Which Claude model to pick in 2026 — Opus 4.8, Sonnet 4.6, Haiku 4.5, fable-5

Anthropic's 2026 lineup has six models, and each one fits a different job. Reaching for Claude Opus 4.8 for every task is like hauling a bag of groceries with a 30-ton dump truck — sure, it moves, but you're burning time and money for no real benefit. This article is a practical matrix: which model to use for which task, where Opus is justified, where Sonnet 4.6 or Haiku 4.5 will do the same job, and what the difference looks like on your bill.

Six models, six roles

All current Anthropic models are available through LiteAI:

Model	Role	When to pick it
claude-opus-4-8	The most powerful	Hard refactors, architecture, long multi-step reasoning
claude-opus-4-8-1m	Opus with 1M context	When you need to fit an entire repo or book into one request
claude-sonnet-4-6	Balance of quality and speed	Daily workhorse. Default for ~80% of tasks
claude-sonnet-4-6-1m	Sonnet with 1M context	Long document + moderate complexity
claude-haiku-4-5	Fastest and cheapest	Classification, renames, templates, short answers
fable-5	Reasoning flagship	Multi-step plans, agentic tasks, precise instructions

Technical parameters on LiteAI:

Model	Context	Speed	Reasoning quality	LiteAI price
claude-opus-4-8	200K	Slow	Maximum	$1 / 1M
claude-opus-4-8-1m	1M	Slow	Maximum	$1 / 1M
claude-sonnet-4-6	200K	Medium	High	$1 / 1M
claude-sonnet-4-6-1m	1M	Medium	High	$1 / 1M
claude-haiku-4-5	200K	Fast	Good	$1 / 1M
fable-5	1M	Slow	Very high	$10 / 1M

What matters: LiteAI charges a flat $1 per 1M tokens on Opus 4.8, Sonnet 4.6, and Haiku 4.5 — regardless of input, output, or context size. Anthropic direct serves the same quality but costs 9–45× more: $15/$75 for Opus, $3/$15 for Sonnet, $0.80/$4 for Haiku. fable-5 is a separate plan at $10/M (also flat), because Anthropic prices it at $5/$25.

Why "Opus for everything" is a bad strategy

Opus 4.8 is objectively the strongest model in the lineup. But "strongest" ≠ "best fit." Three reasons not to use Opus for simple tasks:

1. Higher token spend. Opus likes to "think out loud": its chain-of-thought is longer, and more intermediate reasoning ends up in the output. On the same prompt, Sonnet 4.6 will answer in 200 output tokens, while Opus might use 400–600 to describe its thought process. Cost grows 1.5–3× for the same result.

2. Slower responses. On a request where Sonnet answers in 3 seconds, Opus takes 8–15. That's critical for interactive scenarios: Tab completion in Cursor, autocomplete in an IDE, a website chatbot. The user is waiting, not "enjoying quality."

3. Quality doesn't differ on simple tasks. If you ask to "rename variable x to data", Sonnet and Haiku will do it equally well. Quality differences only show up on tasks that need deep reasoning — and on "rename," the model doesn't really "think" at all.

Rule of thumb: use Opus where its strength actually matters — complex multi-step problems. For everything else, Sonnet and Haiku are there.

The "task → model" matrix

Here's a concrete mapping. Tasks first (hard to simple), then the recommended model.

Deep reasoning and complex code

Task	Best model	Why
Refactoring a large codebase (>50 files)	claude-opus-4-8	Holds context, sees cross-dependencies
Architecture design from scratch	claude-opus-4-8 or fable-5	Trade-off analysis and long reasoning needed
Code review on 1000+ lines	claude-opus-4-8	Catches subtle bugs that weaker models miss
Writing a tricky regex	claude-sonnet-4-6	Opus gives the same answer, slower and pricier
Explaining unfamiliar legacy code	claude-sonnet-4-6	Sonnet is smart enough to understand code without overthinking
Generating tests	claude-sonnet-4-6	Sonnet gives good coverage at a medium pace

