AI models listing

Top AI Models Master Catalog

Filter by company:

Model Name	Developer / Company	Release Date	Parameters	Context Window	Max Output	Input Price ($/1M)	Output Price ($/1M)	Cache Price ($/1M)	Training Data Cutoff	Thinking / Reasoning Mechanism	Key Benchmarks	Strengths (Pros)	Limitations (Cons)	Recommended Use-Case	Official Source
Gemini 4 Argon	Google DeepMind	30/09/2026	~1.8T – 2.2T MoE (~128B active, estimated)	2,000,000+ tokens (~2M+)	1,000,000 tokens	$5.00 (estimated)	$20.00 (estimated)	$0.50 (Context Cache)	July 2026	Autonomous Deep Reasoning + Multi-agent Architecture	DeepSWE v1.1: 76.5% (estimated) Coding Arena: Internal testing FrontierMath Tier 4: Outstanding	Record output of 1 million tokens Super capabilities in cyber defense & frontier reasoning Supports multi-tier agents for self-correction	Currently in experimental limits (Fairwind Program) Much higher cost compared to the Flash line	Enterprise architecture, source code safety audit, complete system generation	Official Article ↗
Gemini 3.8 Flash Cyber	Google DeepMind	15/09/2026	~35B Dense (Cyber-Specialized)	1,048,576 tokens (~1M)	65,536 tokens	$1.00	$5.00	$0.25 (Cache)	June 2026	Targeted Cyber Reasoning & Vulnerability Scanning	CyberSec-Bench: Top 1 DeepSWE v1.1: 72.8% WebDev Arena: 1590	Specialized in security testing & zero-day detection Fixes security bugs without causing regression	Regular UI/UX coding is not as optimized as standard Flash version	DevSecOps pipeline, automatic code scanning before PR merge	Official Article ↗
Gemini 3.8 Flash	Google DeepMind	08/09/2026	~35B Dense (Gen 3 Flash)	1,048,576 tokens (~1M)	65,536 tokens	$0.75 (promo) / $1.50	$3.75 (promo) / $7.50	$0.075 (promo) / $0.15 (Cache)	June 2026	Dynamic Thinking (Effort Control: Low / Med / High)	DeepSWE v1.1: 73.7% WebDev Arena Elo: 1600+ Artificial Analysis Index: 88.5	Leading cost-performance ratio DeepSWE 73.7% surpasses many flagships Extremely fast code generation speed, context 1M 50% discount option available via Batch/Flex API	Requires clear prompts for multi-tier bash shell	Daily Driver: bug fixing, pair programming, agent refactoring	Official Article ↗
Gemini 3.7 Flash	Google DeepMind	12/08/2026	~28B Dense	1,048,576 tokens (~1M)	65,536 tokens	$0.75 (promo) / $1.50	$3.75 (promo) / $7.50	$0.075 (promo) / $0.15 (Cache)	May 2026	Dynamic Thinking (Configurable thinking budget)	DeepSWE v1.1: 65.3% FrontierCode 1.1: 43.6% WebDev Arena Elo: 1588 AutomationBench: 30.4%	High first-pass accuracy Extremely stable in production Low operational costs	DeepSWE score (65.3%) lower than version 3.8	Batch automated test writing, API backend code generation	Official Article ↗
Gemini 3.5 Flash	Google DeepMind	19/05/2026	~28B Dense	1,048,576 tokens (~1M)	32,768 tokens	$0.50	$2.50	$0.15 (Cache)	February 2026	Fast Stream Reasoning	WebDev Arena Elo: 1545 SWE-bench Lite: 48.2% HumanEval+: 92.4%	Extremely fast response time (TTFT < 300ms) Multimodal (UI images, mockups) very smooth	Does not handle overly complex algorithm bugs	Convert Figma/mockup to HTML/Tailwind, IDE auto-complete	Official Article ↗
Gemini 3.5 Flash-Lite	Google DeepMind	19/05/2026	~10B Dense (Ultra-Lightweight)	1,048,576 tokens (~1M)	16,384 tokens	$0.25	$1.00	$0.05	February 2026	Standard Direct Output	HumanEval: 88.0% MBPP+: 85.2%	Record low cost ($0.25/1M in, $1/1M out) Tens of millions of tokens daily	Limited complex reasoning, prone to errors when coding multiple files	Error log classification, docstring generation, writing regex/SQL	Official Article ↗
Gemini 3.1 Pro	Google DeepMind	19/02/2026	~540B MoE (~64B active, estimated)	2,097,152 tokens (~2.1M)	65,536 tokens	$2.50	$10.00	$0.625 (Cache)	November 2025	Multi-Tier Thinking (Low / Med / High)	SWE-bench Verified: 58.7% Arena Elo: 1572 Math-500: 94.8%	2M token context, perfect for needle-in-a-haystack tasks Robust multi-tier reasoning	Slower than Flash; costs 3-4 times more	Microservices architecture analysis, document lookup for millions of pages	Official Article ↗
Gemini 2.5 Pro	Google DeepMind	17/06/2025	~540B MoE (~64B active, estimated)	2,097,152 tokens (~2.1M)	65,536 tokens	$2.50	$10.00	$0.625	January 2025	Implicit Chain of Thought	SWE-bench Verified: 49.5% Arena Elo: 1530	Multi-modal source code analysis Persistent long-context reasoning	Restricted for new accounts since 09/2026; higher cost compared to Flash series	Maintain legacy enterprise pipeline from mid-2025	Official Article ↗
Gemini 2.5 Flash	Google DeepMind	17/06/2025	~25B Dense	1,048,576 tokens (~1M)	65,536 tokens	$0.35	$1.05	$0.0875	January 2025	High-Speed Direct Inference	HumanEval: 89.2% Arena Elo: 1485	Economical for 2025	Complex reasoning capability is more limited compared to Pro series	Lookup, short snippet generation	Official Article ↗
Gemini 2.0 Flash Thinking	Google DeepMind	19/12/2024	~25B Dense + Dynamic Reasoning	1,048,576 tokens (~1M)	65,536 tokens	$0.00 (Experimental) / Standard	$0.00 / Standard	N/A	August 2024	Explicit CoT Reasoning (Thinking)	AIME 2024: 73.3% Arena Elo: 1520	Foundation of Google's Thinking Process mechanism Superior logic compared to standard 2.0 Flash	Experimental, sometimes leads to infinite reasoning loops	Solve LeetCode algorithms / competitive programming	Official Article ↗
GPT-6.1 Sol	OpenAI	29/09/2026	~1.5T – 2T MoE (~64B active, estimated)	1,050,000 tokens (~1.05M)	32,768 tokens	$2.00	$10.00	$0.10 (Super cheap Cached Input)	June 2026	Adaptive Reasoning (Self-regulating reasoning effort)	DeepSWE v1.1: ~74% WebDev Arena Elo: ~1605 Aider Polyglot: Top Tier	Agentic performance on par with flagship Astra but 5 times cheaper Extremely accurate adherence to coding guidelines and formats Cached input price only $0.10/1M	Output tokens more expensive than Gemini Flash Strict rate limits at lower tiers	Agentic Coding: write complete features, automatic bug-fixing loops	Official Article ↗
