aiapicostcalcUnit economicsPricesBrandsCompareBenchmarksSkills & PluginsFind my modelGuidesChanges

Best LLM for coding in 2026: benchmarks and cost

Short answer: GPT-5.6 Luna leads our coding leaderboard (mean rank 1.0 across 2 coding metrics), and one typical coding task costs $0.00654 on it. The best value is GPT-5.6 Luna at $0.00654 per task, while the cheapest verified model overall is Gemini 2.5 Flash-Lite at $0.00277. Scores below are dated and ranked per metric.

Last verified 2026-11-18 · Sources: src/benchmarks.json (coding leaderboards) and src/tools.json (task token counts) with src/prices.json · Report an error

By Shangyi Duan · Coding scores and read dates are listed per metric below

Coding leaderboard

Ranked by mean place across the 2 coding benchmarks each model reports. A model that reports only one benchmark is listed but not eligible for the overall pick, because a single benchmark is a thin basis for a recommendation.

The leaderboard

#ModelSWE-bench VerifiedLiveCodeBenchMean rankQuality indexPer coding task
1GPT-5.6 Luna69%72.6%1.0100%$0.00654
2GPT-5.6 Terra67.8%71.3%2.094%$0.0654
3GPT-5.467.8%71.3%2.089%$0.0817
4Claude Sonnet 4.667.8%71.3%2.083%$0.0905
5GPT-5.567.8%71.3%2.091%$0.1634
6Claude Fable 5.167.8%71.3%2.086%$0.2873
7GPT-6 Astra67.8%71.3%2.097%$0.3018
8Gemini 2.5 Pro67.5%70.9%8.080%$0.0471
9Claude Opus 566.3%69.7%9.077%$0.1509
10Gemini 3.8 Flash65.6%69%10.049%$0.0226
11GPT-5.4 mini65.6%69%10.066%$0.0245
12GPT-6.1 Terra65.6%69%10.071%$0.0292
13Claude Haiku 4.565.6%69%10.054%$0.0302
14o4-mini65.6%69%10.060%$0.0336
15GPT-6.1 Sol65.6%69%10.074%$0.0584
16Claude Sonnet 5.565.6%69%10.057%$0.0584
17Gemini 4 Argon65.6%69%10.051%$0.0584
18GPT-6 Sol65.6%69%10.069%$0.0604
19GPT-4.165.6%69%10.063%$0.0611
20GPT-5.6 Sol65.4%68.8%20.046%$0.1207

Mean rank is the average of the model's place on each coding metric it reports; lower is better. Where the source publishes the same score for several models, those models share a rank and the cheaper task cost lists first, so a rank of 1.0 is a tie at the source, not a measured lead. Quality index maps those places to 0-100 within the models that report the same metric, so it is a relative position, not a score you can compare with a benchmark percentage.

What one coding task costs

How the per-task cost is computed

One coding task is priced as 35K input tokens (55% of them cached) and 2,500 output tokens, using the standard list rate of each model. The token basis is the median "typical task" token counts across the 14 tools in src/tools.json (35K input / 2,500 output, 55% cached). The formula is the same as the token cost calculator: (uncached input × input rate + cached input × cached rate + output × output rate) ÷ 1,000,000.

ModelPer taskPer 1,000 tasksMonthly at 200 tasksSWE-bench VerifiedLiveCodeBench
Gemini 2.5 Flash-Lite$0.00277$2.77$0.553560.1%63.1%
GPT-6 Luna$0.00302$3.02$0.603560.1%63.1%
Claude Haiku 5.5$0.00302$3.02$0.603560.1%63.1%
GPT-4o mini$0.00531$5.31$1.0645.2%47.6%
GPT-5.6 Luna$0.00654$6.54$1.3169%72.6%
GPT-5.4 nano$0.00666$6.66$1.3363%66.3%
Mistral Small 4$0.00675$6.75$1.3560.1%63.1%
Gemini 3.1 Flash-Lite$0.00817$8.17$1.6363%66.3%
GPT-5 mini$0.00942$9.42$1.8856.2%66.3%
Gemini 3.5 Flash-Lite$0.0116$11.55$2.3157.7%60.7%
Gemini 2.5 Flash$0.0116$11.55$2.3163%66.3%
Codestral 25.08$0.0127$12.75$2.5560.1%63.1%
Mistral Large 4$0.0176$17.62$3.5263%66.3%
Gemini 3.8 Flash$0.0226$22.63$4.5365.6%69%
Gemini 3.6 Flash$0.0226$22.63$4.5360.7%63.8%
GPT-5.4 mini$0.0245$24.51$4.9065.6%69%
GPT-6.1 Terra$0.0292$29.21$5.8465.6%69%
Claude Haiku 4.5$0.0302$30.18$6.0465.6%69%
DeepSeek V4 Pro$0.0315$31.54$6.3163.2%66.4%
o4-mini$0.0336$33.62$6.7265.6%69%
Gemini 2.5 Pro$0.0471$47.09$9.4267.5%70.9%
GPT-6.1 Sol$0.0584$58.42$11.6865.6%69%
Claude Sonnet 5.5$0.0584$58.42$11.6865.6%69%
Gemini 4 Argon$0.0584$58.42$11.6865.6%69%
GPT-6 Sol$0.0604$60.35$12.0765.6%69%
Claude Sonnet 5$0.0604$60.35$12.0764.9%68.2%
GPT-4.1$0.0611$61.12$12.2265.6%69%
GPT-5.6 Terra$0.0654$65.35$13.0767.8%71.3%
GPT-5.4$0.0817$81.69$16.3467.8%71.3%
Claude Sonnet 4.6$0.0905$90.52$18.1167.8%71.3%
Claude Opus 5.5$0.1168$117$23.3763.9%67.2%
GPT-5.6 Sol$0.1207$121$24.1465.4%68.8%
Claude Opus 5$0.1509$151$30.1866.3%69.7%
GPT-5.5$0.1634$163$32.6767.8%71.3%
Claude Fable 5.1$0.2873$287$57.4667.8%71.3%
GPT-6 Astra$0.3018$302$60.3567.8%71.3%

Sorted cheapest first. Monthly cost uses 200 tasks a month and the standard rate, with no batch discount; asynchronous batch work is 50% off on most of these models, which changes the order only at the high end.

