Lesson 14 of 22Go deeperintermediate4 min

AI Model Comparison

Cost, speed, and capability: a practical buying guide

Not all models are equal

Models differ in cost, context length, speed, and strengths. Picking the right tier for the job saves money and often improves results. Start with the need picker for a teaching default, then open the cost playground: figures come from the same estimateCost helper the tests exercise, not a live billing API.

Educational pricing only

As of 2026-07-29: Figures are standard public list-price snapshots. The calculator applies documented long-context tiers per request, but excludes caching, cache-write premiums, batch/flex/priority discounts, tool fees, regional uplifts, and taxes. Verify before production spend. Self-hosted open weights show $0 API cost, but GPUs, power, and people still cost money. Recommendations are teaching defaults, not live leaderboard ranks. OpenAI models Anthropic models Google Gemini 3 xAI model migration

Route by job, not brand

Match tier to the work: mini for high-volume simple tasks, frontier for hard multi-step work, open weights when data residency matters.

Cost is tokens times rate

API bills scale with input and output tokens. A cheaper model that needs more retries can still lose. Measure on your evals.

Context is capacity, not free quality

A large window can hold a book. Quality often still drops across very long contexts. RAG plus citations often beats stuffing everything.

Self-host hides cost elsewhere

Open weights can mean $0 per token on the invoice. Hardware, energy, and people are real. Account for them before declaring a win.

As of 2026-07-296 models · sample call 10k in / 2k out: flagship $0.1100 vs mini $0.0220 · flagship is ~ 5× mini on this call

What do you need?

Choose a job shape. The lab returns ordered catalog picks from pure data ( recommendForNeed), not a live eval harness.

Classification, extraction, short rewrites where wrong answers are cheap to catch.

Why these: Token price and latency dominate. Mini tiers and self-hosted small models win when quality is "good enough".

Recommendation is a teaching default. Measure your own eval set before locking a tier.

Teaching picks (best first)

GPT-5.6 LunaOpenAI · sample $0.0220
Open-weight modelSelf-hosted ecosystem

Usually overkill or underpowered here: GPT-5.6 Sol, Claude Opus 5

Workload preset

Scenario: Quick Q&ASimple factual question

Input tokens50

Output tokens100

Requests / month (volume)10,000

At this volume, cheapest API tier (Grok 4.3) ≈ $3.13/mo for this request shape. Self-hosted still shows $0 API bill (you pay hardware).

Grok 4.3xAILowest API $

Per request

< $0.01

~$3.13/mo

Standard · under 200K input

Context

1M tokens

Input

$1.25/1M

Output

$2.50/1M

Speed

Fast

In $0.00006 · Out $0.00025 · < $0.01

Large context with low published token ratesCompetitive coding and general chat qualityProduct integration with X

GPT-5.6 LunaOpenAI

Per request

< $0.01

~$6.50/mo

Standard · up to 272K input

Context

1.05M tokens

Input

$1.00/1M

Output

$6.00/1M

Speed

Very Fast

In $0.00005 · Out $0.00060 · < $0.01

Lower-cost OpenAI tier ($1/$6 per 1M)Fast for classification, extraction, short Q&AGood enough for many production high-volume paths

Gemini 3.1 Pro PreviewGoogle

Per request

< $0.01

~$13.00/mo

Standard · up to 200K input

Context

1M tokens

Input

$2.00/1M

Output

$12.00/1M

Speed

Moderate

In $0.00010 · Out $0.00120 · < $0.01

1M-token context for books and codebasesStrong multimodal and Google ecosystem tiesCompetitive standard-tier pricing

Claude Opus 5Anthropic

Per request

< $0.01

~$27.50/mo

Standard

Context

1M tokens

Input

$5.00/1M

Output

$25.00/1M

Speed

Moderate

In $0.00025 · Out $0.00250 · < $0.01

Frontier reasoning, coding, and agent workflowsLarge 1M-token context windowStrong instruction following and safety posture

GPT-5.6 SolOpenAI

Per request

< $0.01

~$32.50/mo

Standard · up to 272K input

Context

1.05M tokens

Input

$5.00/1M

Output

$30.00/1M

Speed

Moderate

In $0.00025 · Out $0.00300 · < $0.01

Strong general reasoning and code generationLarge context for codebases and long documentsReliable tool calling and structured output

Open-weight modelSelf-hosted ecosystemSelf-hosted

Per request

$0.00 (self-hosted)

~$0/mo

Context

Varies by model

Input

Free (self-hosted)

Output

Free (self-hosted)

Speed

Depends on hardware

Open weights: full control and no per-token API billData can stay on your infrastructureFine-tunable for domain-specific tasks

Choosing a model

Simple tasks (classification, extraction, short answers): mini tiers win on latency and cost.

Hard reasoning (code, analysis, multi-step): frontier or reasoning tiers; wrong answers cost more than tokens.

Privacy / residency: open weights and self-host; the API bill is zero but ops is not.

Huge documents: large-context flagships (e.g. Gemini-class 1M). Still consider RAG for citations and cost.

Costs use estimateCost() on educational mid-2026 price bands. Not a live quote. Verify before production spend.

Routing advice

High-volume classification or extraction
Latency and cost dominate; quality is usually good enough.
Mini / small tiers
Long document or whole-codebase analysis
You pay for window size; still consider RAG for citations.
Large-context flagships (e.g. Gemini-class 1M)
Hard multi-step coding or math
Accuracy beats token price when wrong answers are expensive.
Frontier + optional reasoning tier
Private or offline workloads
Data residency and unit economics; ops cost is real.
Open-weight / self-hosted

Key Insight

Bigger is not always better. For classification and short answers, mini tiers are faster and far cheaper (about 5× less than flagship on the sample call above). Reserve frontier and reasoning tiers for hard multi-step work where a wrong answer costs real money. The most expensive token is the one you did not need to generate.

Check your understanding

1/4

When should you use a smaller, lower-cost model instead of a frontier model?

Sources, scope & review statusVerified Jul 29, 2026

Claims this lesson makes

Review frequentlyModel prices and context tiers are vendor-controlled, time-sensitive inputs to cost comparisons.

Primary reading

Next content review: 2026-08-29. Source links establish the lesson's claims; interactive numbers remain labeled simulations unless explicitly identified as measured data.