GPT-6 Sol	OpenAI	22/09/2026	~1.5T MoE (~64B active, estimated)	1,050,000 tokens (~1.05M)	32,768 tokens	$2.50	$12.50	$0.25 (Cache)	May 2026	Integrated Agentic Reasoning	DeepSWE v1.1: 71.2% Coding Arena Elo: 1595	50% cost reduction compared to GPT-5.6 Sol Tool Calling is very reliable	Quickly replaced by GPT-6.1 Sol after 1 week	Automation agent deployment by the end of September 2026	Official Article ↗
GPT-6 Luna	OpenAI	22/09/2026	~250B MoE / Dense (estimated)	500,000 tokens (~500K)	16,384 tokens	$0.50	$2.00	$0.05 (Cache)	May 2026	Fast Lightweight Reasoning	WebDev Arena Elo: 1555 HumanEval+: 93.8%	Lowest cost in the GPT-6 series Instant response	500K context; does not resolve complex architectural issues	Autocomplete, code explanation, short unit tests	Official Article ↗
GPT-6 Astra (Flagship)	OpenAI	03/09/2026	~3T – 5T MoE (~128B active, estimated)	1,050,000 tokens (~1.05M)	65,536 tokens	$10.00	$50.00	$1.25 (Cache)	May 2026	Frontier Deep Reasoning + Autonomous Computer Use	Terminal-Bench 4.0: Record-breaking FrontierMath Tier 4: 98% Coding Arena Elo: 1625	Smartest of OpenAI Record holder for Terminal-Bench 4.0 and Computer Use FrontierMath 98%	Very expensive ($50/1M output), high latency	Distributed system architecture, kernel/compiler bugs, scientific research	Official Article ↗
GPT-5.6 Sol	OpenAI	09/07/2026	~1.8T MoE (~64B active, estimated)	1,050,000 tokens (~1.05M)	128,000 tokens	$4.00	$20.00	$0.50 (Cache)	March 2026	Max Reasoning / Ultra Mode	DeepSWE v1.1: 53.8% AA Coding Agent Index v1: 80.0 WebDev Arena: 1562	Once the gold standard of 2026 Comprehensive SDK ecosystem	Completely surpassed by GPT-6.1 Sol	Maintain enterprise yet to upgrade to GPT-6	Official Article ↗
GPT-5.6 Terra	OpenAI	09/07/2026	~500B MoE (estimated)	500,000 tokens (~500K)	16,384 tokens	$2.00	$8.00	$0.25 (Cache)	March 2026	Balanced Professional Reasoning	Coding Arena Elo: 1540 SWE-bench Lite: 52.0%	Balances speed and intelligence	Large agentic repo limited compared to Sol	Code review, documentation, standard API integration	Official Article ↗
GPT-5.6 Luna	OpenAI	09/07/2026	~150B MoE / Dense (estimated)	256,000 tokens (~256K)	16,384 tokens	$0.60	$2.40	$0.06 (Cache)	March 2026	Direct High-Throughput	HumanEval: 91.5% Arena Elo: 1515	Cost-optimized for 2026 for GPT-5.6	256K context is small compared to the current 1M standard	Extracting from git commits, generating release notes	Official Article ↗
GPT-5.5	OpenAI	15/05/2026	~1.8T MoE (estimated)	500,000 tokens (~500K)	16,384 tokens	$3.00	$12.00	$0.30	December 2025	Unified Reasoning	SWE-bench Verified: 62.1% Arena Elo: 1548	Was a strong upgrade for ChatGPT and Codex	Set to retire (Retiring 10/14/2026)	Mandatory migration roadmap to GPT-6	Official Article ↗
GPT-5	OpenAI	07/08/2025	~1.8T – 3T MoE (estimated)	400,000 tokens (~400K)	128,000 tokens	$5.00	$15.00	$1.25	March 2025	Foundational Hybrid Reasoning	SWE-bench Verified: 56.4% Arena Elo: 1535	Historic milestone bringing 'Vibe Coding' globally popular	High cost ($15/1M output); surpassed by newer GPT-5.6 / GPT-6 series	Historical models, now replaced	Official Article ↗
o3-mini	OpenAI	31/01/2025	~200B Dense (Reasoning-Specialized)	200,000 tokens (~200K)	100,000 tokens	$1.10	$4.40	$0.55	October 2023	Deliberate Reasoning (Low / Med / High)	AIME 2024: 87.3% Codeforces: 2100+ SWE-bench Verified: 49.3%	Max output 100K Exceptional math/algorithm solving	Not multimodal, relatively slow speed	Solving LeetCode Hard, optimizing heavy algorithms	Official Article ↗
Claude Sonnet 5.5	Anthropic	28/09/2026	~1T MoE (~64B active, estimated)	1,000,000 tokens (~1M)	128,000 tokens	$2.00	$10.00	$0.20 (Cached Input)	June 2026	Adaptive Extended Thinking	DeepSWE v1.1: 73.8% Coding Arena Elo: 1618 Aider Polyglot: Rank 1	New gold standard: fast speed, conventional standards Max output 128K, 1M context Least hallucinations currently	Terminal-Bench still ranks slightly below GPT-6 Astra	Perfect daily coding: Pair programming, Refactoring, PR Review	Official Article ↗
Claude Opus 5.5	Anthropic	22/09/2026	~2T – 3T MoE (~128B active, estimated)	1,000,000 tokens (~1M)	128,000 tokens	$4.00	$20.00	$0.20 (Cached Input)	June 2026	Deep Architectural Extended Thinking + Fast Mode	DeepSWE v1.1: 75.2% SWE-bench Verified: Nearly 100% Coding Arena Elo: 1622	40% cheaper than old Opus 5 Most refined contextual & software architecture thinking Has Fast Mode	Output price is double that of Sonnet 5.5	Senior Code Architect, designing sensitive software systems	Official Article ↗
Claude Fable 5.1	Anthropic	01/09/2026	~3T – 5T MoE (Mythos-Class, estimated)	1,000,000 tokens (~1M)	128,000 tokens	$10.00	$50.00	$0.50 (Cache)	May 2026	Frontier Autonomous Agentic Reasoning	Agentic-SWE Index: Leading Arena Elo: 1612	Dedicated line for long-horizon agentic tasks Self-planning and persistent pursuit of complex goals Mythos-class: highest tier	Expensive; not suitable for simple prompts	End-to-end autonomous agent: self-reading bug reports, building, testing, deploying	Official Article ↗
Claude Opus 5	Anthropic	24/07/2026	~2T MoE (estimated)	1,000,000 tokens (~1M)	128,000 tokens	$5.00	$25.00	$0.50 (Cache)	March 2026	Default Adaptive Thinking	DeepSWE v1.1: 74.0% Coding Arena Elo: 1615	Peak code quality summer 2026 Standardized coding style, clear documentation	Output cost ($25/1M) higher than Sonnet line Opus 5.5 offers better pricing and performance ($20/1M)	Legacy enterprise deployments	Official Article ↗
Claude Opus 4.8	Anthropic	28/05/2026	~1.8T MoE (estimated)	1,000,000 tokens (~1M)	128,000 tokens	$5.00	$25.00	$0.50 (Cache)	February 2026	Deep-Work Reasoning + Dynamic Workflows	SWE-bench Verified: ~90% Coding Arena Elo: 1598	'Honesty upgrade': 4x less code errors missed Dynamic Workflows coordinating multiple sub-agents New Fast Mode is 3x cheaper	Quickly surpassed by generation 5.x	Automated code quality checks, deep-work refactoring	Official Article ↗
