Claude Opus vs Sonnet vs Haiku

Most tier comparisons tell you Opus is the smart one, Sonnet is the balanced one and Haiku is the fast one. That is true, useless, and has been the same sentence for two years. Here is what the current specs actually say — verified against Anthropic's model documentation on 23 August 2026.

	Claude Fable 5	Claude Opus 5	Claude Sonnet 5	Claude Haiku 4.5
Price (in / out per MTok)	$10 / $50	$5 / $25	$2 / $10	$1 / $5
Context window	1M	1M	1M	200k
Max output	128k	128k	128k	64k
Thinking	Adaptive, always on	Adaptive	Adaptive	Extended (older style)
Latency	Slower	Moderate	Fast	Fastest
Reliable knowledge cutoff	Jan 2026	May 2026	Jan 2026	Feb 2025

Three things in that table are worth more than the entire smart vs fast framing.

Opus 5 and Sonnet 5 are the same shape at 2.5× the price

Look at what doesn't differ. Same 1M context window. Same 128k maximum output. Same adaptive thinking. On every spec people actually check before choosing, Opus 5 and Sonnet 5 are identical.

The price is a clean 2.5× on both input and output — $5/$25 against $2/$10. So the decision is never do I need the bigger context or do I need thinking. Those come free at the Sonnet tier. The only question is whether the task is hard enough that the quality difference is worth paying 2.5×.

That reframes how to choose. Don't reason about it in the abstract — run the actual task on Sonnet 5 first. If the output is right, you are done, and you were never going to be able to tell from a benchmark table whether your particular workload sits above or below the line. Benchmarks measure aggregate difficulty; your task has a specific difficulty.

Where the 2.5× usually does pay: long agentic runs where a single wrong turn costs many downstream steps, work where an error is expensive to detect, and anything requiring sustained coherence over a very long context. Where it usually doesn't: classification, extraction, summarisation, formatting, and the enormous category of tasks that are mostly transformation.

The knowledge cutoffs differ, and nobody checks this

Opus 5's reliable knowledge cutoff is May 2026. Sonnet 5's is January 2026. Four months apart, inside the same generation.

This matters in exactly one situation and matters a lot there: when your task depends on the model knowing about something recent without being told. If you are asking about a library version, a regulation, a product that shipped this spring, Sonnet 5 has a materially older picture of the world than Opus 5 does.

It also matters in the opposite direction. If you are supplying the facts in the prompt — which you should be doing for anything time-sensitive anyway — the cutoff is irrelevant and you should not pay for it. The cutoff gap is a reason to use Opus for open-ended research questions, not a reason to use it for work where you provide the context.

Haiku 4.5 is not the small Claude 5

This is the one that trips people up, because the naming hides it. Haiku is on 4.5, a full generation behind Opus 5 and Sonnet 5. The consequences are concrete:

• 200k context, not 1M. Five times smaller. If your pipeline was built assuming a million tokens, Haiku is not a drop-in cheaper option.
• 64k max output, not 128k. Half.
• Reliable knowledge cutoff of February 2025 — roughly fifteen months older than Opus 5.
• Extended thinking rather than adaptive thinking. It uses the older explicit toggle instead of the adaptive mechanism the 5-generation models have.

None of that makes Haiku a bad model — it is the fastest thing Anthropic offers and it is genuinely near-frontier for its class. But step down one tier to save money is a much bigger step between Sonnet 5 and Haiku 4.5 than between Opus 5 and Sonnet 5, and the price ladder ($5 → $2 → $1) makes it look smaller than it is. Sonnet to Haiku halves the price and changes the architecture generation. Opus to Sonnet cuts the price 2.5× and changes nothing structural.

If cost is the pressure, try Sonnet 5 before Haiku 4.5. The saving is bigger and the compromise is smaller.

The tokenizer gotcha that breaks price comparisons

Anthropic's docs carry a note that is easy to miss and changes arithmetic: models from Claude Opus 4.7 onward use a new tokenizer, and the same text produces roughly 30% more tokens than it did on earlier models.

Per-token prices are therefore not directly comparable across that boundary. If you are migrating from a pre-4.7 model and budgeting from the published rates, you will underestimate your bill — the price per token may look flat or lower while your token count quietly rises by about a third. The exact increase depends on your content.

The fix is to measure. Run a representative sample through the new model, count actual tokens, and price from that rather than from a spreadsheet built on your old token counts.

And the tier above Opus

Claude Fable 5 ($10/$50) sits above Opus 5 and is described for long-running agents. Its adaptive thinking is always on and cannot be turned off, and its latency is rated slower — both of which follow from the same design choice. It is not a general upgrade; it is a specific bet that the task runs long enough that thinking quality compounds. For request-response work, the always-on thinking is cost you cannot opt out of.

The decision, compressed

1. Start at Sonnet 5. Same context, same output ceiling, same thinking as Opus, at 40% of the price.
2. Move up to Opus 5 if the task is long-horizon, expensive to check, or depends on knowledge after January 2026.
3. Move to Fable 5 only for genuinely long-running agentic work where you want thinking always on.
4. Drop to Haiku 4.5 for speed-critical, high-volume, short-context work — and re-verify, because you have changed generation, context window, output ceiling and knowledge cutoff all at once.

And do the thing every comparison article should say and almost none do: check these numbers before you rely on them. This table was accurate on 23 August 2026. The last four entries in Anthropic's own legacy-model list were current models within the past year. That is the actual half-life here, which is why the durable version of this question is about how to decide, not about which row wins.

---

Verified against Anthropic's models overview (platform.claude.com) on 23 August 2026. Pricing excludes Batch API discounts and prompt caching rates.