A selection guide, not a ranking
The best AI model for coding is the one that merges cleanly in your repo
Public coding boards disagree because they test different jobs, in different agents, on different dates. This guide shows what the rows on those boards do and do not tell you, then gives you a five-task test and a cost formula so you can choose with your own data.
ZeroTwo publishes this page and sells model access, so we ran no benchmark of our own here. Every number below comes from a source named beside it, checked on 5 October 2026.
SWE-bench Verified, bash-only view
% resolved with mini-SWE-agent 2.0.0, and the average cost the board shows. Bars start at zero.
- Claude 4.5 Opus (high)76.8% · $0.75
- Gemini 3 Flash (high)75.8% · $0.36
- MiniMax M2.5 (high)75.8% · $0.07
- Claude 4.6 Opus75.6% · $0.55
- GPT 5.2 (high)72.8% · $0.47
- GPT 5.2 Codex72.8% · $0.45
Six of the ten rows are shown here. Source: swebench.com.
Three rules we use before trusting a coding score
These are our opinions, built from the rows further down. They will not tell you which model to buy. They tell you what to check before a number decides for you.
Match the benchmark to the work
Issue-fixing in a repository, short multi-language edits and terminal automation are different jobs with different boards. Read the one that looks like your week.
Treat the agent as part of the result
The same model scores differently inside different agents. A number from someone else's harness is a hint about your setup, not a promise.
Judge on cost per accepted change
A cheap run that you reject twice costs more than a dearer run you merge. Count what you merged, not what you sent.
Why one ranking cannot settle it
Three things we found when we read the boards themselves. Each one is checkable on the linked page.
Same model, different agent
Claude 4.5 Opus (medium) scores differently depending on the agent
- live-SWE-agent
- 79.20
- % resolved, run dated 2025-12-15
- mini-SWE-agent
- 74.40
- % resolved, run dated 2025-11-24
Both rows sit in the same SWE-bench Verified table on swebench.com, with a gap of 4.8 points. A score belongs to a model and an agent together, so a number from another setup is only a starting point.
Boards lag releases
The newest polyglot entry is from 2025
The newest run on the Aider polyglot table we read is dated 2025-10-03. Models released since then are absent, so a missing model is not a weak one.
Different jobs
500, 225 and one terminal
SWE-bench Verified is 500 human-filtered issue-resolution instances. The Aider polyglot set is 225 small exercises across six languages. Terminal-Bench scores an agent working in a terminal. A model can lead one and trail another.
Claims we removed
Winners we could not source
An earlier version of this page showed one benchmark column that mixed different SWE-bench variants, a quotation attributed to a researcher, and per-language winners. We could not source them to a row, so they are gone.
What each benchmark measures, and what it leaves out
Read the board that matches your work. The descriptions come from each board's own page; the right-hand column is where to read the current rows rather than a copy that goes stale.
| Benchmark | What it asks a model to do | What it cannot tell you | Where to read current rows |
|---|---|---|---|
| SWE-bench Verified | Resolve real GitHub issues. Verified is a human-filtered subset of 500 instances drawn from the original set of 12 Python repositories. | How the model does outside Python, on your repository layout, or inside a different agent. Rows in the same table can come from different agents. | swebench.com |
| SWE-bench Multilingual | The same kind of task across 300 instances from 42 repositories in 9 programming languages. | Which model is best at one language. The page we read lists overall resolution, not a winner per language. | swebench.com |
| Aider polyglot | Edit code to make 225 Exercism exercises pass, across C++, Go, Java, JavaScript, Python and Rust. | Multi-file work in a large codebase. The exercises are small and self-contained. The board also reports cost per run. | aider.chat leaderboards |
| SWE-bench Pro (Scale) | Longer-horizon software tasks on professional repositories, scored on a public set. | Terminal automation. Scale's page does not report a Terminal-Bench result. | labs.scale.com |
| Terminal-Bench | Agent work in a terminal. Its leaderboard names the agent as well as the model on each row, and showed benchmark version 4.0 when we read it. | Which model is best inside your editor. The agent setup is part of what is being scored. | tbench.ai |
Sources: swebench.com, aider.chat, labs.scale.com, tbench.ai. Checked 5 October 2026.
