Only marked options receive the current limited offer. Extra options and capability calls are billed normally. Voting ends coverage; ongoing use is paid.

Which one should I use for a codebase I already have?

Both read and edit files. The difference people report is in how much direction each needs: Claude Code tends to explore before changing, Codex to act sooner. Send a task against your own repository and read both transcripts rather than taking either vendor's benchmark.

Do they run the same model?

No. Claude Code runs Claude models and Codex runs GPT models, so you are comparing a pairing rather than a harness in isolation. That is the pairing you would actually ship, which is why the page compares it that way.

Can I change the options?

Yes. Change the available models or add another compatible agent in the composer before sending. All options use normal usage billing unless explicitly marked as sponsored.

How does blind comparison work?

You know the candidate set, but results are labelled anonymously until you submit your preference. The original sessions remain available after reveal.

Community agent challenge

One task, different approaches. You choose the result.

Leaderboard

DeepSeek HarnessDeepSeek V4.1 FlashFreeCodexGPT-5.6 SolMediumFreeClaude CodeClaude Opus 5HighFree

Marked options are sponsored for up to 24 hours, ending when you vote. Added options and capability calls are billed normally; connector coverage is limited to 100 calls per account.

By starting a match you agree to the Terms. Task content and transcripts are never published.

How they have gone here

Published pairing: DeepSeek Harness vs Codex vs Claude Code. Human votes on this exact pairing, from blind matches readers judged — the evidence on this page describes that pairing, not whatever the composer above is currently set to.

The board counts a comparison only when it runs two different agents on the lineup's own models, so this combination has no record of its own. The published numbers below are measured independently.

See the full standings

Published reference numbers

What an independent lab measured for the models behind these options.

Intelligence v4.3

General capability across the lab's whole suite.

39.5

47.1

50.7

DeepSeek HarnessDeepSeek V4.1 Flash (Reasoning, Max Effort)

CodexGPT-5.6 Sol (max)

Claude CodeClaude Opus 5 (Adaptive Reasoning, Max Effort)

Coding

The coding subset of the same suite.

Not published

77.4

78.0

DeepSeek HarnessDeepSeek V4.1 Flash (Reasoning, Max Effort)

CodexGPT-5.6 Sol (max)

Claude CodeClaude Opus 5 (Adaptive Reasoning, Max Effort)

Agentic

Tasks the model runs in several steps.

Not published

50.5

56.2

DeepSeek HarnessDeepSeek V4.1 Flash (Reasoning, Max Effort)

CodexGPT-5.6 Sol (max)

Claude CodeClaude Opus 5 (Adaptive Reasoning, Max Effort)

These describe the model on its own, not the harness, the tools or the whole setup this page compares — and the lab runs each model at maximum effort, which is not what the composer above is set to. Scores are only comparable within one index version.

Source: Artificial Analysis — independent measurements, not ours, taken September 16, 2026.

View chart values

Measure	DeepSeek HarnessDeepSeek V4.1 Flash (Reasoning, Max Effort)	CodexGPT-5.6 Sol (max)	Claude CodeClaude Opus 5 (Adaptive Reasoning, Max Effort)
Intelligence index	39.5	47.1	50.7
Input$ / 1M tokens	0.30	4.00	5.00
Output$ / 1M tokens	1.20	20.00	25.00
Output speedtokens / s	214.4	64.7	49.4
Time to first tokens	1.21	94.96	43.51
Cost per task$	0.27	1.99	5.86

Source: Artificial Analysis — independent measurements, not ours, taken September 16, 2026. A dash is a measure they do not publish for that model.

Scored as a whole configuration

An independent benchmark that runs the harness and the model together, which is the unit this page compares.

Accuracy · Terminal-Bench 4.0

Percent of trials the whole configuration solved.

Not published

37.3%

51.8%

DeepSeek HarnessDeepSeek Harness · DeepSeek V4.1 Flash

CodexCodex · GPT-5.6 Sol123 of 330 trials · max effort

Claude CodeClaude Code · Claude Opus 5171 of 330 trials · max effort

The whisker is the 95% confidence interval. Two configurations whose whiskers overlap are not separated by this benchmark, however far apart their bars look.

Source: Terminal-Bench — an independent harness-and-model benchmark, not ours, taken September 17, 2026.

View chart values

Configuration	AccuracyTerminal-Bench 4.0
DeepSeek HarnessDeepSeek Harness · DeepSeek V4.1 Flash	No published run
CodexCodex · GPT-5.6 Sol	37.3% ± 3.8123 of 330 trials · max effort
Claude CodeClaude Code · Claude Opus 5	51.8% ± 3.4171 of 330 trials · max effort

± is the 95% confidence interval. Two configurations whose intervals overlap are not separated by this benchmark, however far apart their percentages look.

Source: Terminal-Bench — an independent harness-and-model benchmark, not ours, taken September 17, 2026. A dash is something they do not publish for that option.

Catalog price and context

What the open model catalog lists for the models behind these options.

Catalog	DeepSeek Harnessdeepseek/deepseek-v4.1-flash	Codexopenai/gpt-5.6-sol	Claude Codeanthropic/claude-opus-5
Input$ / 1M tokens	0.30	2.00	5.00
Output$ / 1M tokens	1.20	10.00	25.00
Context windowtokens	1.0M	1.1M	1M
Providers serving it	19	3	5

List price for the model alone. What a run from this page costs also depends on the harness, the tools it calls and how long it works.

Source: OpenRouter — the open model catalog, not ours, taken September 17, 2026. A dash is something they do not publish for that option.

Limited sponsorship

Coverage belongs to the marked options.

The fixed sponsored lineup is shown in the composer when the offer is available. You can add compatible options at normal usage rates. When sponsorship is unavailable, this page offers a normal paid comparison instead.

Coverage window
Eligible sponsored seats have a limited window of up to 24 hours. Voting ends their coverage. Continued usage is billed normally.
Exclusions
Added options and capability tool calls are not included. Connector coverage is limited to 100 waived calls per account; subsequent calls are billed normally.
Your results stay yours
Task content and transcripts are not published. Customized comparisons do not enter the fixed-lineup standings.

Common questions

Related resources

Activities and leaderboard

Find current challenges and standings.

All comparisons

Start with editable options.

Try the challenge with your task.

Compare your task

Community agent challenge

How they have gone here

Published reference numbers

Scored as a whole configuration

Catalog price and context

Coverage belongs to the marked options.

Coverage window

Exclusions

Your results stay yours

What is sponsored?

Which one should I use for a codebase I already have?

Do they run the same model?

Can I change the options?

How does blind comparison work?

Activities and leaderboard

All comparisons

Try the challenge with your task.