Best LLMs for coding: developer favorites

[Last updated: March 2026]

The best LLMs that developers use for coding stand out by combining deep understanding of programming languages with practical capabilities that enhance a developer's workflow. They solve complex problems and deliver code that can be used to build production applications faster—not just vibe code a prototype.

The landscape shifted fast again in early 2026. Anthropic launched Opus 4.6 and Sonnet 4.6 a few weeks apart in February. Google answered with Gemini 3.1 Pro days later. Then OpenAI dropped GPT-5.4 on March 5—its first general-purpose model with native computer use. Four frontier releases in under five weeks.

Many of these coding LLMs are available to use in developer tools like Claude Code, Cursor, Codex, and Copilot (though Microsoft's GitHub Copilot is notoriously disliked). Software developers tend to have a favorite LLM for code completion and use a few different models depending on the specific task.

Here are some of the LLMs developers use the most for coding.

Developers’ favorite LLMs for coding

Through all of it, Claude's position with developers has only strengthened. The Pragmatic Engineer's March 2026 survey of 906 software engineers found that 46% named Claude Code as the tool they love most—nearly 2.5x the share of Cursor (19%) and more than 5x GitHub Copilot (9%).

Claude Code went from well-liked underdog to the most-used AI coding tool in just eight months since its May 2025 launch, with Anthropic's Claude Sonnet 4.6 and Opus 4.6 dominating coding task model preferences by a significant margin.

There may be a new coding LLM leading the pack by the time you read this, but there are some consistent favorites software engineers rely on. 

Anthropic Claude coding LLMs

Claude Opus 4.6

Claude Opus 4.6: Released February 5, 2026, Opus 4.6 is Anthropic's most capable model for complex, long-horizon work. While Sonnet 4.6 now matches or edges it out on many practical benchmarks, Opus 4.6 remains the go-to for tasks that demand maximum reasoning depth, extended autonomous operation, and multi-agent coordination. 

Key Capabilities:

• SWE-bench Verified: 80.8%—narrowly leading all models on real-world software engineering tasks at time of launch
• Agent teams: Split large tasks across multiple agents that each own their piece and coordinate directly—a meaningful step beyond sequential single-agent workflows
• ARC-AGI-2: 68.8%, up from 37.6% on Opus 4.5—an 83% improvement on novel problem-solving
• BrowseComp: 84.0% on agentic web search, up from 67.8% on Opus 4.5
• Long context: 1M token context window (beta), with 76% accuracy on 8-needle 1M MRCR v2—far ahead of Sonnet 4.5's 18.5% on the same test
• Fast mode: Up to 2.5x faster output generation at premium pricing ($30/$150 per MTok)—same intelligence, faster inference
• Finance Agent benchmark: Leads all models on financial analyst task evaluation
• API pricing: $15 per million input tokens, $75 per million output tokens

Anthropic recommends Sonnet 4.6 for roughly 90% of coding tasks given its price-to-performance ratio. Reserve Opus 4.6 for the top 10%—complex multi-agent pipelines, zero-error financial analysis, large-scale infrastructure planning, or tasks where maximum context utilization is essential.

Claude Sonnet 4.6

Claude Sonnet 4.6:  Released February 17, 2026, Sonnet 4.6 is now the default model on claude.ai for Free and Pro users. It delivers Opus-class performance on most tasks at a fraction of the cost—developers with early access preferred it to Sonnet 4.5 by a wide margin, and even preferred it over Opus 4.5 (Anthropic's previous frontier model) 59% of the time in Claude Code testing. 

Key capabilities:

• SWE-bench Verified: 79.6% in standard mode—a meaningful step up from Sonnet 4.5's 77.2%
• Computer use leadership: 72.5% on OSWorld-Verified, with major improvements in prompt injection resistance vs. prior Sonnet models
• Adaptive thinking: New default reasoning mode that dynamically decides when and how much to think—no need to manually enable extended thinking for most tasks
• 1M token context (beta): Enough to hold entire codebases, lengthy contracts, or dozens of research papers in a single request
• 64K output tokens: Supports generation of complete applications, comprehensive refactors, and large documentation sets
• Better long-session coding: Users report fewer false claims of success, fewer hallucinations, and less tendency to over-engineer or duplicate logic vs. Sonnet 4.5
• OfficeQA: Matches Opus 4.6 on enterprise document comprehension—charts, PDFs, tables—making it practical for non-engineering knowledge work too
• API pricing: $3 per million input tokens, $15 per million output tokens (same as Sonnet 4.5)

In Claude Code testing, users rated Sonnet 4.6 as significantly better at reading context before modifying code, consolidating shared logic rather than duplicating it, and following multi-step instructions through to completion. For most production coding workflows, it's the default choice.

