23 Pair Programmers Compared: Commercial Assistants, Open Models and Chatbots

The pair programmer category hides three different products: assistants that sit in your editor and read your repository, open models you download and run yourself, and chatbots you paste code into. They are priced differently, they fail differently, and several listings on this page describe plans that no longer exist. This guide separates them and flags where the ground has shifted.

1. Amazon Q Developer: AWS's Assistant, Priced Against Your Cloud Bill

Amazon Q Developer is the renamed CodeWhisperer plus agents for Java, Python and AWS infrastructure work. Its differentiator is that it knows IAM, CDK and the service limits of the platform you are deploying to.

Pricing: A free tier with limited agentic requests, then per-user Pro seats.

Where it falls short: Value collapses outside AWS. If your infrastructure is not on Amazon, you are paying for context you cannot use, and the free tier's request cap is reached quickly on real work.

2. GitLab Duo: Assistance Inside the Pipeline

GitLab Duo is unusual in that its AI lives in the merge request as much as in the editor: code suggestions, summarising threads, explaining vulnerabilities, generating tests inside CI.

Pricing: Per-seat add-on on top of a GitLab plan, with a limited free allowance on the lower tiers.

Where it falls short: It only makes sense if GitLab is already your platform. On GitHub it is irrelevant, and self-managed instances need the AI gateway configured before any of it works.

3. Sourcegraph Cody: Enterprise Only Now

Sourcegraph Cody was the codebase-aware assistant - it indexed your whole repository and answered questions with real cross-file context rather than whatever fit in a prompt.

Pricing: Enterprise only. Sourcegraph retired the free and individual Pro plans in July 2025 and now directs individual developers to its separate agentic tool, Amp.

Where it falls short: If you are an individual, the product described in most reviews no longer exists for you. The remaining offering requires an enterprise contract, which makes it a different decision entirely.

4. Codium AI: Renamed Qodo, Same Test-First Idea

Codium AI approached review from the test side: generate tests that expose edge cases, then judge the pull request against them. The company now trades as Qodo.

Pricing: A free tier for individuals, then per-user team plans.

Where it falls short: Test generation is only as good as the behaviour it infers, and inferred behaviour is often wrong in subtle ways. Treat generated tests as a first draft that a human must own, not as coverage.

5. Sourcery: Review Comments That Actually Explain Themselves

Sourcery focuses on refactoring suggestions and PR review, with explanations short enough that reviewers read them rather than skip them.

Pricing: A free tier for individual repositories, then paid team plans.

Where it falls short: It is a reviewer, not a writer. It will not implement the fix, and on large diffs the comment volume can exceed what a human reviewer will process.

6. GitButler: Version Control for People Who Work in Parallel

GitButler is not a code writer - it is a Git client that lets you stage different hunks onto separate virtual branches, so three unrelated changes stop contaminating one commit.

Pricing: Free while in its current stage, with paid plans announced for team features.

Where it falls short: It changes how you think about commits, which is a real cost. Teams with established Git workflows will find the virtual-branch model disruptive rather than liberating.

7. Mintlify: Documentation Generation, Not Code

Mintlify writes the docs most teams never write: it reads your code and produces reference pages, then hosts them as a documentation site.

Pricing: A free tier for open-source and small projects, then paid plans with custom domains and analytics.

Where it falls short: Generated reference docs describe what the code does, not why it does it. Architecture decisions and migration guides still need a human, and inaccurate docstrings propagate straight into the output.

8. Mutable AI: Spec-to-Code, Until Google Acquired It

Mutable AI pitched generating production code from high-level specifications and later built Auto Wiki, a tool that turned a codebase into a readable, cited wiki. Both claims were ambitious enough to put it in the same conversation as GPT Engineer.

Pricing: None applies. Mutable AI was acquired by Google in December 2024 and no longer sells a standalone product, so there is no current price to compare against the tools above.

Where it falls short: Spec-to-code systems produce plausible structure and unreliable detail, which was true here too. The bigger lesson is structural: an acqui-hire can retire a working tool overnight, and the wiki you built on it does not move with you. Treat any single-vendor documentation layer as a dependency with an expiry date, and keep your own architecture notes.

9. CodeGeeX: The Chinese Open Alternative

CodeGeeX is a multilingual code model from Tsinghua's KEG lab with its own IDE plugins, positioned as a free alternative for developers who cannot send code to overseas APIs.

Pricing: Free for individual use through the official plugins; enterprise licensing is separate.

Where it falls short: English-language ecosystem examples are thinner than the competition's, and plugin quality varies by IDE. It is a strong choice under data-residency constraints, a weaker one on pure completion quality.

10. StarCoder2: The Open Model With a Real Licence

StarCoder2 comes from BigCode, trained on permissively licensed code with a data-governance process that most open models lack. That provenance is the point.

Pricing: Free weights under a permissive licence; you pay for inference.

Where it falls short: The 15B model is capable but not frontier, and self-hosting it means owning the serving stack. For teams with licence-compliance requirements it is the safest open option; for raw quality it is not.

11. GLM-4: A General Model That Codes Well in Chinese

GLM-4 is Zhipu's general-purpose model, and its practical strength for many teams is bilingual: it handles Chinese code comments, commit messages and requirements without the translation loss Western models show.

Pricing: A free chat tier with API access sold per token.

Where it falls short: It is a chat model, not an editor-integrated assistant, so repository context is whatever you paste. On English-language framework idioms it trails the dedicated coding models.

