ExpertPal tools · LLM Picker

Pick the model the work actually needs.

Don't burn frontier-model credits on work that doesn't need them.

What are you doing?

Priority

Best possible resultWhen failure costs more than tokens.Save money without getting stupidClose to the best, for noticeably less.Cheapest possibleStill strong enough to trust.

Best value

Anthropic

Claude Fable 5.1

Strongest for serious coding

Our pick when the result matters most.

Cheaper models are below our bar for serious coding.

Claude Opus 5.5 could be just as strong for 58% less. Benchmarks still pending.

Top picks for serious coding

Input / output per 1M · Example costExample cost

Claude Opus 5.5Estimated · benchmarks pending$4.00 / $20.00$1.08−58%

Why this model?

There is a cheaper model, but we don't consider it strong enough for this kind of work.

Prices are current standard API list prices from the provider.

02 / Compare

All models we trust, for serious coding.

Based on public benchmarks, weighted by us – not a benchmark of our own.

All models we trust, for serious coding.
Model	Fit	Input	Cached	Output	Example cost	Context
Claude Fable 5.1Best valueAnthropic	Excellent	$10.00	$0.25	$50.00	$2.60	1M
GPT-6 AstraOpenAI	Excellent	$10.00	$1.00	$50.00	$2.90	1.1M
GPT-6 LunaOpenAI	Good enoughEstimated · benchmarks pending	$0.10	$0.01	$0.50	$0.029	1.1M
Gemini 3.8 FlashCheapestGoogle	Good enough	$0.75	$0.075	$3.75	$0.218	1M
Claude Sonnet 5Anthropic	Good enough	$2.00	$0.20	$10.00	$0.58	1M
GPT-6 SolOpenAI	StrongEstimated · benchmarks pending	$2.00	$0.20	$10.00	$0.58	1.1M
Claude Opus 5.5Anthropic	ExcellentEstimated · benchmarks pending	$4.00	$0.20	$20.00	$1.08	1M
Claude Opus 4.8Anthropic	Good enough	$5.00	$0.50	$25.00	$1.45	1M
Gemini 3.1 Flash LiteGoogle	Not enough data yet	$0.25	$0.025	$1.50	$0.08	1M
Gemini 3.5 Flash LiteGoogle	Not recommended	$0.30	$0.03	$2.50	$0.117	1M
Claude Haiku 4.5Anthropic	Not recommended	$1.00	$0.10	$5.00	$0.29	200K
Claude Sonnet 4.6Anthropic	Not recommended	$3.00	$0.30	$15.00	$0.87	1M
Claude Opus 4.6Anthropic	Not enough data yet	$5.00	$0.50	$25.00	$1.45	1M

Claude Fable 5.1
Anthropic
ExcellentBest value
Example cost
$2.60
Input
$10.00
Cached
$0.25
Output
$50.00
Context
1M
GPT-6 Astra
OpenAI
Excellent
Example cost
$2.90
Input
$10.00
Cached
$1.00
Output
$50.00
Context
1.1M
GPT-6 Luna
OpenAI
Good enoughEstimated · benchmarks pending
Example cost
$0.029
Input
$0.10
Cached
$0.01
Output
$0.50
Context
1.1M
Gemini 3.8 Flash
Google
Good enoughCheapest
Example cost
$0.218
Input
$0.75
Cached
$0.075
Output
$3.75
Context
1M
Claude Sonnet 5
Anthropic
Good enough
Example cost
$0.58
Input
$2.00
Cached
$0.20
Output
$10.00
Context
1M
GPT-6 Sol
OpenAI
StrongEstimated · benchmarks pending
Example cost
$0.58
Input
$2.00
Cached
$0.20
Output
$10.00
Context
1.1M
Claude Opus 5.5
Anthropic
ExcellentEstimated · benchmarks pending
Example cost
$1.08
Input
$4.00
Cached
$0.20
Output
$20.00
Context
1M
Claude Opus 4.8
Anthropic
Good enough
Example cost
$1.45
Input
$5.00
Cached
$0.50
Output
$25.00
Context
1M
Gemini 3.1 Flash Lite
Google
Not enough data yet
Example cost
$0.08
Input
$0.25
Cached
$0.025
Output
$1.50
Context
1M
Gemini 3.5 Flash Lite
Google
Not recommended
Example cost
$0.117
Input
$0.30
Cached
$0.03
Output
$2.50
Context
1M
Claude Haiku 4.5
Anthropic
Not recommended
Example cost
$0.29
Input
$1.00
Cached
$0.10
Output
$5.00
Context
200K
Claude Sonnet 4.6
Anthropic
Not recommended
Example cost
$0.87
Input
$3.00
Cached
$0.30
Output
$15.00
Context
1M
Claude Opus 4.6
Anthropic
Not enough data yet
Example cost
$1.45
Input
$5.00
Cached
$0.50
Output
$25.00
Context
1M

Token prices in USD per 1M tokens.

03 / Background

Why this exists

New models appear constantly. Prices change. Caching rules change.

The cheapest model isn't really cheap if you need five attempts to get the job done. And the strongest model isn't automatically worth paying for either.

We keep a small list of models we trust, track their current prices, and help you pick the one that makes sense for the work.

04 / Method

How we decide

We combine current API prices with our assessment of which models are strong enough for different kinds of work. That assessment is calculated from public benchmark data, weighted by what matters for each task. Cost comparisons use the same representative workload across models. These ratings are practical judgments, not scientific benchmarks. No AI model chooses the recommendation.

Example workloads used for the cost comparison
Task	Input tokens	Cached input tokens	Output tokens
Serious coding	100,000	400,000	30,000
Routine coding	40,000	100,000	10,000
Complex reasoning	40,000	10,000	20,000
Research	100,000	50,000	20,000
Writing / everyday work	20,000	5,000	8,000
Simple / bulk work	100,000	0	10,000

Source: Artificial Analysis (artificialanalysis.ai) via OpenRouter (openrouter.ai/rankings), as of Sep 23, 2026. Licensed under CC BY 4.0.

Source: OpenRouter evals (openrouter.ai) via OpenRouter (openrouter.ai/rankings), as of Sep 24, 2026. Licensed under CC BY 4.0.