aiapicostcalcUnit economicsPricesBrandsCompareBenchmarksSkills & PluginsFind my modelGuidesChanges

Which LLM should I use? Find the best API for your task

Short answer: Tell the picker what you are building and it scores every verified model on quality, price, context and speed, then returns three models with the reasons and a monthly cost estimate. It uses fixed, published rules — no AI call, no account, nothing hidden.

Last verified 2026-11-18 · Sources: src/prices.json, src/benchmarks.json and src/tools.json with the rule set published on the page · Report an error

By Shangyi Duan · Rules, prices and benchmark dates are published on the page

The picker

Describe the job

Showing the coding-agent preset with balanced priorities, scored on the server with the same rules the page script uses. Type a use case or press a preset to rerun it in your browser.

#1 pick95.3 / 100OpenAI

GPT-5.6 Luna

Estimated $1.31 a month at 200 tasks, or $0.00654 per task at list price.

  • Rank 1 of 36 on SWE-bench Verified (69%)
  • $0.00654 per task — 5th cheapest of 42 current models
  • 1050K context window
  • 186 tokens/s output speed

Pairs with: Claude Code · OpenAI Codex CLI

#2 pick72 / 100OpenAI

GPT-5.6 Terra

Estimated $13.07 a month at 200 tasks, or $0.0654 per task at list price.

  • Rank 3 of 36 on SWE-bench Verified (67.8%)
  • $0.0654 per task — 26th cheapest of 42 current models
  • 1050K context window
  • 118 tokens/s output speed

Pairs with: Claude Code · OpenAI Codex CLI

#3 pick68.2 / 100OpenAI

GPT-6.1 Terra

Estimated $5.84 a month at 200 tasks, or $0.0292 per task at list price.

  • Rank 11 of 36 on SWE-bench Verified (65.6%)
  • $0.0292 per task — 17th cheapest of 42 current models
  • 1050K context window
  • 109 tokens/s output speed

Pairs with: Claude Code · OpenAI Codex CLI

The estimate uses the work profile of the selected preset, not your real traffic. Change the tasks-per-month field or use the calculator with your own token counts before you commit a budget.

What the score means

Four weighted factors

Each current model gets a 0-100 score from four ranked factors: quality on the preset's benchmarks, list price for the preset's token shape, context window and output speed. Quality comes from the ranking on this site's benchmark page; price from the verified list rates; context and speed from the provider documentation captured in our specs file. A model missing a benchmark gets a neutral quality factor and the card says so.

The rules, published in full

PresetQualityPriceContextSpeedBenchmarks usedAssumed volume / task size
Coding agent
Agentic coding: a model reads context, edits files and calls tools.
45%35%10%10%SWE-bench Verified, LiveCodeBench200 / 35K in / 2,500 out tokens
Support bot
Short customer replies where answer latency and price per conversation matter.
30%55%5%10%LMArena Elo20,000 / 2K in / 500 out tokens
RAG answers
Retrieved context is pasted into the prompt on every call, so input tokens dominate.
30%45%15%10%LMArena Elo5,000 / 20K in / 1K out tokens
Batch summarization
Asynchronous bulk work where latency does not matter and volume is high.
25%65%5%5%LMArena Elo5,000 / 10K in / 1K out tokens
Long documents
A very large prompt: contracts, transcripts, books, whole repositories.
25%25%45%5%LMArena Elo300 / 200K in / 2K out tokens
Lowest cost
Pick the lowest list price first; quality only as a tie-breaker.
5%90%5%0%LMArena Elo5,000 / 20K in / 1K out tokens
Highest quality
Pick the strongest benchmark position and treat price as secondary.
85%5%5%5%LMArena Elo2,000 / 20K in / 1K out tokens
Lowest latency
Time to first token and output speed matter more than a few points of quality.
20%20%0%60%Output speed5,000 / 20K in / 1K out tokens

Weights are the share of the total score; price weight rises whenever quality is hard to measure, and the quality-first preset flips the emphasis. Priorities on the form multiply one weight by 1.8 before normalising, which is the entire "tuning" of the picker.

Why it is not an AI oracle

Fixed arithmetic, no model call

Nothing is sent anywhere and no language model is called: the page runs the same ranking function that produced the three cards above. You can re-derive every number: prices are the list rates on the model pages, benchmarks are the dated scores on the benchmark page, and the work profiles are printed in the table.

What it does not know

It does not know your prompts, your evaluation results, your region, your negotiated rate or your latency requirements beyond the published speed numbers. Treat the top three as a shortlist: run 20-50 real tasks on each model, then use the comparison page to check the price of the winner at your own token counts.

Related

Questions

Is this recommendation made by an AI?

No. It is a fixed scoring rule printed in the rules table: four weighted factors and a preset work profile. No language model is called, nothing is sent to a server and the same arithmetic runs in your browser.

What does the score out of 100 mean?

It is a relative position, not a quality rating. Each factor is turned into a percentile among the current models, weighted by the preset, and scaled to 100. A score of 80 means the model scores well on those weights, not that it is 80% correct.

How is the monthly cost estimated?

List price × the preset's token shape × your tasks-per-month number. For coding the token shape comes from our tool data (a typical agent task); for other presets it is a published planning assumption in the rules table. Replace it with your logged counts for a real budget.

Why do only three models appear?

Because the page is a shortlist, not a catalogue. The other models are ranked too; open the comparison page to check specific runners-up or the full price table to see every model.

Can I share or reproduce a recommendation?

Yes. The selected preset, priority and task count are written into the page URL, and every factor in the score is printed in the rules table, so the same inputs reproduce the same top three for anyone who opens the link.

Next step → Compare your shortlist side by side

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.