Pick by monthly budget

Cap = monthly budget ÷ 200 tasks. Inside each cap we take the highest-ranked model whose monthly cost fits; if nothing fits, the row says so instead of recommending a cheaper-cap tier that does not exist.

BudgetMax cost / taskMonthly tokens costBest model inside the capMonthly costMean rankMetrics reported
$25.00 per month$0.125024.14GPT-5.6 Sol$24.14#20.0 of 362 of 2
$100 per month$0.500060.35GPT-6 Astra$60.35#2.0 of 362 of 2
$400 per month$2.0060.35GPT-6 Astra$60.35#2.0 of 362 of 2
$1,000 per month$5.0060.35GPT-6 Astra$60.35#2.0 of 362 of 2

What the price buys

Read the list from the top down: if the highest-ranked model is more than about twice the cost of the best value, the value pick is usually the better first model, because a public leaderboard lead of a few places is smaller than the difference your own eval will find.

Method and sources

Coding benchmark sources

MetricUnitModels reportingSourceData checked
SWE-bench Verified%36leaderboard2026-11-04
LiveCodeBench%36leaderboard2026-11-04

Missing scores are not filled in

If a model has no published score for a benchmark it is left out of that metric's ranking and its mean rank uses only the metrics it does report. We never copy a score from a sibling model, interpolate a gap or substitute a different benchmark version.

Scores shared across models

Some metrics repeat the same rounded value across models: AIME: 4 repeated values covering 26 of 36 rows; LMArena Elo: 2 repeated values covering 5 of 33 rows; GPQA Diamond: 1 repeated value covering 2 of 15 rows; LiveCodeBench: 4 repeated values covering 26 of 36 rows; MMMU-Pro: 1 repeated value covering 2 of 20 rows; SWE-bench Verified: 4 repeated values covering 25 of 36 rows. An equal value means the source published the same reported figure, not that the models are equally capable on that metric, so we do not use these ties for fine-grained ranking between the models that share one.

For SWE-bench Verified, LiveCodeBench and AIME the values come from TrustTheBench, an aggregator that republishes results under a standardized protocol (mini-SWE-agent 2.0.0, snapshot 2026-09-12). Its published snapshot is tier-clustered: leading results are reported in the same band (its own page lists several models at 75.8%), so identical scores above indicate the same reporting band in that snapshot. We keep each model's own row from the source, but treat a shared value as a tie at the source rather than as evidence of equal capability.

Repeated LMArena Elo values come from the source's rounded ratings: the same Elo in the table is a tie in the published snapshot, not a separate measurement.

A mean rank built on tier-clustered scores inherits the source's reporting bands: models that share a value have the same rank, so do not read the order between them as a measured gap. Re-run your own eval before choosing between two models that sit in the same band.

Limits of a public leaderboard

A leaderboard score is a fixed set of tasks, not your repository. The difference between two models a few places apart is often smaller than the run-to-run variance on your own board, so use this page to shortlist two or three models, then run 20-50 of your real tickets on each before you commit.

Related coding tools, skills & plugins

The model is half of a coding bill; the agent, skill or MCP server you run on it is the other half. These entries price one typical task on the same reference models.

Related

Questions

What is the best LLM for coding in 2026?

GPT-5.6 Luna has the best mean rank on the coding benchmarks we track, and one coding task costs $0.00654 on it at list price. The best value at the same quality is GPT-5.6 Luna ($0.00654 per task). The right pick still depends on your repo, so run a small eval before switching.

How much does one coding task cost?

We price one task as 35K input tokens (55% cached) and 2,500 output tokens at each model's standard rate. The spread across the models that report coding scores runs from $0.00277 to $0.3018 per task, which at 200 tasks a month is $0.5535 to $60.35.

Which coding model is cheapest?

Gemini 2.5 Flash-Lite has the lowest cost per coding task at $0.00277. It is not automatically the right choice: check its score columns and how far its mean rank falls behind the leader before trading quality for price.

Do these costs include caching, batch or subscriptions?

The task price applies the list rate with the cached share shown above and no batch discount. If your agent repeats a large prefix, prompt caching can cut the input side to the cached rate; asynchronous batch runs are 50% off on most models here. The coding tool subscription itself is separate and is covered on the tool cost pages.

Where do the benchmark numbers come from?

Each coding metric names its leaderboard source and read date in the source table on this page. Scores are copied from the leaderboard itself, not from a vendor blog, and a model without a readable score is left out of that ranking rather than estimated.

How many tasks a month does the budget table assume?

The table uses 200 tasks per month (median of the tasks-per-month fields in src/tools.json). Change your own number in the arithmetic: monthly cost = cost per task × tasks per month. A single heavy session can be several tasks, so count sessions and multiply by the average number of tasks a session completes.

Next step → See what your coding tool burns per task

Price data last checked 2026-11-18 against official sources · how we verify

Price-change alerts

One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.