Claude Opus 4.7	Anthropic	16/04/2026	~1.8T MoE (estimated)	1,000,000 tokens (~1M)	128,000 tokens	$5.00	$25.00	$0.50 (Cache)	December 2025	xHigh Effort Extended Thinking + Self-Verification	SWE-bench Pro: High but controversial Coding Arena Elo: 1592	'xhigh' effort level for superior deep reasoning Improved vision: understands diagrams, screenshots of code	Some cases of git history referencing detected on SWE-bench Pro	Architecture diagram analysis, reviewing code screenshots	Official Article ↗
Claude Opus 4.6	Anthropic	17/02/2026	~1.5T MoE (estimated)	1,000,000 tokens (~1M)	128,000 tokens	$5.00	$25.00	$0.50 (Cache)	October 2025	Adaptive Thinking + Expanded Computer Use	SWE-bench Verified: ~82% Arena Elo: 1578 (Opus 4.6)	MILESTONE: First time Claude achieves 1M context Deep contextual reasoning and multi-step planning Expanded Computer Use for practical GUI interactions	Higher cost compared to Sonnet line Now legacy, replaced by Opus 4.7, 4.8, and 5.x	Transition milestone: Complex architectural analysis in 1M token context	Official Article ↗
Claude Sonnet 4.6	Anthropic	17/02/2026	~700B MoE (estimated)	1,000,000 tokens (~1M)	128,000 tokens	$3.00	$15.00	$0.30 (Cache)	October 2025	Adaptive Thinking + Agentic Tool Calling	SWE-bench Verified: ~80.5% Coding Arena Elo: 1572	MILESTONE: First time Sonnet achieves 1M context Coding performance surpasses predecessor Opus 4.5 Fast speed and optimal cost/performance ratio	Replaced by Sonnet 4.7 and Sonnet 5.5 with superior performance	Daily programming, large project refactoring within 1M token context	Official Article ↗
Claude Opus 4.5	Anthropic	15/11/2025	~1.5T MoE (estimated)	200,000 tokens (~200K)	32,000 tokens	$15.00	$75.00	$1.50	July 2025	Extended Thinking v2	SWE-bench Verified: 75.5% Arena Elo: 1560	Optimizes token-efficiency for long-term tasks High stability in enterprise workflows	200K context is outdated compared to the 1M standard Price is too high	Legacy enterprise (Surpassed by Opus 4.6+)	Official Article ↗
Claude Sonnet 4.5	Anthropic	25/09/2025	~700B MoE (estimated)	200,000 tokens (~200K)	32,000 tokens	$3.00	$15.00	$0.30	May 2025	Agent-Optimized Reasoning	SWE-bench Verified: 77.2% Arena Elo: 1555	Runs agents continuously for 30+ hours without interruption SWE-bench Verified 77.2% surpassing Opus 4 Beta context 200K → 1M	Officially 200K context (1M only in beta)	Long-term running agents needing high stability, legacy workflows	Official Article ↗
Claude Opus 4.1	Anthropic	15/08/2025	~1.5T MoE (estimated)	200,000 tokens (~200K)	32,000 tokens	$15.00	$75.00	$1.50	March 2025	Precision Deep Reasoning	SWE-bench Verified: 74.5% Arena Elo: 1548	High-accuracy multi-file refactoring Very effective agentic search	Officially retired in August 2026 High cost, 200K context is outdated	Legacy model (Retired 08/2026)	Official Article ↗
Grok 4.7 (Flagship)	xAI	21/09/2026	~1.5T – 2T MoE (~80B active, estimated)	500,000 tokens (~500K)	65,536 tokens	$2.20 ($4.40 if >200K)	$6.60 ($13.20 if >200K)	$0.55 ($1.10 if >200K)	June 2026	Configurable Reasoning (Low / Med / High / xHigh) + Self-Verification	SWE-bench Verified: 96.2% Terminal-Bench: Top tier Arena Elo: 1618	Strongest flagship xAI for coding & agentic Self-Verification reduces hidden bugs Colossus supercomputer → extremely fast speed	Price doubles when context >200K Fewer built-in IDE plugins compared to Claude/OpenAI	Agent for fixing bugs & refactoring large modules, complex algorithms	Official Article ↗
Grok 4.6	xAI	12/08/2026	~1.5T MoE (estimated)	500,000 tokens (~500K)	65,536 tokens	$2.50	$7.50	$0.60 (Cache)	May 2026	Agentic Reasoning Engine	SWE-bench Verified: 95.6% Terminal-Bench 2.1: 88.4% Arena Elo: 1602	Record 95.6% SWE-bench Verified (08/2026) Tool-calling and terminal interaction are responsive	Inherited and upgraded by Grok 4.7	Coding agent for resolving GitHub issues, building test suites	Official Article ↗
Grok 4.5	xAI	16/07/2026	~1.5T MoE (estimated)	500,000 tokens (~500K)	32,768 tokens	$3.00	$9.00	$0.75 (Cache)	April 2026	Integrated Frontier Reasoning	SWE-bench Verified: 86.6% Coding Arena Elo: 1585	Strong shift of xAI towards SW engineering Stable long-term task completion	High output cost compared to generation 4.7	Backend, RESTful/GraphQL APIs, database migration	Official Article ↗
Grok 4.3	xAI	15/04/2026	~1.2T MoE (estimated)	1,000,000 tokens (~1M)	32,768 tokens	$3.50	$10.50	$0.85 (Cache)	January 2026	Long-Context CoT Reasoning	SWE-bench Verified: 78.4% Needle-In-A-Haystack: 99.8%	Rare 1M token context from Grok Can read entire medium-sized repositories	High latency when context exceeds 500K	Browsing, analyzing large codebase structures	Official Article ↗
Grok 4.20 (Agentic Special)	xAI	20/04/2026	~1.5T MoE (estimated)	2,000,000 tokens (~2M)	65,536 tokens	$4.00	$12.00	$1.00 (Cache)	January 2026	Multi-Agent Coordinated Reasoning	Terminal-Bench: 82.5% Arena Elo: 1590	2M context supports multiple sub-agents in parallel Coding for many hours	High operational costs; complex memory management	Autonomous Multi-Agent, legacy migration	Official Article ↗
Grok 4.1 Fast	xAI	10/03/2026	~150B MoE (estimated)	2,000,000 tokens (~2M)	16,384 tokens	$0.50	$1.50	$0.125 (Cache)	November 2025	High-Speed Streamlined Inference	HumanEval: 92.0% MBPP+: 88.5% Arena Elo: 1535	Extremely fast speed at a low cost 2M context for quick log browsing	Limited deep reasoning, prone to logical errors in complex algorithms	Auto-complete, generating boilerplate, analyzing massive server logs	Official Article ↗
