Short version of this claude opus review: Opus 4.8 is the model I hand the work I can’t afford to babysit — multi-hour agentic coding runs, gnarly architectural refactors, reasoning problems with real money attached. At $5/$25 per million tokens it earns its price on exactly those jobs, and wastes it on almost everything else. My rating after extensive daily use of the Opus tier (4.7, now 4.8): 4.5/5, with the caveat that most people should not make it their default model.
TL;DR:
• Claude Opus 4.8 costs $5/$25 per MTok with a 1M context window and 128K max output — flagship-tier capability, no longer the top of the range.
• It earns its price on long-horizon agentic work, the hardest reasoning problems, and large multi-file refactors.
• It’s overkill for chat, summaries, routine coding, and anything high-volume — Sonnet 5 covers those at near-Opus quality for less.
• Fable 5 ($10/$50) now sits above Opus as Anthropic’s premium option; Opus 4.8 is the “most capability per dollar at the high end” pick.
• My rating: 4.5/5 as a specialist tool, not a daily driver.
Curious how verdicts like this one are reached? Our how we test page covers the testbed, task selection, and what our ratings do and don’t claim.
Where Opus Sits in the 2026 Lineup
Claude Opus 4.8 (claude-opus-4-8) is the most capable Opus-tier model Anthropic has shipped, and for years “Opus” simply meant “the best Claude.” That changed in 2026: Claude Fable 5, the first of the Claude 5 family, now sits above Opus in capability at double the price ($10/$50 per MTok, with thinking always on). Opus 4.8 has settled into a new role — Anthropic’s default recommendation for demanding work, the model you choose when Sonnet isn’t enough but Fable-level spend isn’t justified. The previous-generation Opus 4.7 remains active on the API at the same price.
Spec-wise, Opus 4.8 matches the rest of the current top tier: a 1 million token context window, 128K max output, vision input, and support for fast mode — in Claude Code, /fast runs Opus at noticeably higher output speed, which changed how often I actually reach for it. If you want the whole family mapped out first, my full Claude AI review covers the ecosystem; this piece is strictly about whether Opus deserves your harder jobs and bigger budget.
How I Tested Opus 4.8
I’ve leaned on the Opus tier heavily — first Opus 4.7, now 4.8 — and since Sonnet 5’s June 30 launch I’ve run Opus alongside it, deliberately routing the same categories of work to both: agentic sessions in Claude Code against a production TypeScript/Python codebase, contract and strategy analysis over very long contexts, and a rotating set of hard reasoning problems where I already knew the answer. No formal benchmarks here — my judgments are qualitative patterns from repeated real work, and I’ll flag where the gap between models was obvious versus where I had to squint.
Claude Opus Review Scorecard
My personal assessment after months on the Opus tier, paired against Sonnet 5 since its June 30 launch — not benchmark scores:
| Category | My rating | Notes |
|---|---|---|
| Reasoning | 5/5 | The most reliable finisher of genuinely hard problems short of Fable 5 |
| Coding | 5/5 | Big refactors and long agentic runs are where it separates from Sonnet |
| Writing | 4.5/5 | Excellent, but the gap over Sonnet is smallest here |
| Speed | 3.5/5 | Deliberate by nature; fast mode meaningfully helps in Claude Code |
| Value | 4/5 | Outstanding on the right tasks, poor as an everything-model |
| Overall | 4.5/5 | A specialist that rewards deliberate use |
Where Opus 4.8 Earns Its Price
Long-horizon agentic work
This is the clearest win. In multi-hour Claude Code sessions — migrations, dependency upgrades that ripple through dozens of files, “make the test suite green” marathons — Opus holds a plan in its head longer than Sonnet does. It backtracks less, re-reads files less redundantly, and is better at noticing that step 14 invalidated an assumption from step 3. In my experience, the failure mode where an agent confidently drifts off-course mid-task is meaningfully rarer with Opus, and on long runs that difference is worth more than the per-token premium, because a derailed session costs you the whole session.
The hardest reasoning problems
On ambiguous, multi-constraint problems — a pricing model with interacting edge cases, a legal clause interacting with three others, debugging from symptoms across systems — Opus more often finds the actual crux instead of a plausible-sounding one. Sonnet 5 gets there most of the time; Opus gets there more of the time and shows sturdier work. When the cost of a wrong answer is high, that margin is the product.