OpenAI GPT LLMs for coding

GPT-5.4: Released March 5, 2026, GPT-5.4 is OpenAI's most capable and efficient frontier model to date. It absorbs the industry-leading coding capabilities of GPT-5.3-Codex and extends them into a general-purpose model built for sustained professional workflows—the first OpenAI model with native, state-of-the-art computer use capabilities baked in. 

Key capabilities:

• Native computer use: First general-purpose OpenAI model that can autonomously navigate applications, operate computers, and execute multi-step workflows across software environments via Codex and the API
• 1M token context: Available in the API and Codex—enabling agents to plan, execute, and verify tasks across long horizons
• Token efficiency: Uses significantly fewer tokens (up to 47% fewer on tool-heavy tasks) compared to GPT-5.2, translating to faster speeds and lower costs at scale
• Upfront planning: In ChatGPT, GPT-5.4 Thinking displays its reasoning plan before completing a response—users can course-correct mid-response without starting over
• Tool search: New API tool management system that helps agents find and use the right tools more efficiently without sacrificing intelligence
• Reduced hallucinations: Individual claims are 33% less likely to be false compared to GPT-5.2; full responses are 18% less likely to contain any errors
• Knowledge work: 83% on OpenAI's GDPval benchmark, and record scores on OSWorld-Verified and WebArena Verified
• Legal and finance: 91% on BigLaw Bench; preferred by 87.3% of investment banking modelers over GPT-5.2
• GPT-5.4 Pro: Higher-performance variant for ChatGPT Pro and Enterprise users—89.3% on BrowseComp, 83.3% on ARC-AGI-2 

GPT-5.4 is rolling out across ChatGPT (as GPT-5.4 Thinking for Plus, Team, and Pro users), the API (as gpt-5.4), and Codex—where it replaces GPT-5.3-Codex as the primary model. For front-end development, document-heavy professional workflows, and agentic tasks requiring computer interaction, it's a top contender. 

Google LLMs for coding 

Gemini 3.1 Pro: Released in preview on February 19, 2026, Gemini 3.1 Pro is Google DeepMind's most capable model to date and a significant competitive leap. It's the first time Google has shipped a ".1" increment between major Gemini versions—a naming choice that reflects a focused intelligence upgrade rather than a broad feature expansion.

The standout headline: it more than doubled its reasoning performance over Gemini 3 Pro.

• ARC-AGI-2: 77.1%—more than double Gemini 3 Pro's 31.1%, and the highest score of any model at launch; tests novel pattern recognition that can't be trained away
• SWE-bench Verified: 80.6% on real-world software engineering tasks, competitive with the best models available
• GPQA Diamond: 94.3% on graduate-level science reasoning
• LiveCodeBench Pro: 2887 Elo—significantly ahead of GPT-5.2 (2393) and Gemini 3 Pro (2439)
• MCP Atlas: 69.2% on multi-tool coordination—leading all models tested; strong signal for production agentic pipelines
• Three-tier thinking: Low, Medium, and High compute modes—the new Medium level lets developers optimize cost per request without sacrificing quality
• 1M token context: Natively multimodal across text, audio, images, video, and entire code repositories; 64K output tokens
• Same pricing as Gemini 3 Pro: $2 per million input tokens, $12 per million output tokens—roughly 7.5x cheaper than Opus 4.6 on input
• Status: Currently in preview via Gemini API (AI Studio, Vertex AI, Gemini CLI, AntiGravity); rolling out to Google AI Pro and Ultra users in the Gemini app

The ARC-AGI-2 jump from 31.1% to 77.1% in a single generation isn't a benchmark footnote—it's the clearest signal yet that Gemini's reasoning architecture made a genuine qualitative leap. Developers building agents that need to handle novel, unpredictable problems are noticing. 