Long context

Task	Best model	Why
Analyzing a 300+ page PDF	claude-sonnet-4-6-1m or fable-5	1M context instead of 200K, Sonnet is cheaper
Analyzing an entire repo at once	claude-opus-4-8-1m	Opus + 1M = max quality on large volumes
Summarizing a long log / transcript	claude-sonnet-4-6-1m	Summarization doesn't need flagship reasoning
Reading a 50-page contract	claude-sonnet-4-6	200K context is enough, Sonnet handles it

Agentic and multi-step

Task	Best model	Why
Agent with a 10-step plan (tool use)	fable-5	Best at holding instruction structure
Cursor Composer on a complex task	claude-sonnet-4-6	Good balance of price, speed, quality
Claude Code on a refactor project	claude-opus-4-8	When you need to "think" before acting
Auto-fix linter (rename, import sort)	claude-haiku-4-5	Mechanical work, no reasoning needed

Classification, short answers, routine

Task	Best model	Why
Summarizing 5 sentences into 1	claude-haiku-4-5	Trivial task, no need for Sonnet's power
Ticket classification (bug / feature / question)	claude-haiku-4-5	Fast and cheap
Generating an email / commit template	claude-haiku-4-5	A template is a template, Haiku is enough
Parsing JSON from text	claude-haiku-4-5	Structural task, not creative
Sonnet said "I don't know" — what to try next?	claude-opus-4-8	If Sonnet failed, there's a chance Opus can

Snappy questions and IDE

Task	Best model	Why
Tab completion in Cursor / IDE	claude-haiku-4-5	Autocomplete must be instant
Inline chat in VS Code	claude-haiku-4-5 or claude-sonnet-4-6	Short question, short answer
Explain a stack trace error	claude-sonnet-4-6	Needs context and understanding, not just classification
"How do I do X in Python"	claude-sonnet-4-6	Typical question, Sonnet answers well

How much you save by picking the right model

Let's compare a scenario: 1000 requests per day, average volume — 1K input + 800 output tokens. That's a typical workload for an IDE assistant or chatbot.

Scenario 1: "Everything on Opus 4.8"

• 1M input + 0.8M output tokens per day
• On LiteAI: 1.8M × $1 = $1.8/day ≈ $54/month

Scenario 2: "Right model per task" (50% Haiku, 30% Sonnet, 20% Opus)

• 0.5M on Haiku + 0.4M on Sonnet + 0.18M on Opus = same 1.8M, same price — $1/M
• Sounds like there's no savings. But!

Scenario 3: "Accounting for Opus thinking longer"

On simple tasks (classification, templates) Opus spends 2–3× more output tokens because it generates detailed reasoning. Real measurements:

Task	Opus output	Sonnet output	Haiku output
Rename a variable	350 tokens	50	30
Ticket classification	280 tokens	80	20
Email generation	600 tokens	250	150
Paragraph summary	400 tokens	180	120

If you run all of these on Opus, output grows 2–4×. So at $1/M flat per day:

• "Everything on Opus": ~1.8M tokens → $1.8/day
• "Right model": ~0.7M tokens → $0.7/day

That's 2.5× cheaper, or ~$33/month saved. At higher volumes (10K+ requests/day) — hundreds of dollars per month.

The economics of speed

Beyond money, response time is user experience. Average response time on LiteAI (recent measurements):

Model	TTFT (time to first token)	Full answer at 500 tokens
claude-haiku-4-5	~0.4 sec	1.5–2 sec
claude-sonnet-4-6	~0.6 sec	3–4 sec
claude-sonnet-4-6-1m	~0.7 sec	3–5 sec
claude-opus-4-8	~1.5 sec	8–12 sec
claude-opus-4-8-1m	~1.8 sec	10–15 sec
fable-5	~1.5 sec	8–14 sec

If you're doing Tab completion in an IDE — the user won't wait 10 seconds for Opus when Haiku delivers in 0.4 sec. For real-time scenarios, Haiku is the only sensible choice.

When Opus 4.8 is actually needed

Concrete scenarios where Opus pays for itself:

1. Complex multi-step refactoring. "Rewrite the auth layer to match the new RBAC model" — Opus handles it, Sonnet might miss nuances.
2. Architectural decisions with trade-offs. "Pick between Postgres, ClickHouse, and DuckDB for analytics given the load, budget, and team experience" — Opus dissects variants in more depth.
3. Non-trivial edge cases. "What happens in this distributed system if three nodes fail simultaneously" — Opus holds complex scenarios better.
4. Deep code review. Large PR, subtle race conditions, places where Sonnet says "looks ok" — Opus finds the issue.
5. Long agents with branching. If an agent builds a 20-step plan with backtracking — Opus or fable-5 is more reliable.