Big refactors and whole-repo comprehension
Give Opus a 1M-token context stuffed with a real codebase and ask for an architectural change, and it produces plans that respect the code that actually exists — including the ugly parts — rather than an idealized version of it. The deep-dive comparison in my Claude Opus vs Sonnet guide walks through this in detail, but the summary is: the bigger and messier the change, the more the Opus premium pays for itself.
Where Opus Is Overkill
- Everyday chat and drafting. Emails, summaries, blog outlines, quick explanations — Sonnet 5’s output is indistinguishable in practice and costs less.
- Routine coding. Writing a component, a test file, a small bug fix: my Claude Sonnet review explains why Sonnet is the better daily driver here, especially at its intro pricing.
- High-volume API workloads. Classification, extraction, routing — at $5/$25 the economics collapse immediately. That’s Haiku territory.
- Latency-sensitive anything. Opus thinks before it speaks. Fast mode narrows the gap in Claude Code, but if raw responsiveness is the requirement, you’ve picked the wrong tier.
Opus 4.8 vs Fable 5 vs Sonnet 5
| Model | Price (in/out per MTok) | Context | When I pick it |
|---|---|---|---|
| Fable 5 | $10 / $50 | 1M | The single hardest problems, where a wrong answer is expensive |
| Opus 4.8 | $5 / $25 | 1M | Long agentic runs, big refactors, demanding analysis |
| Sonnet 5 | $3 / $15 (intro $2 / $10 through Aug 31, 2026) | 1M | Everything else — my actual default |
The honest framing: Fable 5 has taken the “money is no object” crown, which makes Opus 4.8 the value play within the premium tier — most of the top-end capability at half the top-end price. My routing rule after all this testing: Sonnet by default, Opus when the task is long, interconnected, or expensive to get wrong, Fable when Opus visibly struggles. Current per-token prices for the whole lineup are on Anthropic’s official pricing page.
What People Get Wrong About Opus
- “Opus is the best Claude model.” Not anymore — Fable 5 sits above it. Opus 4.8 is the most capable Opus-tier model and the sweet spot for demanding work.
- “Defaulting to Opus guarantees better results.” On routine tasks the quality difference versus Sonnet 5 is usually invisible, and you pay for tokens either way. Model choice is a routing decision, not a status symbol.
- “Opus is slow, full stop.” It’s deliberate, but fast mode in Claude Code runs Opus at much higher output speed, and for agentic work throughput matters less than not derailing.
- “You need a Max plan to touch Opus.” Every current model, Opus included, is available on the $20/month Pro plan within its usage budget — heavier use just drains the budget faster.
FAQ
Is Claude Opus worth the price?
For long agentic coding sessions, large refactors, and high-stakes reasoning, yes — at $5/$25 per million tokens the reliability gain over cheaper models pays for itself. For everyday chat, drafting, and routine coding it is overkill; Sonnet 5 delivers near-identical results for less.
Is Opus still Anthropic’s best model?
No. Claude Fable 5, priced at $10/$50 per million tokens, now sits above Opus as Anthropic’s most intelligent generally available model. Opus 4.8 remains the most capable Opus-tier model and the default recommendation for demanding work.
What is the difference between Opus 4.8 and Opus 4.7?
Opus 4.8 is the current generation and Anthropic’s recommendation; Opus 4.7 is the previous generation, still active on the API at the same $5/$25 pricing. Both offer a 1M context window and 128K max output; fast mode is a research preview on Opus 4.8 and deprecated on 4.7. New projects should start on 4.8.
Can I use Claude Opus on the Pro plan?
Yes. The $20/month Pro plan includes all current models, Opus 4.8 among them, within its usage budget. Opus consumes that budget faster than Sonnet or Haiku, so heavy Opus users often end up on a Max plan.
How big is Claude Opus 4.8’s context window?
1 million tokens, with a 128K maximum output. In practice that is enough to load a substantial codebase or a long document set into a single conversation, which is exactly where Opus’s whole-repo comprehension is strongest.
Verdict
Opus 4.8 is a superb tool used deliberately: 4.5/5 for the work it was built for, an expensive habit everywhere else. Route your hard, long, high-stakes tasks to it and let cheaper models carry the rest. If you’re deciding between the two most common tiers, read my Opus vs Sonnet comparison next — or browse the full reviews section for the rest of the lineup, including the Haiku 4.5 review at the opposite end of the price spectrum.