Other coding LLMs worth knowing 

The top three dominate most developers' daily workflows, but there's a rich ecosystem worth knowing—especially for cost-sensitive, open-source, or specialized use cases.

DeepSeek V3.2 / R1: The cost-to-performance leader. DeepSeek V3.2 remains the go-to for high-volume workloads where API economics matter, and R1 is the reasoning variant favored for math and algorithms. V4 is expected imminently—multimodal, 1M context, SWE-bench scores that could challenge the frontier.

• API pricing starts at $0.27/M input tokens, dropping to $0.028/M on cache hits
• R1 excels on AIME and competition math; V3.2 leads on general coding and tool use
• Open-weight; self-hostable with full data privacy
• V4 benchmarks unverified at time of writing—evaluate at launch

Kimi K2.5 (Moonshot AI): Strongest open-source coding model at launch, with real developer traction. Usage hit 50B+ tokens/day on OpenRouter within hours of release—developers integrating it into real workflows, not just experimenting.

• 76.8% SWE-Bench Verified; 1T parameter MoE at $0.60/M input (10x cheaper than Claude Opus 4.6)
• Standout visual coding: turns UI designs and screen recordings into working front-end interfaces
• Agent Swarm mode for parallel sub-agent execution on complex tasks
• Kimi Code CLI is a direct Claude Code competitor; open-source, works with VSCode, Cursor, Zed

Qwen 3.5 (Alibaba): The open-weight model that's forcing a conversation about whether closed models are worth the premium. Released Feb 16, 2026; 397B total / 17B active parameters, Apache 2.0, 201 languages.

• IFBench instruction following: 76.5—beats GPT-5.2 (75.4) and Claude (58.0)
• LiveCodeBench v6: 83.6; AIME 2026: 91.3
• $0.40/M input—10–17x cheaper than Claude or GPT for comparable tasks
• Self-hostable under Apache 2.0; smallest variants run on consumer hardware

GLM-4.7 (Z.ai):  The quiet workhorse for local and agentic coding. GLM-4.7 introduces thinking-before-tool-use behavior, improving reliability on long and complex agentic tasks—a pattern developers have found more consistent than larger models on multi-file work.

• Strong performance in real-world coding environments and popular coding agents
• Available in Google Cloud Vertex AI Model Garden for core coding, tool use, and complex reasoning
• Cost-effective; accessible via Z.ai, OpenRouter, and major agent frameworks
• Preferred local model for multi-file refactoring among developers on constrained hardware

Mistral Large 3/ Devstral 2: The European open-source option—strong general performance with a real deployment story for regulated industries. Mistral Large 3 is a sparse MoE with 41B active / 675B total parameters, Apache 2.0 licensed.

• ~92% pass@1 on HumanEval; second-ranked non-reasoning model on the LMArena open-source leaderboard
• Devstral 2 powers Mistral Vibe 2.0—a terminal-native coding agent with custom subagents, multi-file orchestration, and slash-command skills
• EU-based infrastructure; strong fit for teams with data residency requirements
• Runs on a single 8×GPU node; fine-tunable on internal codebases

Multi-model approach to coding with LLMs

Many developers use multiple models through platforms like Cursor or specialized IDE integrations like Codex CLI – leveraging different models' strengths for specific coding tasks rather than relying on a single LLM for every task.

• Code completion and FIM: Claude Sonnet 4.6 provides exceptional code completion with advanced pattern recognition, adapting precisely to existing code styles. GPT-5.4 is also frequently used for fast, high-quality completion.
• Architecture and design: Claude Sonnet 4.6 excels at system design, planning large refactors, and maintaining consistency across complex codebases.
• Algorithm development: DeepSeek R1 and GLM-4.7 are particularly strong for mathematically intensive algorithms and optimization problems.
• Security auditing: GPT-5.4 and Claude models demonstrate superior capabilities in identifying security vulnerabilities.
• Documentation generation: Gemini 3.1 Pro with its thought summarization features creates exceptionally clear and comprehensive documentation.
• Front-end development: GPT-5.4 and Claude Sonnet 4.6 both excel at creating beautiful, responsive UIs with aesthetic sensibility.
• Multi-language projects: Models with strong cross-language capabilities like GPT-5.4 and Gemini 3.1 Pro help with projects spanning multiple programming languages.
• Educational settings: More affordable options like GLM-4.7 or DeepSeek's distilled models provide excellent learning tools.
• Long-running agentic tasks: Claude Sonnet 4.6 and Opus 4.6's agent teams feature stand out for complex autonomous workflows; Gemini 3.1 Pro excels at iterative development requiring extended reasoning.