Grok 4.0	xAI	15/07/2025	~600B MoE (estimated)	256,000 tokens (~256K)	16,384 tokens	$5.00	$15.00	$1.25	April 2025	Early Agentic Thinking	SWE-bench Verified: 54.2% Arena Elo: 1540	Launch of Grok 4 generation, competing with GPT-5 in 2025	Retired during xAI system cleanup	Historical model (Retired)	Official Article ↗
Grok 3 (Think / Big Brain)	xAI	17/02/2025	~1T MoE (Colossus-Trained, estimated)	131,072 tokens (~128K)	32,768 tokens	$3.00	$12.00	$0.75	November 2024	Explicit CoT Reasoning ('Think Mode')	AIME 2025: 83.5% SWE-bench Verified: 48.0% Arena Elo: 1555	Historical milestone: first time xAI competes evenly with o1/o3 Significant leap in solving math/algorithms	Limited context of 128K; retired on API	Solving complex algorithms, logic analysis	Official Article ↗
Grok 3 mini	xAI	25/02/2025	~70B MoE / Dense (estimated)	131,072 tokens (~128K)	16,384 tokens	$0.60	$2.40	$0.15	November 2024	Lightweight Deliberate Reasoning	HumanEval: 91.0% Arena Elo: 1515	Low-cost mini version, fast reasoning	Small context of 128K	Short code explanation, writing single tests, learning programming	Official Article ↗
Grok 2	xAI	13/08/2024	314B MoE (estimated Grok architecture)	131,072 tokens (~128K)	8,192 tokens	$2.00	$10.00	$0.50	July 2024	Standard Multimodal Pre-Reasoning	HumanEval: 82.5% Arena Elo: 1470	First generation integrating images + X data	Basic coding, prone to hallucination with new frameworks	Early legacy model (Legacy)	Official Article ↗
Muse Spark 1.3 (Flagship)	Meta	02/09/2026	~400B – 500B MoE (estimated internal Meta)	1,000,000 tokens (~1M)	65,536 tokens	$1.25 (Standard) / $0.10 (Contributor)	$4.25 (Standard) / $0.20 (Contributor)	$0.15 (Cache)	May 2026	Proprietary Multimodal Agentic Reasoning (Meta Superintelligence Labs)	Terminal-Bench 2.1: 88.8% DeepSWE v1.1: 75.4% SWE-Atlas QnA: 59.4% Coding Arena Elo: 1612	NEW FLAGSHIP CODING OF META: Transitioning to a closed commercial model Directly operates engine for 'Muse Code' CLI Reduces tool calls by 20% and tokens by 25% compared to Spark 1.2 Terminal-Bench 2.1 reaches 88.8%, DeepSWE 75.4%	Closed model (closed weights); Contributor package shares data with Meta	Top engine for terminal coding agent, resolving real repo bugs, long-term autonomous agent	Official Article ↗
Muse Glimmer (Open-Weights)	Meta	10/08/2026	30B Dense (Official)	500,000 tokens (~500K)	32,768 tokens	$0.30 (via Cloud) / $0 (Self-host)	$0.90 (via Cloud) / $0 (Self-host)	$0.05	April 2026	Efficient Open-Weight Reasoning (30B Dense)	HumanEval+: 93.2% SWE-bench Lite: 58.0% Arena Elo: 1570	Rare open representative of the Muse line: 30B size runs smoothly on a single internal GPU Inherits coding agent capabilities from the Muse family Fully self-hosted secure freedom	Complex reasoning capabilities not as strong as the closed-weight Spark 1.3	Self-hosted coding agent on enterprise workstations, offline code assistant	Official Article ↗
Muse Spark 1.2	Meta	05/08/2026	~400B MoE (estimated)	1,000,000 tokens (~1M)	65,536 tokens	$1.25 (Standard) / $0.10 (Contributor)	$4.25 (Standard) / $0.20 (Contributor)	$0.15 (Cache)	April 2026	Agentic Terminal CoT Reasoning	Terminal-Bench 2.1: 82.9% DeepSWE v1.1: 59.3%	Released simultaneously with the 'Muse Code' terminal tool Achieved 82.9% on Terminal-Bench 2.1 at launch Automatically shrinks and compresses context to prevent memory overflow	Inherited and optimized by Spark 1.3, which outperforms in tool call frequency	Terminal coding, automatic build/test in CLI environments	Official Article ↗
Muse Spark 1.1	Meta	09/07/2026	~350B MoE (estimated)	1,000,000 tokens (~1M)	32,768 tokens	$1.25 (Standard) / $0.10 (Contributor)	$4.25 (Standard) / $0.20 (Contributor)	$0.15 (Cache)	March 2026	Multimodal Foundation Reasoning	Arena Elo: 1585 HumanEval: 92.5%	Marks a historical turning point: Meta officially launches a closed model line competing directly with OpenAI/Claude Multimodal combines reading images with UI design	First version of the Muse line, upgraded to 1.2 and 1.3	First generation foundational model of the Muse system	Official Article ↗
Llama 4 Scout (10M Context)	Meta	05/04/2025	109B MoE (17B active – 16 experts, Official)	10,000,000 tokens (~10M)	65,536 tokens	$0.40 (via Cloud) / $0 (Self-host)	$1.20 (via Cloud) / $0 (Self-host)	$0.10 (Cache)	December 2024	Open-Weight MoE (109B Total / 17B Active) – Ultra-Long Context	Needle-In-A-Haystack: 99.9% at 10M tokens SWE-bench Verified: ~68%	WORLD RECORD OPEN: 10 MILLION TOKENS context champion Loads entire source code repository and annual git commits 17B active runs super smoothly	Algorithmic reasoning extremely difficult compared to specialized models in the Muse Spark line	Reading and auditing entire massive code repositories, querying multi-year commit history	Official Article ↗
Llama 4 Maverick	Meta	05/04/2025	400B MoE (17B active – 128 experts, Official)	1,000,000 tokens (~1M)	65,536 tokens	$0.80 (Cloud) / $0 (Self-host)	$2.40 (Cloud) / $0 (Self-host)	$0.20 (Cache)	December 2024	MoE High-Capacity (400B Total / 17B Active – 128 Experts)	SWE-bench Verified: 75.8% Coding Arena Elo: 1588 HumanEval+: 93.5%	Reasoning power comparable to major commercial flagships 128 experts enable deep specialization for multiple programming languages Standard context of 1M tokens.	Requires large server hardware for full 400B weights self-hosting.	Self-host on private enterprise cloud, build internal AI coding assistant.	Official Article ↗
Llama 3.3 70B	Meta	06/12/2024	70B Dense (Official)	131,072 tokens (~128K)	8,192 tokens	$0.12 (Cloud) / $0 (Self-host)	$0.30 (Cloud) / $0 (Self-host)	$0.03	December 2023	Dense Instruction-Tuned (Optimized from 405B)	HumanEval: 88.6% Arena Elo: 1530 Math-500: 91.2%	Performance on par with the previous 405B version but only 70B in size Smooth operation on a workstation with 2x RTX 4090 / A100 GPUs Global open-weight standard.	Limited context of 128K; max output of 8k restricts large file generation.	On-premise deployment for startups and R&D departments, daily coding.	Official Article ↗
