youmightnotneedfrontier.com

A cheaper-model field guide · v0.1

You might not need frontier.

Frontier models are for the work that earns them. For bounded, repeatable jobs, start lower—then verify and escalate when the evidence says to.

See the 12 tasks When frontier earns its keep
12curated task shapes
4lower-cost routes
0magic model scores

Curated, not ranked by frequency. Reviewed 2026-07-25.

The field guide

Start here before you reach for the top tier.

These are workload classes, not claims that providers or models are equivalent. “Local” always means: on hardware that fits the chosen model, with an eval for your task.

Showing all 12 tasks

Use a tool Data

Count, sort, deduplicate, or calculate

“Count these rows, deduplicate by email, and calculate the totals.”

Use SQL, a spreadsheet formula, or a short script. Ask a model to write the operation only if you need help.

The answer is deterministic. A language model adds variability where ordinary code can be exact and repeatable.

Guardrail + evidence

Keep frontier when The hard part is deciding what the numbers mean, reconciling ambiguous definitions, or designing the analysis—not executing the arithmetic.

Verify it Re-run the operation from the source data; include row counts and invariants.

Evidence OpenAI · Hamel Husain and Shreya Shankar

Local / open Data

Extract fields into a schema

“Pull the names, dates, totals, and invoice number from this document as JSON.”

Start local or efficient hosted with a strict output schema. Escalate only for difficult layouts or domain judgment.

The input and required fields bound the job, and schema validation gives you a cheap, objective first check.

Guardrail + evidence

Keep frontier when Scans are poor, tables span pages, fields require legal/accounting interpretation, or missing values have costly consequences.

Verify it Validate the schema, require source spans or page references, and sample against the original.

Evidence Google · Simon Willison · Stanford University

Local / open Operations

Classify, tag, or route incoming text

“Label each ticket by topic, urgency, and destination team.”

Use a small local or efficient model against a fixed label set. Let low confidence abstain.

Classification is bounded and high-volume—the shape where efficient models and task-specific tuning are strongest.

Guardrail + evidence

Keep frontier when Labels are subjective, policy-sensitive, adversarial, or trigger irreversible action without human review.

Verify it Measure precision and recall on held-out, human-labeled examples; monitor abstentions and drift.

Evidence Google · Anthropic · Hamel Husain and Shreya Shankar

Efficient hosted Text

Translate routine text

“Translate these support replies into Spanish and return only the translated text.”

Use an efficient hosted model—or a tested local model for privacy and offline work.

Official efficient-model guidance explicitly includes high-volume translation with constrained output.

Guardrail + evidence

Keep frontier when The text is legal, medical, literary, dialect-heavy, safety-critical, or must preserve subtle brand voice.

Verify it Back-translate a sample, preserve named terms with a glossary, and use native-speaker review for consequential copy.

Evidence Google · Google AI Edge

Local / open Media

Transcribe clear audio

“Transcribe this meeting recording with speaker labels and timestamps.”

Start with a local speech model or an efficient multimodal model. Keep the audio private when that matters.

Transcription is a specialized, bounded transform; current efficient and on-device model guides name it directly.

Guardrail + evidence

Keep frontier when Audio is noisy, speakers overlap, vocabulary is specialized, or a verbatim record has legal or safety significance.

Verify it Spot-check names, numbers, speaker changes, and low-confidence segments against the audio.

Evidence Google · Google AI Edge

Local / open Text

Rewrite, proofread, or change tone

“Make this email clearer and friendlier without changing the facts.”

Use a local or efficient model with explicit facts to preserve and a short style constraint.

On-device guidance demonstrates rewriting and voice-driven editing; the task does not require frontier discovery.

Guardrail + evidence

Keep frontier when The copy carries legal commitments, crisis communications, sensitive negotiation, or a distinctive creative voice that needs expert judgment.

Verify it Diff facts, names, numbers, dates, promises, and links—not just grammar.

Evidence Google AI Edge · OpenAI

Efficient hosted Text

Summarize material you provide

“Summarize this report in five bullets for an executive reader.”

Use local for privacy or efficient hosted for speed. Specify the audience, length, and facts that must survive.

Document processing and summarization are named efficient-model tasks; the supplied source bounds the answer.

Guardrail + evidence

Keep frontier when Synthesis spans conflicting sources, requires current external research, or omission could materially change a decision.

Verify it Require page or section references and compare every decision-relevant claim with the source.

Evidence Google · Google AI Edge · Anthropic

Local / open Operations

Turn notes into action items

“Convert these meeting notes into decisions, owners, due dates, and open questions.”

Use local or efficient hosted. Tell it never to invent an owner or date; mark missing fields as unknown.

This is structured extraction plus light summarization, with a clear abstention rule for information the notes do not contain.

Guardrail + evidence

Keep frontier when The meeting is politically sensitive, commitments are disputed, or deciding the real priority requires context outside the notes.