12. Magicoder: Synthetic Data, Surprisingly Effective

Magicoder showed that training data quality matters more than quantity: its OSS-Instruct approach generated diverse coding problems from open-source snippets rather than scraping more repositories.

Pricing: Free and open source, research-oriented release.

Where it falls short: It is a research artefact - no hosted endpoint, no IDE plugin, no support. You run it yourself or you use something else.

13. WizardCoder: Evol-Instruct and a Broken Homepage

WizardCoder applied Evol-Instruct - automatically rewriting training prompts into harder variants - and produced a model that beat contemporaries on benchmarks at the time of release.

Pricing: Free and open source weights.

Where it falls short: The project's original site is gone; the code now lives in the WizardLM repository. It is also a snapshot of 2023-era capability - useful as a baseline, not as a daily driver.

14. PolyCoder: A 2022 Experiment Worth Reading, Not Running

PolyCoder was an early study of code model scaling from CMU, trained on 12 languages with a 2.7B parameter flagship. Its paper remains a useful reference.

Pricing: Free and open source, distributed as a research release.

Where it falls short: The public site is unreachable and the model predates instruction tuning becoming standard. Treat it as literature, not as a tool to install.

15. InCoder: Fill-in-the-Middle Before It Was Fashionable

InCoder from Meta made infilling a first-class training objective, which is why it could insert code at a cursor rather than only continue a prefix.

Pricing: Free weights on Hugging Face under a research licence.

Where it falls short: The original project page is gone and the weights are dated. Its infilling idea won; the model did not.

16. CodeT5: Encoder-Decoder for Code Understanding

CodeT5 takes a different architectural bet: an encoder-decoder built for code understanding tasks - summarisation, defect detection, translation between languages - not just generation.

Pricing: Free and open source from Salesforce Research.

Where it falls short: It is a fine-tuning base, not a product. The useful part is the checkpoint you adapt to your own task, which means ML work before any benefit.

17. CodeGen: The Pre-ChatGPT Generation

CodeGen is Salesforce's 2022 family of autoregressive code models, notable for scaling studies and for shipping sizes that would fit on a single GPU.

Pricing: Free and open source.

Where it falls short: It predates the instruction-following era, so it completes code rather than answering requests. Nothing here will replace an assistant that understands a task.

18. AlphaCode: Competitive Programming, Not Production

AlphaCode from DeepMind produced solutions ranked in the top 54% of participants in Codeforces contests - a genuine milestone, achieved by generating and filtering thousands of candidates per problem.

Pricing: Research publication; not a purchasable product.

Where it falls short: Contest problems are self-contained and precisely specified. Production code is neither. The sampling budget that makes AlphaCode work also makes it unusable interactively.

19. ChatGPT: The Default, for Better and Worse

ChatGPT is what most people mean when they say they use AI to code. It has no repository access unless you supply it, which is both the limitation and the reason it works everywhere.

Pricing: A free tier with usage limits, plus paid subscription tiers that raise limits and unlock newer models.

Where it falls short: Pasting context is lossy - you choose what the model sees, and you usually choose wrong. It also cannot run your tests, so confident wrong answers survive until you paste them into an editor.

20. DeepSeek Chat: Strong Reasoning at a Low Price

DeepSeek Chat built its reputation on delivering competitive reasoning at a fraction of frontier pricing, with an API that is genuinely cheap per token.

Pricing: Free chat access with rate limits; API billed per million tokens at rates well below Western frontier models.

Where it falls short: Availability and policy have changed repeatedly, and data-handling terms differ from US providers - a blocker for regulated codebases regardless of quality.

21. Qwen Chat: Alibaba's Model With a Free Tier

Qwen Chat gives access to the Qwen family, which has become one of the more reliable open-weight lineages for coding tasks and is widely available through third-party APIs.

Pricing: Free chat access; API usage sold per token, with open weights for several sizes.

Where it falls short: The chat UI is thin compared with a proper assistant - no repository indexing, no multi-file editing. Its real value is the weights and the API, not the interface.

22. Perplexity: Citations Instead of Code

Perplexity is a search-first assistant, and that is its coding use case: finding the current API signature or the thread where someone hit your exact error, with sources attached.

Pricing: A free tier with limited advanced searches, then paid subscription tiers.

Where it falls short: It answers questions, it does not write your codebase. Asking it to produce a nontrivial implementation gets you plausible text with citations that do not support it.

23. Phind: Built for Developers, Then Nerfed

Phind was engineered around technical queries - it searched documentation and Stack Overflow, then synthesised an answer with code. Its free tier was once the best deal in this category.

Pricing: A limited free allowance, then paid subscription tiers.

Where it falls short: The free allowance shrank as costs rose, and the product's direction has shifted toward paid usage. It remains good at technical search and weaker as a code author.

How to Choose

If you want an assistant in your editor today: Amazon Q inside AWS, GitLab Duo inside GitLab, Qodo for test-first review, Sourcery for review comments. If you need enterprise-only codebase context and have a contract: Cody.

If code cannot leave your network: StarCoder2 for licence-clean weights, CodeGeeX for a Chinese-hosted option, GLM-4 or Qwen for bilingual work. The 2022-era research models (PolyCoder, InCoder, CodeGen, CodeT5, AlphaCode) are references, not tools.

If you just want to ask questions: DeepSeek for price, Perplexity for citations, ChatGPT for breadth, Phind when the question is technical and you have allowance left. Skip MutableAI until its site responds.