Llama 3.2 90B Vision	Meta	25/09/2024	90B Multimodal Dense (Official)	131,072 tokens (~128K)	8,192 tokens	$0.35 (Cloud) / $0 (Self-host)	$0.90 (Cloud) / $0 (Self-host)	$0.08	December 2023	Multimodal Vision-Language Reasoning	DocVQA: 90.1% WebDev Arena: 1515	Understands UI design images, data flow diagrams, architectural photos Can generate frontend code from open-source image mockups.	Code generation speed is slower than the pure text 70B version; context limited to 128K.	Convert UI mockup images to React/HTML internally.	Official Article ↗
Llama 3.1 405B	Meta	23/07/2024	405B Dense (Official)	131,072 tokens (~128K)	8,192 tokens	$1.50 (Cloud) / $0 (Self-host)	$3.00 (Cloud) / $0 (Self-host)	$0.30	December 2023	Dense Frontier Foundation Architecture	HumanEval: 89.0% Arena Elo: 1540	Largest open-source dense model in AI history (405 billion parameters) Deep foundational programming knowledge, used for synthetic data generation.	Requires an 8x H100 cluster for self-hosting (extremely high initial infrastructure cost).	Distill smaller models, serve as a teacher model for generating synthetic datasets for companies.	Official Article ↗
Mistral Large 4 (Flagship – nicknamed Le Chonk)	Mistral AI	06/10/2026	1T MoE (52B active – Granular MoE, Official)	1,000,000 tokens (~1M)	65,536 tokens	$2.00	$6.00	$0.20 (Cache)	June 2026	Granular Mixture-of-Experts (1T Total / 52B Active) + Frontier Reasoning	Terminal-Bench 4.0: Leading DeepSWE v1.1: 74.8% Coding Arena Elo: 1616+ SWE-bench Verified: 92.5%	LATEST FLAGSHIP OF MISTRAL (10/2026): MoE architecture with 1 trillion parameters (52B active) Extremely powerful for multi-layer agentic reasoning and solving complex software engineering tasks Leads Terminal-Bench 4.0 and achieves DeepSWE v1.1 74.8%.	Requires a large server cluster for full 1T parameter self-hosting; full open-weights release at the end of October.	Distributed software architecture, autonomous agents handling complex repo bugs, enterprise systems.	Official Article ↗
Devstral 2 (Software Engineer)	Mistral AI	09/12/2025	123B Dense (Official)	256,000 tokens (~256K)	32,768 tokens	$0.40	$2.00	$0.10 (Cache)	August 2025	Agentic Software Engineering & Tool-Calling Engine (123B Dense)	SWE-bench Verified: 62.5% HumanEval+: 93.8% Aider Polyglot: Top Tier	Highly specialized for software engineers and agentic workflows Capable of self-planning, debugging, and running automated test terminals Very economical operational costs.	Context of 256K is smaller than the standard 1M of Large 4; max output of 32k.	Automate CI/CD development cycles, automatically resolve GitHub issues, refactor modules.	Official Article ↗
Codestral (codestral-2508)	Mistral AI	15/08/2025	22B Dense (Official)	256,000 tokens (~256K)	32,768 tokens	$0.30	$0.90	$0.05 (Cache)	March 2025	Fill-in-the-Middle (FIM) + Code Specialization (22B Dense)	HumanEval: 91.5% RepoBench: Top Tier FIM Benchmark: 96.0% Arena Elo: 1560	Speed king and Fill-in-the-Middle (FIM): supports over 80 programming languages Can generate code between two code snippets with 96% accuracy Extremely lightweight 22B size, runs smoothly on a single internal GPU.	Not optimized for large system architecture problems or complex multi-step reasoning.	Real-time inline autocomplete on IDEs (VS Code, JetBrains), rapid unit test writing, code snippet generation.	Official Article ↗
Mistral Medium 3.5	Mistral AI	14/05/2026	~70B MoE (estimated)	128,000 tokens (~128K)	16,384 tokens	$0.80	$2.40	$0.15 (Cache)	January 2026	Balanced Enterprise Reasoning Engine	Coding Arena Elo: 1575 SWE-bench Lite: 51.0% HumanEval: 89.2%	Perfect balance between intelligence and cost for daily technical tasks Ability to comply with JSON schema and call API function calling very accurately.	128K context is limited when analyzing large repositories.	Daily programmers, internal code reviews, technical data analysis and extraction.	Official Article ↗
Pixtral Large (124B)	Mistral AI	18/11/2024	124B Multimodal Dense (Official)	128,000 tokens (~128K)	8,192 tokens	$2.00	$6.00	$0.30 (Cache)	July 2024	Multimodal Vision-Language Reasoning (124B Dense)	DocVQA: 93.3% WebDev Arena Elo: 1530 MathVista: 69.0%	Leading multimodal model from Mistral: understands architectural images, flowcharts, UI mockups Analyzes system diagrams and autonomously converts them into structured code.	Code generation speed is slower than pure coding versions; max output of 8,192 tokens.	Convert mockups and diagrams into source code, analyze screenshots of web/app interfaces.	Official Article ↗
DeepSeek-V4 Pro	DeepSeek	13/08/2026	~1T MoE (CSA/HCA, ~48B active, estimated)	1,000,000 tokens (~1M)	384,000 tokens	$0.30 – $1.30 (per hour)	$1.00 – $2.00	$0.08 (Very cheap Cache)	April 2026	Hybrid Attention (CSA/HCA) + Multi-level Reasoning (Think/Non-think)	DeepSWE v1.1: 72.5% SWE-bench Verified: 91.4% Coding Arena Elo: 1602	Max output of up to 384K tokens (record for continuous code generation) CSA/HCA architecture reduces KV cache memory by 70% when running large repos Super cheap API pricing, supports native agentic.	Peak hours may occasionally experience API congestion on the official server.	Top choice for price/performance for backend, agent refactoring large repos, optimizing TCO.	Official Article ↗
DeepSeek-V4 Flash	DeepSeek	13/08/2026	~67B MoE (12B active, estimated)	1,000,000 tokens (~1M)	65,536 tokens	$0.15	$0.60	$0.04 (Cache)	April 2026	Fast Compressed Stream Inference	HumanEval+: 93.0% Arena Elo: 1565	The cheapest model on the planet achieving 1M context Extremely fast token generation (>120 tokens/second) Very strong in Python/Bash scripting.	Reasoning for distributed architecture design problems is still limited.	Automate CI/CD, write bulk unit tests, generate migration scripts at minimal cost.	Official Article ↗