Score next to cost, from one board and one agent
Ten rows from the SWE-bench Verified bash-only table, all run with mini-SWE-agent 2.0.0, so the models sit in the same setup. Use it as a template for your own sheet: model, benchmark version, agent, source row, date and cost.
| Model and effort | % resolved | Avg. $ on the board | Run date |
|---|---|---|---|
| Claude 4.5 Opus (high) | 76.80 | $0.75 | 2026-02-17 |
| Gemini 3 Flash (high) | 75.80 | $0.36 | 2026-02-17 |
| MMiniMax M2.5 (high) | 75.80 | $0.07 | 2026-02-17 |
| Claude 4.6 Opus | 75.60 | $0.55 | 2026-02-17 |
| GPT 5.2 (high) | 72.80 | $0.47 | 2026-02-17 |
| GPT 5.2 Codex | 72.80 | $0.45 | 2026-02-19 |
| Claude 4.5 Sonnet (high) | 71.40 | $0.66 | 2026-02-17 |
| KKimi K2.5 (high) | 70.80 | $0.15 | 2026-02-17 |
| DeepSeek V3.2 (high) | 70.00 | $0.45 | 2026-02-17 |
| Gemini 3 Pro (high) | 69.60 | $0.96 | 2026-02-26 |
Source: swebench.com, read 5 October 2026. The board does not define its cost column on the page, so treat it as a relative figure. Models released after these runs are not in the table, and which of these you can use in ZeroTwo changes, so check the model library.
Start from your kind of work
A decision matrix that names jobs, not winners. Each row says which board to read, what to test yourself and the usual way the shortcut misleads.
| If your work looks like | Read this first | Then test | Watch for |
|---|---|---|---|
| Fixing bugs across several files | SWE-bench Verified and SWE-bench Pro rows for the agent you will use | Five closed issues from your own repository that have a failing test | Rows measured in a different agent, and models newer than the board |
| Terminal and CI automation | Terminal-Bench rows that name the same agent you run | A scripted task in a throwaway sandbox, with a time and token cap | Commands that touch files or services outside the sandbox |
| Short edits in many languages | Aider polyglot, which covers six languages | Two or three small tasks in each language you ship | Small exercises that flatter every model |
| Inline completion while you type | Latency and price in your editor, more than issue-resolution scores | A typing session on real files, counting accepted suggestions | Choosing a model for the board when the job is speed |
| Large batches of cheap, routine changes | The cost column next to the score, as in the table above | Cost per accepted change over 20 or more tasks | Savings that vanish when more runs need a rewrite |
| Code that cannot leave your network | The model's license and the hardware it needs | The same bake-off, run on your own infrastructure | Treating a hosted model's board score as a self-hosted score |
Once you have a shortlist, the next choice is where the model runs. Best AI for coding compares tools and editors, and AI for developers covers day-to-day workflows.
Compare cost per merged change, not price per token
Price per million tokens is only the first term. Add how many tokens a task burns and how often you accept the result.
The formula
(input tokens × input price + output tokens × output price) × runs ÷ changes you merge
Count every run, including the ones you throw away.
Worked arithmetic, illustrative
One run with 200,000 input and 20,000 output tokens
| Model | Input / output per 1M tokens | Cost of one run |
|---|---|---|
| Haiku 4.5 | $1 / $5 | $0.30 |
| Sonnet 4.6 | $3 / $15 | $0.90 |
| Opus 4.7 | $5 / $25 | $1.50 |
Hypothetical, to show the division: if the $0.30 run were accepted 4 times in 10 and the $1.50 run 9 times in 10, the cost per merged change would be $0.75 against about $1.67. Those acceptance rates are invented for the example and describe no real model.
Prices are Anthropic's published list API rates on claude.com/pricing, read 5 October 2026, with no caching or other modifiers. That page lists Opus 4.7 under legacy models and shows newer models at other rates.
A five-task test you can run this week
This is the part no leaderboard can do for you. It takes an afternoon and uses work you already trust.
Pick five tasks you already solved
Closed issues or pull requests with a test that failed before the fix. Keep a spread: one small, one cross-file, one in the language you worry about.
Freeze the setup
Same agent or editor, same tools, same context for every candidate. If you change two things at once you cannot tell which one mattered.
Run each candidate once per task
Write down pass or fail against the test, how many lines you had to fix by hand, tokens used and the wall-clock time.
Divide spend by merged changes
Total spend for the candidate divided by the number of changes you would actually merge. That is the figure to compare.
Repeat on a second set
A model that wins one set of five may not win the next. If the ranking flips, you have learned that the gap is small for your work.