Aside from developer’s preferences, you can get an idea of how good an LLM is at coding with industry benchmarks. 

Coding LLM benchmarks

These are the leading benchmarks for evaluating LLMs in software development—from Python-specific tasks to real-world software engineering and agentic terminal work.

One important caveat in 2026: scores increasingly depend not just on the model but on the agentic scaffold used (Claude Code, Codex CLI, custom builds). Always note the evaluation conditions when comparing numbers across sources.

1. SWE-bench Verified:  The industry standard for agentic code evaluation, measuring model performance on real GitHub issues. Underwent a major scaffold upgrade in February 2026 (v2.0.0)—scores before and after aren't directly comparable.

2. SWE-Bench Pro: A harder, contamination-resistant variant sourced from complex real-world codebases. Even top frontier models score only ~23% here versus 70%+ on Verified Scale—a useful reality check on how much standard scores are inflated by training data leakage.

3. Terminal-Bench 2.0: 89 tasks in real terminal environments drawn from actual developer workflows, covering system architecture, dependency management, and environment configuration. 

4. LiveCodeBench: Continuously pulls new competition-level coding problems so models can't train their way to a high score. Best signal for algorithmic and competitive programming ability.

5. Aider Polygot: Tests models' ability to edit source files across multiple programming languages using diffs. Practical signal for real-world code editing workflows.

6. HumanEval: OpenAI's classic benchmark: 164 Python problems with unit tests. Frontier models now score 90%+ so it's losing discriminatory power at the top end, but remains widely cited.

7.  BigCodeBench: The functional successor to HumanEval. More complex function-level tasks across a wider range of real-world libraries and use cases.

8. WebDev Arena: Head-to-head human preference Elo for frontend and full-stack tasks. Best signal for UI quality and real-world web development. Gemini 3.1 Pro currently leads.

9. ARC-AGI-2: Tests novel pattern recognition that models can't memorize their way through. One of the most meaningful benchmarks for raw reasoning ability. Gemini 3.1 Pro leads at 77.1%.

10. Chatbot Arena:  Blind Elo voting from real humans across coding and general tasks. The most reliable proxy for real-world output quality.

These benchmarks help developers compare key metrics such as debugging accuracy, code completion capabilities, and performance on complex programming challenges.

Which LLM should my team use for coding?

Different models excel in different areas:

• Production development at scale: Claude Sonnet 4.6
• Architecture & system design: Claude Sonnet 4.6
• Complex multi-agent pipelines: Claude Opus 4.6
• Front-end & beautiful UIs: GPT-5.4
• Native computer use / agentic automation: GPT-5.4
• High-volume or cost-sensitive workloads: DeepSeek V3.2 or Qwen 3.5
• Multimodal & full-stack: Gemini 3.1 Pro
• Mathematical optimization & algorithms: DeepSeek R1
• Regulated industries / data residency requirements: Mistral Large 3 (self-hosted)
• Local deployment on constrained hardware: GLM-4.7 or Qwen 3.5

Most experienced developers take a multi-model approach leveraging specialized strengths for specific tasks through platforms like Claude Code, Cursor, and Copilot.

Need help choosing which coding LLMs to use at your company?

We can help you decide which coding LLMs are best for your teams, tech stack, and goals. Our senior engineering teams already use AI coding tools to build production-ready software.

And if you need to hire an AI-augmented team of senior software engineers to speed up delivery, we can assemble them for you in 4-6 weeks. It’ll be faster to get started, more cost-efficient than internal hiring, and we’ll deliver high-quality results quickly.

Zappos, Twilio, and Veho are just a few companies that trust us to build their software and enterprise systems.

You can schedule a time to talk with us here. No hassle, no expectations, just answers.