DeepSeek-R1 (Flagship Reasoning)	DeepSeek	20/01/2025	671B MoE (37B active, Official)	128,000 tokens (~128K)	64,000 tokens	$0.55	$2.19	$0.14 (Cache)	July 2024	Large-Scale Reinforcement Learning (Pure CoT Reasoning)	AIME 2024: 79.8% Codeforces: 2050+ SWE-bench Verified: 49.2%	The first pure reasoning open-source model to outperform many closed models Extremely proficient in solving problems, tackling LeetCode Hard algorithms, and debugging hidden logic.	Code generation speed is slow due to long thinking mechanisms; lacks multimodal capabilities.	Solve challenging algorithm problems, debug race conditions and complex logic.	Official Article ↗
DeepSeek-V3	DeepSeek	26/12/2024	671B MoE (37B active, Official)	163,840 tokens (~164K)	8,192 tokens	$0.27	$1.10	$0.07 (Cache)	July 2024	Multi-head Latent Attention (MLA) + DeepSeekMoE (671B / 37B Active)	HumanEval: 82.6% SWE-bench Verified: 42.0% Arena Elo: 1550	'Earthquake' AI at the end of 2024: performance approaching Claude 3.5 at 1/10 the cost Fully open weights (MIT License), huge community support.	Max output limited to 8,192 tokens; context of 164K does not meet the current 1M standard.	Deploy low-cost production API, daily backend coding.	Official Article ↗
DeepSeek-Coder-V2.5	DeepSeek	05/09/2024	236B MoE (21B active, Official)	128,000 tokens (~128K)	8,192 tokens	$0.14	$0.28	$0.03	June 2024	Code-Specialized MoE (236B / 21B Active)	HumanEval: 90.2% Aider Polyglot: Top Tier in 2024	Highly specialized for 338 programming languages Supports professional code context structures (Fill-in-the-Middle).	Older generation than V3/V4; lacks multi-step agentic reasoning capabilities.	Autocomplete source code in IDEs, write multi-language boilerplate.	Official Article ↗
Qwen 3.8-Max (Flagship)	Alibaba Cloud	18/06/2026	~1.2T MoE (estimated)	1,000,000 tokens (~1M)	65,536 tokens	$2.00	$6.00	$0.20 (Cache)	March 2026	Autonomous Agentic Reasoning + Native Multimodal Vision	SWE-bench Verified: 82.4% WebDev Arena Elo: 1605 HumanEval+: 94.2%	The most versatile flagship of Alibaba Cloud: extremely powerful in multimodal and agentic Most accurate conversion of Figma designs/mockup images to React/Tailwind Excellent multilingual support.	Highest API pricing among Chinese model lines.	Frontend coding from UI designs, large-scale agentic tool-calling, multilingual.	Official Article ↗
Qwen 3-Coder-Turbo	Alibaba Cloud	15/05/2026	480B MoE (35B active, Official)	262,144 tokens (~262K)	32,768 tokens	$0.30	$1.00	$0.05 (Cache)	January 2026	Streamlined Code Agent Engine	Aider Polyglot: Top 3 in the industry SWE-bench Lite: 56.5%	Designed specifically for high-speed coding agents Excellent adherence to git diff format and JSON schema Competitive pricing	Context 262K is lower than Max 1M version	Integrated as an engine for CLI coding tools (Aider, OpenCode, Continue)	Official Article ↗
Qwen 2.5-Coder-32B-Instruct	Alibaba Cloud	12/11/2024	32.5B Dense (Official)	131,072 tokens (~128K)	8,192 tokens	$0.66 (Cloud) / $0 (Self-host)	$1.00 (Cloud) / $0 (Self-host)	$0.05	June 2024	Dense Code-Specialized (Open Weights – Apache 2.0)	HumanEval: 92.7% SWE-bench Verified: 39.8% EvalPlus: Top 1 open 32B	'OPEN-SOURCE CODING ICON': runs smoothly on a single RTX 4090 GPU (with quantization) Coding performance comparable to other 70B dense models Completely free for commercial use	Max output only 8k; 128K context limits for large repos	The number 1 model for developers to self-host offline on personal machines or company servers	Official Article ↗
Qwen 2.5-Coder-7B-Instruct	Alibaba Cloud	12/11/2024	7.6B Dense (Official)	131,072 tokens (~128K)	8,192 tokens	$0.20 (Cloud) / $0 (Self-host)	$0.40 (Cloud) / $0 (Self-host)	$0.02	June 2024	Ultra-Lightweight Open Weights	HumanEval: 88.4% MBPP+: 81.5%	Runs smoothly on common laptops (Mac M-series, 8GB VRAM) Extremely fast code generation (>80 tok/s) Well integrated with Ollama/LM Studio	Cannot solve complex system architecture or logic problems	Inline real-time code completion on VS Code / Cursor	Official Article ↗
Qwen 2.5-Max	Alibaba Cloud	20/01/2025	~1T MoE (estimated)	131,072 tokens (~128K)	8,192 tokens	$1.60	$4.80	$0.40	November 2024	Frontier Cloud Foundation MoE	Arena Elo: 1560 MMLU-Pro: 85.5%	Alibaba's flagship cloud platform for early 2025 High general reasoning and multi-domain logical inference capabilities	Surpassed by Qwen 3.8-Max in both context 1M and agentic capabilities	Enterprise business processing combined with multi-functional programming	Official Article ↗
GLM-5.3 (Flagship)	Zhipu AI (Z.ai)	28/08/2026	~700B MoE (estimated)	1,000,000 tokens (~1M)	128,000 tokens	$1.40	$4.40	$0.26 (Prompt Cache)	May 2026	Long-Horizon Systems Reasoning + Open Weights (MIT)	SWE-bench Verified: 88.2% Terminal-Bench: 84.0% Arena Elo: 1598	Expert in 'Vibe Coding' and long-term software engineering techniques No context drift during lengthy sessions MIT open-source for commercial use	Moderate code generation speed (not as fast as DeepSeek Flash)	Long-term agentic engineering, refactoring complex module architectures, enterprises needing open-source code	Official Article ↗
GLM-5.2	Zhipu AI (Z.ai)	10/06/2026	~700B MoE (estimated)	1,000,000 tokens (~1M)	65,536 tokens	$1.20	$3.80	$0.20 (Cache)	February 2026	Systems Engineering Agentic Mode	SWE-bench Verified: 83.5% Arena Elo: 1582	Excellent task load handling for continuous technical tasks Deep integration into domestic and international software development tools	GLM-5.3 has a more complete self-correction feature	Building agent pipelines for enterprise software development	Official Article ↗