Record sheet
One row per task and model
- Task
- Model
- Test passed
- Hand fixes
- Tokens in / out
- Minutes
A template for your own notes. It holds no results.
Two questions the public boards answer badly
Language
Is there a best model for Rust or Go?
The Aider polyglot score pools six languages into one number, and SWE-bench Multilingual pools nine. Neither page we read gives a winner per language. If you ship Rust, put Rust tasks in your five, and do the same for any other language that matters to you.
Self-hosting
Does an open-weight model fit?
Running a model yourself replaces per-token fees with hardware, operations and a quality check on your own code. Read the model's license, run the same five tasks on your own machines, and do not assume a hosted board score carries over to your setup.
Where ZeroTwo fits, and where it does not
We sell access to models, so weigh this section knowing that.
What it can do for the test
- Keep 60+ models from OpenAI, Anthropic, Google, xAI and others in one account, with the live list at /models.
- Free plan: 100 free credits to start, +20 bonus credits each day you log in, limited model selection, no card. Plus is $14.99/mo with a code agent. Pro is $29.99/mo with ZeroCode.
- Let you paste the same task into several models and compare the answers yourself.
What it does not do
- Run SWE-bench, Aider or Terminal-Bench for you. This page contains no ZeroTwo benchmark results.
- Promise that a model named on a board is available on your plan. Check the model library and your plan first.
- Replace testing on your own repository, which is the step that decides the choice.
Questions people ask before choosing a coding model
What is the best AI model for coding?
There is no single answer, and the published boards show why. SWE-bench, the Aider polyglot board and Terminal-Bench measure different jobs, and rows on one board can come from different agents and different dates. Shortlist two or three models from the board that looks like your work, then run the same five tasks from your own repository on each and compare cost per merged change.
Why do coding leaderboards rank models differently?
They test different tasks. SWE-bench Verified is 500 human-filtered issue-resolution instances, the Aider polyglot board is 225 small Exercism exercises in six languages, and Terminal-Bench scores agent work in a terminal. They also differ in harness, effort setting and run date, so a model can lead one board and sit lower on another.
Does the agent matter as much as the model?
It can matter a lot. On the SWE-bench Verified table we read, Claude 4.5 Opus at medium effort shows 79.20 with live-SWE-agent and 74.40 with mini-SWE-agent. The model is the same on both rows; the agent differs. When you compare a model, compare it in the agent you will use.
How much does Claude Opus 4.7 cost through the API?
Anthropic's pricing page listed Opus 4.7 at $5 per million input tokens and $25 per million output tokens, under legacy models, when we checked it on 5 October 2026. The same page lists newer models at other rates, and optional modifiers can change a bill, so check the page before you budget.
Which model is best for Python, Rust or Go?
The boards we read do not publish a winner per language. The Aider polyglot board covers C++, Go, Java, JavaScript, Python and Rust in one score, and SWE-bench Multilingual reports tasks across nine languages as one result. If one language matters most to you, put tasks in that language in your five-task test.
Is a cheaper model good enough for coding?
Sometimes. On the same SWE-bench Verified table, MiniMax M2.5 shows 75.80 at $0.07 and Claude 4.5 Opus (high) shows 76.80 at $0.75, both with mini-SWE-agent. That is a leaderboard comparison, not a promise about your code. Run your own tasks and compare cost per merged change before you switch.
Can I try several coding models in ZeroTwo without separate subscriptions?
ZeroTwo lists 60+ models from OpenAI, Anthropic, Google, xAI and others in one account, and the live list is at /models. The Free plan starts with 100 free credits, adds 20 bonus credits each day you log in, has limited model selection and needs no card. Plus is $14.99/mo and Pro is $29.99/mo. Check /models for which models a plan includes before you rely on one.
Which tool should I use once I have picked a model?
That depends on where you write code: an editor, a terminal agent or a chat window. The best AI for coding guide compares tools and environments, and the AI for developers page covers professional workflows.
Next steps after you pick a model
- Best AI for coding
Picked a model? This compares the tools and environments that can run it.
- AI for developers
Professional workflows such as review, debugging and documentation, with the model choice left open.
- AI code generator
From a snippet to a small app, and what to check in the output.
- Claude vs ChatGPT for coding
A two-way comparison for coding tasks, separate from this model guide.
- Model library
The current list of models you can use in ZeroTwo.
- ZeroTwo pricing
Plan prices and allowances, for when you run a bake-off at volume.