benchr
"Which Claude is best" is the wrong question. There's no single best one — there's a best one for the job in front of you and the budget you're spending, and Anthropic's lineup is built around exactly that trade. Pick wrong upward and you pay double for accuracy you didn't need. Pick wrong downward and you babysit a model that keeps missing. This guide sorts the lineup so you land on the right tier the first time.
All prices below are the published per-million-token rates from Anthropic's pricing page, and the benchmark figures come from Anthropic's model documentation. The exact numbers shift; the shape of the decision holds. If you'd rather think in dollars-per-workload than per-model, the companion piece on price per use case runs the math.
The lineup at a glance
Four tiers you can pay for, cheapest to most capable. Read this table top to bottom as a ladder: each rung buys more capability at a higher price, and the right rung is the lowest one that clears your task.
| Model | Price /1M (in/out) | Best for | Headline benchmark |
|---|---|---|---|
| Claude Haiku 4.5 | $1 / $5 | High-volume, simple, well-scoped tasks | SWE-bench Verified 73.3 |
| Claude Sonnet 4.6 | $3 / $15 | The everyday default — best value | SWE-bench Verified 79.6 |
| Claude Opus 4.8 | $5 / $25 | Hard pro coding & reasoning | SWE-bench Verified 88.6 |
| Claude Fable 5 | $10 / $50 | The hardest agentic work; top tier | SWE-bench Pro 80.3* |
One caveat on the Fable 5 benchmark above. The launch table reports SWE-bench Pro as the higher of Mythos 5 and Fable 5, and the starred rows reflect Mythos 5 — so treat 80.3% as the family ceiling rather than a clean Fable-5-only number. The honest read: Fable 5 is the strongest model you can book, and Mythos-class capability is the reason to book it. The full breakdown is in the Fable 5 launch piece.
Claude Haiku 4.5 — the high-volume workhorse
Haiku 4.5 is the cheap tier, and "cheap" here doesn't mean "toy." At $1 input / $5 output per million tokens (batch $0.5/$2.5, cache reads $0.10) with a 200K-token context window, it posts SWE-bench Verified 73.3 — genuinely strong for a model at this price. That combination is what makes it the right call for work measured in millions of calls, not hundreds.
Reach for Haiku when the task is high-volume and well-scoped: classification, extraction, routing, tagging, content moderation, simple chat, the inner loop of a pipeline that fires constantly. On that kind of work the marginal quality you'd buy by moving up a tier doesn't change the outcome, and the price difference compounds fast across volume.
Where Haiku stops being the right answer is ambiguity. If the task needs the model to reason through a genuinely uncertain situation, or to write code with non-trivial structure, you'll feel the ceiling — and that's your signal to move up to Sonnet. The full picture is in the Haiku 4.5 review.
Claude Sonnet 4.6 — the everyday default
Sonnet 4.6 is the model most people should run most of the time, and it's the one this guide defaults to. At $3 input / $15 output per million tokens (batch $1.5/$7.5, cache reads $0.30) it sits in the value sweet spot: cheap enough to leave running, strong enough that you rarely hit a wall. SWE-bench Verified 79.6, GPQA Diamond 89.9, AIME 2025 (no tools) 95.6, tau2-bench Telecom 97.9 — that's a model that handles real coding, hard math, and tool-use agents without flinching.
The case for Sonnet as your default is economic as much as technical. It's a fifth of Fable 5's output price and three-fifths of Opus 4.8's, so the right pattern is to run Sonnet by default and escalate the specific prompts that stump it, rather than paying a flagship rate on every call. Most teams that think they need a flagship are paying for headroom they use on one task in twenty.
Note one structural difference: Sonnet has no fast mode — that's an Opus-only option. If low latency at scale is a hard requirement, that pushes you toward Opus 4.8 even when Sonnet would otherwise be the pick. The full picture is in the Sonnet 4.6 review.
Start on Sonnet. Measure where it actually fails. Then pay for the upgrade only on the tasks that earned it — not on the whole workload.