GLM-5	Zhipu AI (Z.ai)	20/02/2026	~400B MoE (estimated)	200,000 tokens (~200K)	32,768 tokens	$1.00	$3.20	$0.18	November 2025	Advanced Agentic Reasoning	SWE-bench Verified: 74.0% Arena Elo: 1565	The first generation bringing Zhipu AI into the international agentic coding race High accuracy in function calling (Tool Calling)	200K context is limited compared to versions 5.2/5.3 (1M)	Writing test cases, auto-generating code according to API specs	Official Article ↗
GLM-4-Plus	Zhipu AI (Z.ai)	16/08/2024	~130B Dense (estimated)	128,000 tokens (~128K)	8,192 tokens	$1.40	$1.40	$0.15	June 2024	Flagship Foundation LLM	Arena Elo: 1540 HumanEval: 85.0%	A solid foundational model of Zhipu equivalent to GPT-4 in 2024 Flat input/output pricing, extremely high stability	Lacks modern Extended Thinking mechanism; small max output	Maintaining legacy enterprise systems of Zhipu AI	Official Article ↗
GLM-4-Flash / CodeGeex-4	Zhipu AI (Z.ai)	05/07/2024	9B Dense (Official)	128,000 tokens (~128K)	4,096 tokens	$0.00 (Free basic API) / $0.10	$0.00 / $0.10	N/A	April 2024	Ultra-Fast Free/Cheap Inference (Open Weights)	HumanEval: 82.0%	Free on Zhipu AI's web/basic API CodeGeex-4 specialized for writing code for VS Code/JetBrains extension Super lightweight and fast	Weak in deep logical reasoning and multi-layer architecture problems	Automatic code completion extension (Free Copilot)	Official Article ↗
Kimi K3 (Flagship)	Moonshot AI	16/07/2026	2.8T MoE (16/896 active experts – Record MoE, Official)	1,000,000 tokens (~1M)	65,536 tokens	$3.00	$15.00	$0.30 (Prompt Cache)	March 2026	2.8T MoE (Stable LatentMoE) + Native Multimodal Video/Image	LiveCodeBench: Top SWE-bench Verified: 89.5% Arena Elo: 1610	RECORD 2.8 TRILLION PARAMETERS: The largest open model in the world from Moonshot AI Native multimodal: view debug screen videos and analyze UI interfaces Optimized for GPU kernel, chip design, and game development	Requires a minimum load to unlock API; output price of $15 is relatively high	Frontier engineering: writing CUDA/Triton kernels, chip design, developing complex games	Official Article ↗
Kimi K2.5	Moonshot AI	15/03/2026	~1T MoE (estimated)	1,000,000 tokens (~1M)	32,768 tokens	$2.00	$10.00	$0.20	November 2025	Long-Context Engineering Agent Engine	SWE-bench: 76.0% Needle-In-Haystack: 99.9%	Milestone achieving 1 million tokens context with high data fidelity Initial experimental platform of Kimi Code CLI	Officially discontinued support from 31/08/2026 to focus on K3	Transitional model (Deprecated 08/2026)	Official Article ↗
Kimi K2	Moonshot AI	20/11/2025	~500B MoE (estimated)	500,000 tokens (~500K)	16,384 tokens	$1.50	$6.00	$0.15	July 2025	Reinforcement Learning + Long Context	HumanEval: 89.5% Arena Elo: 1550	The starting model for the K-series focused on software engineering Optimizes long reasoning chains	Support has been discontinued along with the system cleanup in 08/2026	Legacy model (Deprecated)	Official Article ↗
Kimi k1.5	Moonshot AI	15/01/2025	~70B MoE / Dense (estimated)	131,072 tokens (~128K)	32,768 tokens	$1.00	$4.00	$0.10	October 2024	Long-to-Short Context RL Reasoning	AIME 2025: 78.5% Math-500: 93.0% LiveCodeBench: 45.2%	A breakthrough for Moonshot AI in Reinforcement Learning training STEM scores and algorithm solving significantly improved compared to generation v1	128K context has been completely surpassed by K3	Algorithm solving, mathematical logic analysis, and data science	Official Article ↗
Moonshot-v1-128k	Moonshot AI	20/03/2024	~100B Dense (estimated)	131,072 tokens (~128K)	4,096 tokens	$1.50	$1.50	$0.15	January 2024	Pioneer Ultra-Long Context Architecture	Needle-In-Haystack: 99.5% (2024) Arena Elo: 1510	HISTORICAL ICON: The first model to usher in the era of ultra-long context AI in China Can read and understand documents and codebases hundreds of pages long without losing information	Older generation; officially discontinued on 31/08/2026	Classic historical model (Deprecated 08/2026)	Official Article ↗

Benchmark Standard / Metric	Host / Organization	Evaluation Method & Criteria	Significance & Top Models
Terminal-Bench (2.1 & 4.0)	Open-Source Community	Evaluates bash/zsh interactions, package installation, build, and runtime debugging in real shell environments.	Mistral Large 4 leads in 4.0; Muse Spark 1.3 achieves 88.8%. Gold standard for autonomous Terminal Coding Agents.
DeepSWE v1.1	Datacurve & Artificial Analysis	113 long-term software engineering problems on large repositories; strict anti-memorization sandboxing.	Currently the most reliable benchmark: Muse Spark 1.3 (75.4%), Opus 5.5 (75.2%), Mistral Large 4 (74.8%), Gemini 3.8 Flash (73.7%).
SWE-bench Verified (& Lite / Pro)	Princeton NLP	500 real GitHub bug fixes verified by human software engineers in isolated Docker environments.	Classic benchmark for open-source bug-fixing. Grok 4.7 achieves 96.2%, Mistral Large 4 reaches 92.5%, DeepSeek-V4 hits 91.4%.
Fill-in-the-Middle (FIM)	IDE & Editor Benchmarks Community	Inserts code between existing prefix and suffix blocks to test logical cohesion and contextual syntax matching.	Decisive benchmark for inline auto-complete quality in VS Code, Cursor, and JetBrains. Codestral leads with 96.0%.
AA Coding Agent Index	Artificial Analysis (Independent)	Comprehensive composite index: DeepSWE v1.1 + Terminal-Bench 4.0 + SWE-Atlas-QnA.	Evaluates end-to-end AI Agent autonomy: code generation, command line operations, runtime debugging, and repo architecture comprehension.
WebDev & Coding Arena	LMSYS Organization	Blind Elo evaluation: 2 models solve the same real-world coding prompt, voted anonymously by developers worldwide.	Directly reflects global developer satisfaction regarding code formatting, conventions, completeness, and response vibe.
