AI coding benchmarks, explained for vibe coders

Quick answer

WebDev Arena

People vote on front-end work, including tasks that take more than one step. The number is a preference score with a published interval. It is not a test that your app builds, logs in, or saves data. The page header says 18 October 2026, 900,155 votes, and 146 models. These are the top 8 rows.

Independentarena.ai WebDev boardBoard header 18 Oct 2026. Read 18 Oct 2026.

1. 1Claude Opus 5.5 (Max)2,870 votes1806 +14/-14
2. 2GPT-6 Astra (Max)6,172 votes1776 +10/-10
3. 3Claude Sonnet 5.5 (xHigh)3,342 votes1768 +13/-13
4. 4GPT-6.1 Sol (Max)3,021 votes1754 +13/-13
5. 5Claude Fable 5.1 (Max)7,684 votes1742 +9/-9
6. 6Claude Sonnet 5.5 (High)5,105 votes1709 +10/-10
7. 7Claude Opus 5 (Max)18,473 votes1690 +6/-6
8. 8GPT-6 Sol (Max)4,552 votes1682 +10/-10

Gemini 4 Argon (High) is rank 9 at 1678 +15/-15 and is marked Preliminary on that page. It is outside this top 8.

Terminal-Bench 4.0

Each row is an agent, a model, and an effort setting. The percent is how many of 330 terminal tasks that pair finished. The same model shows up more than once when the effort changes. These are the first eight rows on the public board, in that board's order.

Independenttbench.ai leaderboardBoard updated 6 Oct 2026. Read 18 Oct 2026.

1. 1Opus 5.5 · Claude Codemax effort · 214/330 · $4,718.65 · run 22 Sep 202664.8% ± 3.1%
2. 2Sonnet 5.5 · Claude Codemax effort · 204/330 · $7,333.64 · run 28 Sep 202661.8% ± 2.9%
3. 3GPT-6 Astra · Codexmax effort · 192/330 · $3,267.18 · run 3 Sep 202658.2% ± 2.8%
4. 3GPT-6.1 Sol · Codexmax effort · 192/330 · $634.24 · run 29 Sep 202658.2% ± 3.1%
5. 5Fable 5.1 · Claude Codemax effort · 191/330 · $6,243.50 · run 1 Sep 202657.9% ± 3.8%
6. 5GPT-6 Astra · Codexxhigh effort · 191/330 · $2,350.51 · run 3 Sep 202657.9% ± 2.7%
7. 5GPT-6 Astra · Codexhigh effort · 191/330 · $2,269.42 · run 3 Sep 202657.9% ± 3.0%
8. 5Fable 5.1 · Claude Codexhigh effort · 191/330 · $4,872.04 · run 1 Sep 202657.9% ± 3.4%

Other rows checked on the same board

The board has more rows than these. Ranks jump where a row was not copied onto this page. Grok 4.7 is a partial run: the cost field is 3683.287146 dollars, and the row says 324 of 330 trials.

Independenttbench.ai leaderboardBoard updated 6 Oct 2026. Read 18 Oct 2026.

1. 11Opus 5 · Claude Codexhigh effort · 178/330 · $6,086.22 · run 24 Jul 202653.9% ± 3.2%
2. 13Opus 5 · Claude Codemax effort · 171/330 · $5,969.11 · run 24 Jul 202651.8% ± 3.4%
3. 16GPT-6 Sol · Codexmax effort · 163/330 · $1,451.32 · run 22 Sep 202649.4% ± 3.2%
4. 21Grok 4.7 · Grok Buildxhigh effort · 124/330 · 3683.287146 USD, partial 324/330 · run 21 Sep 202637.6% ± 3.5%
5. 22GPT-5.6 Sol · Codexmax effort · 123/330 · $2,541.70 · run 26 Jun 202637.3% ± 3.8%
6. 31GPT-6 Luna · Codexmax effort · 54/330 · $72.14 · run 22 Sep 202616.4% ± 2.7%

Anthropic prints a different Opus 5.5 run: 66.4% with a footnote of ±2.6 points, at xhigh, on Anthropic's setup. That number is vendor-reported. It is not this public 64.8% row, and this page does not average them.

Vendor-reportedanthropic.com/claude-opus-5-5Read 18 Oct 2026.

What the vendor printed

These scores come from Anthropic's Opus 5.5 announcement. They are not the public boards above. A footnote says the GPT-6 Astra and GPT-5.6 Sol Terminal-Bench figures were reported by OpenAI. No separate OpenAI benchmark page was opened. CursorBench is Anthropic's eval name. It is not a score for the Cursor product. A cell the announcement left as a dash is left out of these lists.

Terminal-Bench 4.0, on Anthropic's page

Vendor-reportedanthropic.com/claude-opus-5-5Read 18 Oct 2026.

