Claude Opus vs Sonnet: Which Model for Which Job (2026)

Most comparisons of Claude‘s model tiers read like a product page: Opus is the clever one, Sonnet is the fast one, Haiku is the cheap one. That framing quietly assumes the expensive model is the better model for whatever you happen to be doing. It usually isn’t. This guide is organised around the decision rather than the specifications, and it starts from an unflattering premise: if you reach for Opus by reflex, you are probably overspending — in money on the API, and in usage headroom on a consumer plan.

Quick answer

The lineup, as of August 2026

The family has grown a tier since most comparison posts were written. Above Opus now sits Claude Fable 5, released on 9 June 2026 and described by Anthropic as its most capable widely released model. All prices below are per million tokens on the first-party Claude API.

One line there deserves attention: Opus and Sonnet now share an identical 1M-token context window and an identical 128k maximum output. Context length used to be a real differentiator between tiers and no longer separates them at all — so “I need a big context window” is not a reason to pay for Opus. Haiku is the odd one out, still a 4.5-generation model on a 200k window.

The price gap is smaller than nearly every guide says

Any Claude comparison written before late 2025 will tell you Opus costs $15 in / $75 out — five times Sonnet’s $3/$15. That was Opus 4.1 pricing. It has not held since the Opus 4.5 generation, and Anthropic kept Opus 5 flat at $5/$25 when it launched on 24 July 2026. Sonnet 5, meanwhile, launched on 30 June at an introductory $2/$10, and on 10 August Anthropic made that permanent, cancelling the planned rise to $3/$15.

The real gap today is 2.5x, not 5x — and that moves the argument. When Opus cost five times as much, the case against it was simply sticker price. At 2.5x, sticker price settles nothing: a model that costs 2.5x but finishes in one attempt instead of three is the cheaper model. The question becomes empirical — does Opus reduce your retry rate on your workload? For most workloads the honest answer is no.

One footnote that trips up cost forecasts: Claude 4.7-generation models and later, including both Opus 5 and Sonnet 5, use an updated tokeniser producing roughly 1.0 to 1.35 times as many tokens for the same text. Haiku 4.5 still uses the older one. Haiku’s per-word advantage over Sonnet is therefore wider than the 2x sticker difference implies, and any spreadsheet built on pre-4.7 token counts will understate 5-series spend.

Cost per task: three worked examples

Abstract per-token prices are hard to reason about. Here is the same arithmetic on three realistic workloads.

1. High-volume support triage (10,000 tickets)

Anthropic’s documentation works this example at roughly 3,700 tokens per conversation and puts the bill at about $37 on Haiku 4.5. Sonnet 5 is exactly twice Haiku’s rate on both input and output, and Opus 5 exactly five times, so the same volume costs about $74 on Sonnet 5 and about $185 on Opus 5 — before the tokeniser difference pushes the 5-series figures higher still. Triage is a narrow classification problem. Paying $148 more per 10,000 tickets for Opus to do it is very hard to justify.

2. Long-document analysis (200k in, 8k out, 500 runs a month)

Sonnet 5: 200,000 × $2/M = $0.40 input plus 8,000 × $10/M = $0.08 output — $0.48 per run, $240 a month. Opus 5: $1.00 + $0.20 = $1.20 per run, $600 a month. The $360 delta buys better reasoning over the document. Whether that is worth it depends on what you do with the answer: for a digest nobody acts on directly, no; for a due-diligence memo a partner signs, plausibly yes.

3. A one-hour agentic coding session

Using Anthropic’s published Managed Agents example — 50,000 input tokens, 15,000 output, one hour of runtime at $0.08/hour — Opus 5 comes to $0.705. The identical session on Sonnet 5 costs $0.33. But this is exactly where per-token maths misleads, because agent sessions are not fixed-length. If Opus resolves the task in one pass and Sonnet needs two and a half, Opus wins outright. Anthropic’s early-access partners lean on this: one reported comparable accuracy with a third fewer turns and tool calls, another 26% fewer tokens at equivalent quality. Credible for hard agentic work; not credible as a blanket claim, because most sessions do not fail twice on Sonnet in the first place.

Two underused levers cut all of these numbers: prompt caching, where a cache hit costs 10% of the base input rate, and the Batch API, which takes 50% off both input and output for anything that need not be synchronous. Applying both to a nightly job saves far more than switching tiers.

Where Opus actually earns its money

Anthropic’s positioning for Opus 5 is worth taking at face value because it is narrower than marketing usually is. The model is pitched at long-running agents and complex agentic coding, and the standout claims concern persistence and self-verification rather than raw knowledge: on Frontier-Bench v0.1 it more than doubles Opus 4.8’s score at a lower cost per task, and on the novel-problem-solving evaluation ARC-AGI 3 Anthropic reports a score three times the next-best model. Translated into work you might actually be doing, Opus is the right call for:

• Long multi-step agent runs. Twenty-plus dependent steps, where an error at step four silently poisons step nineteen. The advantage compounds with horizon length; on a three-step task it is invisible.
• Hard debugging and root-cause analysis. The recurring distinction is fixing a symptom versus fixing a cause. Cognition, which builds Devin, specifically flagged difficult debugging and root-cause analysis as where Opus 5 separates itself.
• Dense analytical work with consequences. Financial modelling, legal redlining, scientific analysis — where a plausible-but-wrong answer is worse than none, and you cannot cheaply verify the output yourself.
• Nuanced writing. Not “write a blog post” — Sonnet does that fine. Rather, writing that must hold an argumentative line over length, manage an awkward tone, or know what to leave out. The least measurable category and the most over-claimed, but real.
• Front-end and visual output. Several early-access partners independently reported better animations, layouts, and interactive artefacts, and Opus catching its own layout faults by inspecting rendered pages.