Aider Polyglot	Aider AI	Measures bug fix success rate and clean git diff generation across 6 core languages (Python, JS, TS, Go, Rust, C++).	Evaluates minimal regression and disciplined git diff formatting. Qwen 3 Coder, Devstral 2, and Claude Sonnet 5.5 lead.
Needle-In-A-Haystack (NIAH)	Long Context Evaluation Community	Inserts a random fact or code snippet inside millions of context tokens to test 100% retrieval recall and precision.	Crucial benchmark for ultra-long context models: Llama 4 Scout (10M), Gemini 3.8/4 (1M-2M), Kimi K3 (1M), Mistral Large 4 (1M).
LiveCodeBench	UC Berkeley, MIT & Cornell Team	Continuously updated programming benchmark sourcing fresh problems from LeetCode, Codeforces, and AtCoder to eliminate data contamination.	Unbiased test of genuine algorithmic reasoning on novel problems never seen in pre-training data.
HumanEval / HumanEval+	OpenAI & EvalPlus Team	164 Python coding challenges based on docstrings, augmented by EvalPlus with 80x more test inputs to catch edge-case bugs.	Foundational industry standard for syntax-free function generation. Most modern top-tier models score above 90%.
MBPP / MBPP+ (Mostly Basic Python Problems)	Google Research & EvalPlus Team	974 practical Python coding problems spanning data structures and basic algorithms with strict unit testing.	Measures accuracy in standard Python programming and everyday developer workflows.
CyberSec-Bench	Meta & Purdue University	Evaluates cybersecurity capabilities: zero-day vulnerability detection, preventing insecure code generation, and secure bug remediation.	Crucial standard for DevSecOps and enterprise application security. Gemini 3.8 Flash Cyber and GPT-6 Astra lead.
FrontierMath (Tier 1 – 4)	Epoch AI	Hundreds of original PhD-level research math problems developed by leading mathematicians, unpublished to prevent memorization.	The ultimate frontier benchmark for advanced mathematical and abstract reasoning. Only frontier models (Gemini 4 Argon, Claude Opus 5.5, GPT-6 Astra) break through Tier 4.
Math-500 & AIME (2024 / 2025)	American Invitational Math Exam & Hendrycks Math	500 challenging problems from MATH dataset and standard US AIME competition exams requiring rigorous step-by-step proofs.	Gold standard for mathematical reasoning and Chain-of-Thought thinking mechanisms. DeepSeek-R1, QwQ-32B, and OpenAI o-series excel.
RepoBench & SWE-Atlas QnA	Salesforce Research & Data-Centric AI Community	Evaluates context retrieval across large code repositories (repository-level code retrieval and cross-file in-context learning).	Measures capability to comprehend complex multi-file architectures and identify precise files/functions across tens of thousands of lines of code.
AutomationBench & FrontierCode	Artificial Analysis & Open Agent Benchmarking	Simulates real-world software workflows where AI Agents execute multi-step automation: reading specs, calling APIs, configuring Docker, fixing tests, and committing code.	Measures reliability and autonomy of AI Agents in end-to-end software development workflows without manual step-by-step guidance.
Codeforces Rating / Competition Elo	Codeforces Community	Evaluates models on live competitive programming problem sets and Elo rating system on the world-renowned Codeforces platform.	Measures competitive problem-solving under time constraints and optimal algorithmic time & space complexity.
Multimodal Benchmarks (DocVQA, MathVista)	Stanford, UCLA & AI Vision Consortia	Evaluates comprehension of system architecture diagrams, UI screenshots, workflow flowcharts, and converting Figma designs into frontend code.	Key indicator for frontend/UI-UX engineering and visual technical documentation comprehension. Qwen 3.8-Max, Pixtral Large, and Gemini Flash lead.

Application Scenario / Use-Case	Primary Recommendation #1	Alternative / Open-Source #2	Selection Rationale & Technical Criteria	TCO Cost Strategy
1. Daily Pair-Programming	Claude Sonnet 5.5	Mistral Large 4 / DeepSeek-V4 Pro	DeepSWE > 73.7%, low hallucination, code adheres to conventions; Mistral MoE 1T is extremely intelligent.	Enable Prompt Cache on Sonnet 5.5 ($0.20/1M) or use Mistral Large 4 ($0.20 Cache).
2. Inline Auto-complete & Fill-in-the-Middle (FIM) on IDE	Codestral (codestral-2508)	Qwen 2.5-Coder-7B	FIM accuracy reaches 96.0%, supports >80 languages, extremely low latency, lightweight 22B size.	Self-host Codestral on 1 internal GPU or use API at $0.30 in / $0.90 out.
3. Autonomous Terminal Coding Agent (Agent CLI)	Muse Spark 1.3 (Meta)	Devstral 2 (Mistral) / GPT-6.1 Sol	Terminal-Bench achieves >88%, autonomously runs builds, tests, and fixes code on failure, closely follows git diff.	Use Devstral 2 at $0.40 in / $2.00 out or Meta's Contributor package ($0.10 in).
4. Absolute Security / Self-host On-Premise	Qwen 2.5-Coder-32B	Codestral 22B / Llama 4 Maverick	Runs entirely on internal GPUs (RTX 4090 / A100), no source code leakage outside.	API cost = $0; leverage existing company server hardware.
5. Personal Projects / Ultra Budget Startup	DeepSeek-V4 Flash	GLM-5.3 / Gemini 3.8 Flash	Costs 5-10 times less than US models while achieving 85-90% of flagship performance.	Use DeepSeek-V4 Flash ($0.15/1M in) for 90% of tasks; only call larger models when logic is blocked.
6. Refactor Huge Codebase / Read Entire Large Repo	Llama 4 Scout (10M context)	Mistral Large 4 (1M) / Gemini 3.8 Flash (1M)	Llama 4 Scout can handle 10 million tokens, Mistral Large 4 MoE 1T deeply analyzes architecture.	Load repo once + utilize context caching.
7. Core System Architecture Design & Complex Algorithm	Claude Opus 5.5	Mistral Large 4 / GPT-6 Astra / DeepSeek-R1	Abstract thinking, system modeling, and high-level multi-layer reasoning. DeepSeek-R1 excels in solving problems.	Activate only for Design RFC phase and core algorithm.
8. Automate CI/CD, Generate Tests & Scan Server Logs	DeepSeek-V4 Flash	Devstral 2 / Grok 4.1 Fast	Super fast token generation speed (>120 tok/s), minimal cost for large volumes.	Utilize Batch API or DeepSeek during off-peak hours for an additional 50% reduction.
9. Rapid UI/UX Creation from Mockup / Figma Design	Qwen 3.8-Max	Pixtral Large / Claude Sonnet 5.5	Multimodal deeply understands image layout, generates clean and responsive Tailwind/React code.	Send mockup images at moderate resolution to save image tokens.
10. DevSecOps & Zero-Day Vulnerability Scanning	Gemini 3.8 Flash Cyber	GPT-6 Astra	Gemini Cyber is specialized for network defense and secure vulnerability patching.	Integrate as an automated step in the CI pipeline before deploying to production.