@ahmetcantryk
As of August 2026, Anthropic's model lineup spans four distinct tiers, each sitting at a different cost-quality point. Picking the right one is no longer "just grab the strongest model" — it's a deliberate trade-off between task complexity, latency tolerance, and budget. This guide is a task-by-task decision tool, not a cross-vendor comparison — we cover GPT and Gemini elsewhere, deliberately not here.
The timing also matters right now: Sonnet 5's introductory pricing ends August 31, 2026. If your team runs meaningful volume on it, that's a concrete reason to audit your bill today, not next quarter.
The 2026 Lineup at a Glance
Claude Haiku — the cheap, fast, high-volume tier. Best for simple classification, extraction, latency-sensitive UI features, and any workload where API call volume dominates the cost equation.
Claude Sonnet 5 — Anthropic's workhorse mid-tier model, launched June 30, 2026. It brought a substantial jump in coding and agentic-workflow quality over its predecessor. For most production workloads it's the "strong enough, cheap enough" sweet spot.
Claude Opus 5 — launched July 24, 2026, positioned by Anthropic as "frontier AI at half price." It's Anthropic's strongest reasoning tier — for the hardest coding, reasoning, and analysis tasks — and it has already become the default model on Claude Max while remaining the strongest option available on Claude Pro. Reach for it when quality matters more than raw cost.
Fable and Mythos — Anthropic's creative-writing-oriented tiers, distinct from the general-purpose Haiku/Sonnet/Opus line. They're not built for coding or analysis; they're positioned for narrative and long-form creative-writing use cases.
None of these four tiers substitutes for another — each targets a different workload shape. The real question usually isn't "which is best" but "which is needlessly expensive for this task, and which will underdeliver."
Decision Table by Task
Task Type | Recommended Model | Why |
|---|---|---|
Drafting copy, email, summarization | Sonnet 5 | Cost-to-quality ratio is enough for most content work |
Coding — multi-file features, refactors | Opus 5 | The hardest reasoning and coding tasks are where it earns its cost |
Coding — single-file, small edits | Sonnet 5 | Opus 5's edge shrinks on simple, scoped edits |
Agent / automation workflows (multi-step) | Sonnet 5 or Opus 5 (via a router) | Escalate to Opus 5 as complexity and stakes rise |
Deep analysis, financial or legal review | Opus 5 | Accuracy outweighs cost here |
High-volume classification, extraction | Haiku | Low cost, low latency; quality bar is comparatively forgiving |
Latency-sensitive UI features | Haiku | Speed and cost beat top-tier quality |
Creative writing, long-form narrative, character consistency | Fable / Mythos | Outside the general-purpose line, trained specifically for narrative work |
Treat this as a starting point, not a rulebook. Testing a tier down on your own workload and measuring whether quality actually drops is almost always cheaper than defaulting to the top of the stack.
Cost-Versus-Quality Trade-offs and Token Budgets
The numbers are concrete: Sonnet 5's introductory pricing runs through August 31, 2026, 11:59 PM UTC, at $2 per million input tokens and $10 per million output tokens. Starting September 1, 2026, standard pricing kicks in — $3 per million input tokens, $15 per million output tokens — a 50% increase on both dimensions. This pricing applies across the Claude Platform, Claude Code, and Amazon Bedrock.
That deadline isn't an abstract announcement — it's a concrete action item. If your usage on Sonnet 5 is heavy, audit your bill now and lock in the savings before September 1. Nothing about your architecture needs to change; the calendar does the work if you act on it.
Given Opus 5's "frontier AI at half price" positioning, some teams may find it makes sense to run Opus 5 directly instead of Sonnet 5, especially for workloads that are already complex. But for simple, repetitive work, that jump is usually an unnecessary cost increase.
When a Smaller Model Wins
Task complexity and cost-sensitivity should drive the pick — not a reflex to reach for the strongest model available. Prompt caching offers up to 90% cost savings and batch processing up to 50%, per Anthropic's own guidance. Combine the two and a smaller model — or even the same model run at a lower effort setting — frequently beats the larger one on total cost-per-outcome.
Repetitive, high-volume work (bulk classification, archive scanning, daily report generation) falls into this category almost every time. My honest take here: if you're already running the same task thousands of times a day, ignoring the caching and batch savings and calling the most expensive model on every request isn't a quality decision — it's an engineering shortcut that shows up directly on the invoice.
The Router Pattern: Mixing Tiers in One Workflow
Most production systems don't pick a single model — they mix tiers within one workflow. The common pattern: Haiku handles a first-pass triage or classification step, escalating only ambiguous or high-stakes cases to Sonnet or Opus. This keeps cost under control while preserving quality exactly where it matters.
A simple router might look like this, selecting a model by task type:
{
"router": [
{ "task": "triage_classification", "model": "claude-haiku-4-5", "escalate_if": "confidence < 0.7" },
{ "task": "draft_writing", "model": "claude-sonnet-5" },
{ "task": "multi_file_coding", "model": "claude-opus-5" },
{ "task": "high_stakes_analysis", "model": "claude-opus-5" },
{ "task": "creative_narrative", "model": "claude-fable-5" }
]
}This kind of routing layer pairs well with the caching and batch strategies covered in our guide on cutting LLM token costs. If you're designing multi-step automation or agents, our piece on multi-agent orchestration patterns is a useful extension of the same routing logic.
Quick-Pick Cheat Sheet
- Need fast and cheap, full stop? Haiku.
- Need a balanced default for most daily workflows? Sonnet 5 — but lock in the price before August 31.
- Facing the hardest coding, reasoning, or analysis work? Opus 5.
- Writing long-form fiction or creative narrative? Fable or Mythos.
- Running a workflow with mixed task types? Build a router: triage on Haiku, escalate to Sonnet 5 or Opus 5 as needed.
We covered Opus 5's launch in detail earlier; if you're migrating off Opus 4.1, our retirement and migration guide walks through that transition. Looking for a cross-vendor comparison instead? Our Sonnet 5 vs. GPT-5.6 vs. Gemini 3.5 benchmark covers that ground — this piece deliberately stays focused on choosing within Anthropic's own lineup.
Sources: Claude Sonnet 5 announcement, Claude Opus 5 announcement, Anthropic pricing page.
Frequently Asked Questions
When does Sonnet 5's introductory pricing end?
August 31, 2026, at 11:59 PM UTC. After that, pricing rises from $2 to $3 per million input tokens and from $10 to $15 per million output tokens — a 50% increase on both. If you have heavy usage, review your bill now.
Does Opus 5 replace Sonnet 5?
No, they serve different tiers. Opus 5 is for the hardest reasoning and coding work; Sonnet 5 remains the more cost-effective choice for most day-to-day, moderate-complexity workflows. Opus 5 became the default on Claude Max and the strongest option on Claude Pro, but that doesn't mean it should be the default for every task.
How are Fable and Mythos different from the Haiku/Sonnet/Opus line?
Fable and Mythos are trained specifically for creative writing, not coding or analysis. They're the right pick for long-form narrative, character consistency, and fiction generation; general-purpose tasks are better served by Haiku, Sonnet 5, or Opus 5.
Does it make sense to start with a smaller model and escalate when needed?
Yes — that's the core logic of the router pattern. Running a cheap first-pass triage on Haiku and escalating only ambiguous or high-stakes cases to Sonnet 5 or Opus 5 cuts cost while preserving quality where it counts. For a deeper look at when the smaller model is genuinely enough, see our piece on when small language models win.