When Sonnet 4.6 is your best friend

Sonnet 4.6 is the default you should start from. Move to Opus only if Sonnet clearly failed. Scenarios where Sonnet 4.6 is optimal:

• Writing new code of moderate complexity (features, endpoints, migrations)
• Code review of a normal PR (up to 500 lines)
• Explaining unfamiliar code
• "How do I do X" questions
• Long context via claude-sonnet-4-6-1m (analyzing documents)
• Inline chat in IDE
• Most tasks in Claude Code by default

When Haiku 4.5 is exactly right

Haiku is simple and fast. Not every task needs smart — some need fast and cheap:

• Inline hints in IDE (Tab complete)
• Classification and routing
• Generating uniform templates
• Extracting structured data (JSON parsing, regex)
• Short answers to typical questions
• When Sonnet already answered "I don't know" — don't jump to Opus, rephrase the prompt first, and only escalate if it's truly broken

fable-5 — the special case

fable-5 is Anthropic's reasoning flagship, available through LiteAI as a separate plan at $10 per 1M tokens. It's not part of the $1/M package. That's because Anthropic prices fable-5 at $5/$25, and there's no direct overlap with Opus/Sonnet.

When fable-5 is justified:

• Agentic tasks with long plans. fable-5 is noticeably more accurate than Sonnet on multi-step reasoning.
• Strict instruction following. If Sonnet skips steps in a long instruction — fable-5 holds the structure.
• Minimal hallucination in API calls. fable-5 less often invents non-existent methods.
• Deep refactoring that picks up project style. Adopts codebase conventions faster.

When not needed:

• Simple dialogue and IDE assistance
• Classification and short answers
• Anywhere Sonnet 4.6 already works

Detailed breakdown of fable-5 — in «The fable-5 model».

How to switch models on the fly

In Claude Code:

# Inside a session — slash command
/model claude-opus-4-8
/model claude-sonnet-4-6
/model claude-haiku-4-5
/model claude-opus-4-8-1m
/model fable-5

# Or at launch
claude --model claude-sonnet-4-6

In Cursor (via LiteAI API key):

{
  "models": [
    { "id": "claude-opus-4-8", "name": "Claude Opus 4.8" },
    { "id": "claude-sonnet-4-6", "name": "Claude Sonnet 4.6" },
    { "id": "claude-haiku-4-5", "name": "Claude Haiku 4.5" }
  ]
}

In LiteAI Chat, switching is a click in the model selector at the top of the window. Between Opus / Sonnet / Haiku / fable-5 you flip in one click.

Daily strategy

A ready-to-print cheat sheet:

What I'm doing	Which model I pick
Writing a new feature	Sonnet 4.6
Debugging a hard bug	Opus 4.8
Reviewing a PR > 500 lines	Opus 4.8
Reviewing a PR < 500 lines	Sonnet 4.6
Tab completion in IDE	Haiku 4.5
Inline chat in IDE	Sonnet 4.6 or Haiku 4.5
Agent with 10+ steps	fable-5
Agent with 3–5 steps	Sonnet 4.6
Classification / parsing	Haiku 4.5
Long document (50+ pages)	Sonnet 4.6 1M
Entire repo in context	Opus 4.8 1M
Don't know what to pick	Sonnet 4.6 (default)

Bottom line

Anthropic ships six models — and each one fits its own job. Opus 4.8 is not "the best Claude," it's "the smartest Claude for hard tasks." Sonnet 4.6 is the workhorse, Haiku 4.5 is speed and cheapness, fable-5 is the reasoning heavyweight for agents.

The habit of using Opus for everything is 2–3× overspending and 3–4× slower responses with no quality gain. Through LiteAI, all of these models are available at $1 per 1M tokens flat (fable-5 at $10/M), so pick by task, not by "what's the most expensive."

Try it — buy the 5M package for $5, run a typical query on each model, and see where Sonnet is enough and where you really need Opus. For $5 you'll get the full picture, and the rest of your savings build on that.