Claude Opus 4.8 — the flagship daily driver
Opus 4.8 is the pro flagship, and it has the deepest published benchmark record in the lineup: SWE-bench Verified 88.6, SWE-bench Pro 69.2, Terminal-Bench 2.1 74.6, GPQA Diamond 93.6, OSWorld-Verified 83.4, and a GDPval-AA Elo of 1890. It runs $5 input / $25 output per million tokens (batch $2.5/$12.5, cache reads $0.50), with a 1M-token context window and 128K output (300K in beta).
This is the tier to choose when an answer being wrong costs you real time or real money — hard production coding, multi-step debugging, agentic workflows where a subtle mistake propagates, dense reasoning where you'll pay for the accuracy and be glad you did. The jump from Sonnet's 79.6 to Opus's 88.6 on SWE-bench Verified is exactly the kind of gap that matters on the work where correctness is the whole deliverable.
Opus is also the only tier with an optional fast mode: $10/$50 per million tokens for roughly 2.5x output speed. That's the same price as Fable 5's base rate, so it's a real decision — pay it only when latency is the constraint, not the capability. If you don't need the speed, run Opus at base; it's the identical model. The full picture is in the Opus 4.8 review.
High-volume simple
Haiku 4.5 $1/$5 per 1MEveryday default
Sonnet 4.6 Best valueHard pro coding
Opus 4.8 Pay for accuracyHardest agentic
Fable 5 Top tier you can runNeed fast mode
Opus 4.8 Opus-only optionMost capable class
Mythos 5 Not for saleClaude Fable 5 — the top tier you can actually use
Fable 5 is the newest and most capable model you can book, and Anthropic frames it plainly: Mythos-class capability you can actually use. It runs $10 input / $50 output per million tokens (cache-hit $1, with up to a 90% caching discount), with a 1M-token context window and 128K output. The premium is real — double Opus 4.8's base rate — and it's priced for the work that justifies it: the hardest agentic and coding tasks, and the cases where you simply want the strongest tier available.
On the benchmark above, Fable 5 anchors the family ceiling at SWE-bench Pro 80.3% (with the caveat that the starred figure reflects Mythos 5). The practical takeaway isn't the decimal — it's that when Opus 4.8 is leaving capability on the table on your hardest workload, Fable 5 is the next rung, and it's a rung you can actually reach.
One restriction worth knowing before you build on it: cyber, bio, and distillation requests fall back to Opus 4.8 rather than running on Fable 5. For ordinary agentic and coding work that's invisible; if your workload lives near those categories, plan for the fallback. The full launch breakdown is in the Fable 5 launch piece.
Claude Mythos 5 — the one you can't buy
Mythos 5 is Anthropic's most capable class, and it's the one to set expectations on early: it's not generally available. Access is gated under Project Glasswing, so it isn't something you can sign up for or route to through the API. Treat it as out of reach for normal use.
That's exactly why Fable 5 exists and is positioned the way it is. The whole pitch of Fable 5 — "Mythos-class capability you can actually use" — only lands because Mythos 5 itself is locked down. If you've been waiting for the top of the lineup to open up, the realistic answer is that Fable 5 is as high as you can go today, and it's built to be the usable version of that ceiling.
How to choose, in four questions
Skip the spec sheets and answer these in order. The first one that fits is your model.
Classification, extraction, routing, simple chat at scale → Haiku 4.5. Cheapest tier, and it clears the work.
Real coding, reasoning, drafting, agents — but not your hardest case → Sonnet 4.6. The default, best value.
Hard production coding, multi-step debugging, dense reasoning → Opus 4.8. Pay for the accuracy.
The hardest agentic work, or you just want the top → Fable 5. Mythos 5 isn't for sale.
The single recommendation that holds for most readers: make Sonnet 4.6 your default, keep Haiku 4.5 wired in for the high-volume inner loops, and reserve Opus 4.8 and Fable 5 for the specific tasks that genuinely need them. That mix covers the overwhelming majority of work at a fraction of the cost of running a flagship on everything, and it scales cleanly as your volume grows.
And don't over-buy. The most common Claude mistake in 2026 isn't picking a model that's too weak — it's paying flagship rates on a workload Sonnet would have handled. Start one rung lower than your instinct, watch where it actually fails, and let the failures, not the fear of them, decide when you move up.