Where Sonnet is genuinely indistinguishable

The list of tasks where a blind A/B test would leave you guessing is longer, and it covers most of what most people do all day:

• Summarising documents, meetings, and threads.
• Drafting emails, briefs, specs, and first-pass documentation.
• Routine coding: writing a function to a clear spec, adding tests, refactoring a file, translating between languages or frameworks.
• Structured extraction — pulling fields out of contracts, invoices, or scraped pages into JSON.
• Retrieval-augmented question answering, where the reasoning burden sits mostly in the retrieval.
• Most tool-calling and MCP server workflows, where the model’s job is to pick the right tool and pass sane arguments.

Anthropic’s own cost-optimisation guidance in the platform documentation says the same thing in fewer words: Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning. It is unusual for a vendor to steer buyers down-market in its own docs, and it is the clearest signal available that the default should not be the flagship.

The decision table

The lever most people never touch: effort

Before you upgrade tiers, change the effort setting. Both Opus 5 and Sonnet 5 accept an effort parameter controlling how much thinking the model spends, defaulting to high on the Claude API and in Claude Code. Anthropic’s Sonnet 5 launch charts show it covering a much wider cost-performance range than its predecessor and, at higher effort, matching Opus 4.8 on some tasks outright.

This reorders the escalation path. The right sequence is: Sonnet at medium effort → Sonnet at high or extra-high → Opus at medium → Opus at high. Two of those four steps cost nothing extra per token, and one of them frequently resolves the problem. Jumping straight from “Sonnet failed” to “use Opus” skips the cheapest fix available. The logic runs downward too: Anthropic reports that even at its lowest effort setting, Opus 5 passes more tasks on Zapier’s AutomationBench than any other model, so if you are already on Opus, lowering effort is often a better economy than changing tier.

Rate limits on consumer plans

On a subscription rather than the API, the currency is not dollars but usage headroom — and the arithmetic is harsher, because Opus consumes it considerably faster than Sonnet.

The plans are Pro at $20/month, Max 5x at $100 and Max 20x at $200. Usage runs on a rolling five-hour window plus a seven-day cap, shared across Claude on the web and Claude Code. Sonnet 5 is the default on Free and Pro; Opus 5 became the default on Max at its July launch and is the strongest option available on Pro.

Three consequences. On Pro, a long Opus session eats the same pool your Sonnet work draws from for the rest of the window, so an afternoon of Opus can leave you rate-limited for routine tasks you could have run for a fraction of the budget. In Claude Code, Anthropic has historically switched Max users from Opus down to Sonnet automatically past a usage threshold — meaning that on heavy days you end up on Sonnet regardless, so you may as well choose it deliberately and keep the headroom. And Anthropic does not publish fixed prompt counts; the figures in its own Pro/Max support article have not been refreshed for the 5-series, so treat any specific hour count you read anywhere as approximate.

[Insert live content] current Pro/Max 5-hour and weekly allowances stated in hours or prompts for Sonnet 5 and Opus 5 — Anthropic’s support article still cites Sonnet 4 / Opus 4 figures

[Insert live content] the Claude Code auto-switch threshold from Opus 5 to Sonnet 5 on Max 5x and Max 20x, if still in force

When Haiku is the right answer instead

Haiku gets overlooked because it is a generation behind, but for the right shape of task the gap barely registers. Use it when the answer space is narrow and verifiable: classification, routing, sentiment, entity extraction, moderation pre-screening, converting unstructured text into a fixed schema. The schema or label set does most of the constraining, and a more capable model has nothing extra to contribute.

Its real limits are the 200k context window — a fifth of Opus and Sonnet’s — a 64k maximum output, and a mid-2025 training cutoff that rules it out for anything touching recent events. The sensible production pattern is a Haiku pre-filter routing the genuinely ambiguous minority up to Sonnet, with Opus reserved for a narrow escalation path.

On benchmark numbers

A note on rigour. Anthropic published detailed cost-performance charts for Opus 5, but its launch benchmark table for Sonnet 5 compares against Opus 4.8 rather than Opus 5, and third-party reporting is inconsistent — we found the same evaluation quoted with materially different figures across several outlets. We have therefore avoided reproducing head-to-head scores and stuck to claims Anthropic states directly. The two system cards, linked below, are the authoritative source.

[Insert live content] direct Opus 5 vs Sonnet 5 scores on a shared evaluation (SWE-bench Verified, OSWorld 2.0, or GDPval-AA) taken from the two system cards