1. Opus 5.5xhigh. Footnote: ±2.6 pts. Anthropic's setup. Anthropic, self-reported.66.4%
2. Fable 5.1Same table. Footnote: ±1.6–2 pts for the other Claude models. Anthropic, self-reported.55.8%
3. Opus 5Anthropic says the public Claude Code board reports 51.8% and their setup reproduces 52.3%. Anthropic, self-reported.52.3%
4. GPT-6 AstraFootnote: high effort, as reported by OpenAI. OpenAI figure, quoted on Anthropic's page.57.9%
5. GPT-5.6 SolFootnote: as reported by OpenAI. Effort is not printed on this cell. OpenAI figure, quoted on Anthropic's page.37.3%

FrontierCode v1.1 (Main)

Vendor-reportedanthropic.com/claude-opus-5-5Read 18 Oct 2026. A later sentence on the same page says Opus 5.5 is 54.6% at default effort (medium). The list uses the table cell, 54.4%.

1. Opus 5.5Table cell. A later sentence on the same page says 54.6% at default effort (medium). Anthropic, self-reported.54.4%
2. Fable 5.1Table cell. Anthropic, self-reported.50.3%
3. Opus 5Table cell. Anthropic, self-reported.48.0%
4. GPT-6 AstraPrinted on Anthropic's table. This page did not open a separate OpenAI write-up. Printed by Anthropic.53.3%
5. GPT-5.6 SolPrinted on Anthropic's table. Printed by Anthropic.47.5%

CursorBench 4.0

Vendor-reportedanthropic.com/claude-opus-5-5Read 18 Oct 2026. A later sentence says Opus 5.5 is 52.5% at default effort (medium). The list uses the table cell, 57.8%.

1. Opus 5.5Table cell. The page's max-effort note covers Opus 5.5 unless a line says otherwise. A later sentence says 52.5% at default effort (medium). Anthropic, self-reported.57.8%
2. Fable 5.1Table cell. The later sentence calls this the max setting. Anthropic, self-reported.51.8%
3. Opus 5Table cell. The later sentence calls this the max setting. Anthropic, self-reported.46.6%
4. GPT-5.6 SolTable cell. The later sentence calls this Sol's top score. Printed by Anthropic.41.7%

What this means for you

Beginners

Start with a tool you can open and a price you can see on its tool page. WebDev Arena is the closest of these boards to “the screen looks like an app,” and it is still a vote, not your login flow. A model at the top of a terminal board can be a poor fit if the product that wraps it is closed, expensive, or assumes you already live in a repository. The quiz on Find your stack points at the tool pages. It does not use these scores.

Experienced vibe coders

Read the agent and the effort before you treat two Terminal-Bench percents as a model ranking. Claude Code plus Opus 5.5 at max is the public 64.8% row. Anthropic's own 66.4% is a different setup. Opus 5 at max is 51.8% ± 3.4%, and Opus 5 at xhigh is 53.9% ± 3.2% on the same public board, so the max row is not always the highest row. SWE-bench bash-only does not yet list the models on today's Claude Code and Codex pages. Aider does not list them either.

Caveats

• Contamination. SWE-bench issues and Exercism exercises are public. A model may have seen the task while training. A high score is not proof it can build a private app from a prompt it has never seen.
• Scaffolding. The SWE-bench full board mixes custom agents. Bash-only holds the model to mini-SWE-agent, and version 2.x is not the same setup as version 1.x. Terminal-Bench scores an agent, a model, and an effort. Aider scores a model plus an edit format. WebDev Arena is a blind preference, not a unit test.
• Vendor self-reporting. Anthropic's table is Anthropic's, except the Terminal-Bench cells the footnote attributes to OpenAI. A launch-post number can sit next to a different number on the public board for the same benchmark name. This page keeps both. It does not average them, and it does not estimate a score the page left blank.
• Intervals. WebDev Arena prints a plus and minus around each score. Terminal-Bench prints a band around the accuracy. Rank order inside that band is a weak reason to switch tools.

Coming soon: vibestacklab hands-on build tests

The numbers above are other people's measurements. vibestacklab has not run a hands-on test. The upcoming monthly test uses the same prompt in each tool: a habit tracker web app with login. Each run will report time to a working app, fixes needed, and cost. None of those fields are filled in yet.

Get the Tuesday email if you want the first log when it exists. The email already covers Radar tools, verified changes, and one practical tip. The build test is not in it until a run is logged.

Questions

Sources

• WebDev Arena on arena.ai — Independent vote board. Page header: 18 Oct 2026, 900,155 votes, 146 models. Read 18 Oct 2026.
• Terminal-Bench 4.0 — Independent public leaderboard of an agent plus a model. Leaderboard updated_at 6 Oct 2026. Row dates are the run dates below. Read 18 Oct 2026.
• Aider polyglot leaderboard — Independent of the model vendor. It is Aider's own harness.. Page footer: last updated 20 Nov 2025, by Paul Gauthier. Read 18 Oct 2026.
• SWE-bench leaderboards — Official board. Some rows are marked team-checked. Others are not.. Verified scores are in the homepage JSON. Newest Verified date in that file: 26 Feb 2026. Read 18 Oct 2026.
• SWE-bench Verified explainer — Benchmark definition. Method text only. This HTML did not include the numeric table. Read 18 Oct 2026.
• Claude Opus 5.5 announcement — Vendor self-report, except where the page says a figure was reported by OpenAI. Vendor table. Read 18 Oct 2026.