thumbnails, product variants, mockups, programmatic art.

Luma Photon Flash

Luma Labs

~$0.002 per 1080p image

Universal Transformer (non-diffusion) at radical cost-efficiency. ~$0.002 per 1080p image — orders of magnitude cheaper than Imagen Ultra or Nano Banana Pro, with quality that's competitive for non-portrait work.

Best open-weight image

Best image model you can download and run on your own GPU.

HiDream-O1-Image

HiDream AI

8B unified transformer · free self-host

May 2026 release; claimed to beat GPT Image 2 and Seedream 4.0 on quality benchmarks. Open weights — true SOTA self-host. 8B unified transformer architecture.

Aesthetic / artistic style

Best at painterly, illustrative, stylized output — for art, concept work, distinctive look.

Midjourney V8.1

Midjourney

subscription only · Basic $10/mo

Aesthetic leader for 3+ years. V8.1 incrementally improves realism + consistency over V7 while keeping the distinctive Midjourney look. App-gated by design.

Audio-visual co-generation

Best at generating video + synchronized audio (dialogue, SFX, music) in a single pass — table stakes since Veo 3.

Veo 3.1

Google DeepMind

$0.40/s w/ audio · $0.03/s Lite

Native synchronized 48kHz audio with tight lip-sync; 4/6/8s clips at 24fps, 720p/1080p (4K premium). Made A/V co-generation table stakes; still the highest-quality result when audio matters.

Character / world consistency

Best at keeping the same character / setting consistent across cuts, shots, and durations — the hard problem for narrative video.

Runway Gen-4

Runway

$0.01/credit · plans $12–$188/mo

Breakthrough single-image character/world consistency at Gen-4's launch (Mar 2025); Motion Brush and Director Mode camera control. Still the go-to when you need a recognizable face or setting across multiple shots.

Cinematic HDR

Best for cinematic camera work, depth, dynamic range — film-quality output, not social-clip output.

Luma Ray3.14

Luma Labs

Plus $30 / Pro $90 / Ultra $300 per mo

First model with native 16-bit HDR (EXR export). Reasoning-driven refinement, cinematic camera motion, durations up to 20s. HDR leader — though no native audio generation yet.

Best $/sec for non-premium video

Best video model at the budget tier — for social clips, drafts, prototyping at scale.

Hailuo 02

MiniMax

~$0.045/s at 768p · ~$0.28/generation

Native 1080p, strong physics, consistently top-3 on AA's video arena. Cheapest mainstream video model at this quality level. 6s or 10s clip durations.

Best open-weight video

Best video model you can run on your own hardware.

LTX-2

Lightricks

open weights · API also available

First production-ready truly open-weight model with single-pass synchronized audio. Native 4K@50fps, up to 20s. Asymmetric dual-stream transformer. The open answer to Veo-style A/V generation.

Most expressive narration

Best at conveying emotion, nuance, character voice — for audiobooks, narration, dialogue.

Eleven v3

ElevenLabs

character credits · API + app

Inline Audio Tags for fine-grained emotion control. 70+ languages. Best in class for expressive narration / dialogue work, especially with the v3 GA release in 2026.

Real-time / lowest-latency

Best for live voice agents, conversational AI, call centers — where first-chunk latency matters more than expressiveness.

Cartesia Sonic 3

Cartesia

~$35/1M chars

Sub-100ms first-chunk latency. State-space model architecture. 42 languages, natural laughter, voice cloning. The latency leader for voice agents.

Broadest multilingual coverage

Best when you need lots of languages with consistent voice quality — global apps, accessibility tools.

Azure AI Speech (Neural / HD)

Microsoft

~$15/1M chars

400+ neural voices across 140+ languages — broadest enterprise coverage. Custom neural voice for brand voices. Emotionally-aware HD voices for the latest tier.

Best open-weight TTS

Best TTS you can self-host — for privacy-sensitive deployments, offline use, edge.

Kokoro-82M

hexgrad (open)

Apache-2.0 · <$1/1M chars hosted

82M params, 8 languages, 54 voices. Punches well above weight at this size. Apache-2.0 license — true commercial-friendly self-host.

Highest batch accuracy

Best for offline batch transcription where you need the lowest word error rate — legal, medical, archival.

ElevenLabs Scribe v2 (Batch)

ElevenLabs

premium per-min

Lowest WER on benchmarks; 90–99 languages; ~97% accuracy on English; built-in diarization. The accuracy leader when latency doesn't matter.

Real-time streaming

Best for live captioning, voice agents, meeting bots — where low latency matters more than perfect accuracy.

Deepgram Nova-3

Deepgram

~$0.0077/min streaming

~5.3% English WER with sub-second streaming. Speed + price leader. Real-time customization for domain vocabularies.

Multilingual / open-weight

Best open or multilingual STT — when you need wide language coverage or offline operation.

NVIDIA Parakeet TDT 0.6B v3

NVIDIA

open weights (NeMo / HF)

#1 on HuggingFace Open ASR Leaderboard with the highest throughput. FastConformer architecture. The cheap-high-throughput batch option when you can self-host.

Song generation (vocals + instrumentation)

Best at producing a full song with coherent structure, vocals, and instrumentation.

Suno v5.5

Suno

Pro / Premier subscription · credit-based

Most expressive Suno yet. Personal voice cloning, custom-style fine-tuning, My Taste personalization. Leading consumer AI song generator with coherent multi-minute output.

Instrumental composition

Best for production-grade instrumental tracks — film scoring, background, ambient, royalty-free use.

Google Lyria 2

Google DeepMind

Vertex AI API

High-fidelity 48 kHz stereo instrumental. Broad genre control + SynthID watermarking. Enterprise-grade for commercial production.

Text-to-SFX / sound design

Best for generating sound effects, ambient soundscapes, foley from text prompts.

ElevenLabs Sound Effects

ElevenLabs

per-generation credits

Leading text-to-sound-effects generator. Strong for game audio, video sound design, ambient soundscapes.

OCR / document understanding

Best for converting documents (PDFs, scans, screenshots) into structured text + tables + math.

Mistral OCR 3

Mistral AI

$2/1k pages ($1 batch)

Markdown output with HTML table reconstruction; handles forms, handwriting, math, multilingual. 74% win rate over OCR 2 at launch. Best price point for high-volume document work.

General-purpose embeddings

Best vector embeddings for semantic search, RAG, recommendation, similarity tasks.

voyage-3-large

Voyage AI

first 200M tokens free

SOTA general/multilingual. Matryoshka dims, 32K context, int8/binary quantization for cheap serving. Default for serious RAG pipelines in 2026.

Reranker

Best for the rerank stage of a retrieval pipeline — second-pass scoring of query/document pairs.

Cohere Rerank 3.5

Cohere

$2.00 per 1,000 searches

Cross-encoder; 100+ languages, 4K context. The leading multilingual enterprise reranker with predictable pricing.

3D generation

Best for generating textured 3D meshes from text or image — for games, AR, product viz.

Rodin Gen-2.5 (Hyper3D)

Deemos Tech

web + API (fal) · credit/subscription

Diffusion Transformer (CLAY), quad-mesh + PBR, sculpt-level detail above 10M polygons. Top quality-and-speed in 2025-2026 comparisons.

Image / video segmentation

Best for separating objects from backgrounds — masks, video object tracking, scene parsing.

SAM 3.1 (Segment Anything 3.1)

Meta AI

open weights · free

Unified image + video with text / exemplar / visual prompts. Real-time object multiplex tracking. The default for any vision pipeline needing masks. Open weights.