Verify it Link each item to a quote or timestamp and leave absent owners/dates explicitly unknown.

Evidence Google · Simon Willison

Efficient hosted Knowledge

Answer from supplied context

“Using only this handbook, answer the question and cite the relevant section.”

Use an efficient model with retrieval, citations, and an explicit “not in the source” response.

Good retrieval narrows the reasoning surface. The model should explain the supplied material, not invent missing policy.

Guardrail + evidence

Keep frontier when The corpus is huge or contradictory, the question needs cross-document inference, or the answer drives high-stakes action.

Verify it Open every cited passage; test unanswerable questions and conflicting-source cases.

Evidence Anthropic · Hamel Husain and Shreya Shankar · Sam's Lab

Efficient hosted Media

Draft alt text or image metadata

“Write concise alt text for these product images and tag the visible attributes.”

Use an efficient multimodal model for the first pass, then human-review context and purpose.

The visual input is bounded and the output short. The expensive part is often review, not frontier reasoning.

Guardrail + evidence

Keep frontier when Images are medically, legally, or technically specialized, or subtle relationships are essential to accessibility.

Verify it Review with the page context; confirm the text conveys purpose rather than listing every visible object.

Evidence Google · Anthropic

Efficient hosted Code

Generate known-pattern boilerplate

“Create three components that match this existing component and its tests.”

Use an efficient model when the pattern, files, constraints, and check command are explicit.

The repository supplies the pattern and deterministic tools can catch formatting, type, and test failures.

Guardrail + evidence

Keep frontier when The framework is unfamiliar, APIs are uncertain, architecture is changing, or generated code crosses security boundaries.

Verify it Run formatter, typecheck, tests, and a focused review for copied assumptions.

Evidence Anthropic · OpenAI · Sam's Lab

Balanced hosted Code

Make a small, scoped code edit

“Rename this API field in the named files, update tests, and change nothing else.”

Start balanced; step down to efficient when the edit is mechanical and the verifier is strong.

A clear boundary plus executable checks does more for safe downgrading than prompt length or a model leaderboard score.

Guardrail + evidence

Keep frontier when The task is ambiguous root-cause debugging, a cross-system migration, security-sensitive, destructive, or long-horizon autonomous work.

Verify it Inspect the diff, run the narrow test suite, and escalate on failure rather than asking the cheap model to improvise indefinitely.

Evidence OpenAI · Anthropic · Stanford University · Sam's Lab

Do not downgrade blindly

Frontier earns its keep here.

Prompt length is not difficulty. A two-line incident can demand deeper reasoning than a 3,000-word extraction request.

How this guide decides

Task first. Model second.

An embedding can find similar prompts. It cannot prove the cheapest route that will pass in your harness.

01 / Demand

How hard is the work itself?

Scope, ambiguity, reasoning depth, tools, context, and verification burden.

02 / Readiness

Can the prompt be executed as written?

Clear outcome, constraints, source context, success criteria, and non-goals.

03 / Risk

What happens if it is confidently wrong?

Stakes, reversibility, privacy, approvals, and the strength of the verifier.

04 / Route

What is the cheapest plausible pass?

Try it, check it, and escalate. Your measured outcomes outrank a static leaderboard.

Evidence ledger

Read the basis, not just the badge.

Official intended use is a starting prior. Practitioner patterns add texture. Your own held-out eval and live verifier should make the final call.

  1. Official guide Using GPT-5.6

    OpenAI. Separates Sol, Terra, and Luna workload roles; recommends choosing effort intentionally and validating on representative work.

  2. Official guide Choosing the right model

    Anthropic. Recommends an efficiency-first start for high-volume, straightforward work, with task-specific evals before upgrading.

  3. Official model guide Gemini 3.1 Flash-Lite

    Google. Names translation, transcription, extraction, classification, document processing, and summarization as efficient-model workloads.

  4. Official engineering post Bringing Gemma 4 12B to your laptop

    Google AI Edge. Demonstrates local rewriting, transcription, data analysis, and bounded coding with an on-device open model.

  5. Practitioner Structured data extraction from unstructured content using LLM schemas

    Simon Willison. Shows schema-constrained extraction across hosted and local models; useful as a concrete practitioner pattern, not a universal quality claim.

  6. Practitioner LLM Evals FAQ

    Hamel Husain and Shreya Shankar. Centers real prompts, human labels, held-out tests, and scoped checks when deciding whether a cheaper model is good enough.

  7. Research FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance

    Stanford University. Establishes the cheap-first cascade pattern: try a cheaper route, verify, and escalate when needed.

  8. Audited local decision Routing agentic-coding work by AA Index score — crux

    Sam's Lab. Concludes that model-level scores shortlist models; task shape, harness, and a live verifier decide safe routing.

Source set reviewed 2026-07-25. Model names and provider guidance change; the route logic should be rechecked before material use.