Section 2 of 7 — 18 models
Coding & Agentic ModelsCoding & Agentic Models
Models and agents built specifically to write, review, and ship code — from IDE copilots to fully autonomous cloud agents.
Ranked by score, highest first · 12 under evaluation follow
Anthropic
- +Strong at inferring intent from vague prompts and architectural context
- +1M-token context supports coherent multi-file, cross-repo refactors
- −Can use many more tokens than leaner coding models on routine work
- −Narrates its reasoning at length, which slows down quick tasks
Google
- +Resolves 75.8% of SWE-bench Verified issues on the standard open harness, the highest figure measured in this directory
- +Holds a million tokens of context at the cheaper Flash tier, so a large repository fits in one pass
- −Rated well below the Claude and GPT flagship tiers on the WebDev arena, where people judge working web apps
- −Its coding evidence is one harness run: nothing here measures how it behaves across a long agent session
Zhipu AI
- +Resolves 72.8% of SWE-bench Verified issues, the best open-weights result measured in this directory
- +Open weights served by several independent hosts, so a team can run it inside its own environment
- −Rated below the closed frontier tiers on the WebDev arena
- −A 200k-token context is limiting next to the million-token tiers when working across a large codebase
Google
- +Resolves 69.6% of SWE-bench Verified issues on the standard open harness
- +Stronger general reasoning than the Flash tier, which shows on problems that need architectural judgement
- −Scores below its own Flash tier on the same SWE-bench harness, so paying for Pro doesn’t buy better issue-fixing
- −Not served through OpenRouter, so its price and context window aren’t independently measured here
Anthropic
- +Resolves 66.6% of SWE-bench Verified issues, unusually high for a small, cheap tier
- +Anthropic’s cheapest tier with tool calling and structured outputs, which suits high-volume automated fixes
- −The lowest SWE-bench figure among the scored coding models here
- −A 200k-token context caps how much of a large codebase it can hold at once
Alibaba
- +Open-weight performance within striking distance of proprietary leaders
- +Free to self-host, appealing for cost-sensitive or air-gapped teams
- −Requires serious infrastructure to run at full size
- −Tooling and IDE integrations are less mature than Copilot, Cursor, or Codex
StepFun
- +600B mixture-of-experts with 27B active and a 1M-token context, built for long-horizon agentic work
- +$1.00 input and $2.70 output per million tokens, with open weights promised for October 2026
- −Generated roughly 1.7× the median output tokens on Artificial Analysis's index, so real cost runs well above the headline rate
- −A preview with no independent coding-benchmark result yet, so it carries no overall score here
OpenAI
- +Excellent terminal automation, git operations, and CI/CD debugging
- +Far more token-efficient than reasoning-heavy rivals on routine tasks
- −Needs detailed, unambiguous instructions; struggles with vague requests
- −Smaller context window than some rivals, a constraint on huge monorepos
Cursor
- +Tight, real-time feedback loop; you see and steer every change
- +Affordable flat-rate pricing for all-day assistance
- −Needs a developer actively driving; not built for unattended runs
- −Background and async agent mode is still early and limited
MiniMax
- +Low-cost alternative that holds up reasonably on coding benchmarks
- +Good option for parallel, high-volume exploratory tasks
- −Clearly behind the frontier coding leaders on hard problems
- −Smaller ecosystem and community support than bigger rivals
Cognition AI
- +Runs autonomously in the cloud, including overnight and unattended
- +Strong on large, well-defined jobs like migrations and refactors
- −Usage-based pricing gets expensive fast for heavy workloads
- −Needs clear upfront specs; weak at open-ended, exploratory work
Microsoft
- +Widest IDE support of any coding assistant, with the largest installed base
- +Accessible free tier and tight integration with pull-request review workflows
- −Less specialized for complex multi-file refactors than purpose-built AI IDEs
- −Suggestion quality depends heavily on which underlying model you select
Amazon
- +Strong for AWS-centric teams, including automated Java version upgrades
- +Deep integration with CloudFormation and infrastructure-as-code workflows
- −Limited value for teams outside the AWS ecosystem
- −Less compelling than Cursor or Copilot for general-purpose development
Codeium
- +Budget-friendly pricing with automatic indexing of large codebases
- +Flexible deployment options, including self-hosted setups
- −Smaller ecosystem and mindshare than Cursor or Copilot
- −Fewer third-party integrations than the bigger platforms
Tabnine
- +Built for privacy-first enterprises, with on-premises deployment and zero data retention
- +Supports bring-your-own-model for teams with strict compliance needs
- −More conservative suggestions than aggressive agentic competitors
- −Smaller feature set for autonomous, multi-step tasks
Replit
- +Goes from prompt to a running, deployed app inside one browser-based environment
- +Approachable for non-professional developers building small tools quickly
- −Less suited to large, established codebases than IDE-based agents
- −Deployment and hosting are tied to Replit's own platform
- +Free, open-source, terminal-based pair programmer that works with almost any model
- +Git-aware, auto-committing changes with clear diffs for easy review
- −No polished GUI; requires comfort with the command line
- −Lacks the managed infrastructure and support of commercial agents like Devin
OpenAI
- +Rated #1 of 128 models on LMArena’s WebDev arena, where people judge working web apps — 42 points clear of the next model
- +A 1,050,000-token context window, so a large codebase fits in one pass
- −No result yet on SWE-bench Verified’s standard harness, so its skill at fixing real repository issues isn’t measured here
- −Priced at the top of the market, at about $20 per million tokens blended