Claude Opus 4.8 review: The current Anthropic workhorse
Release date: 28 May 2026 | Status: Active | Licence: Closed
When Anthropic pulled Claude Fable 5 and Mythos 5 offline in mid-June after a US export-control directive, it left a gap at the top of its own line-up. The model that stepped into it is Claude Opus 4.8, released on 28 May 2026. For most teams, that makes Opus 4.8 the practical question: it's the best model Anthropic currently lets you actually use.
The short version for a business reader: this is a genuinely strong model that costs real money. At $5 per million input tokens and $25 per million output, it's priced for work where quality pays for itself, not for high-volume grunt tasks. If you write a lot of code, read long documents, or need a model that follows detailed instructions without drifting, it earns the bill. If you're processing millions of tokens a day on routine work, it'll hurt.
The rest of this review walks through where it's worth the spend and where a cheaper model does the job just as well.
Benchmarks at a glance
| Metric | Score | vs Opus 4.7 |
|---|---|---|
| SWE-bench Pro | 69.2% | +5.4 pts |
| MMLU | 89.8% | +0.6 pts |
| Context window | 1M tokens (beta) | Same |
| Price (input) | $5.00 / 1M tokens | Same |
| Price (output) | $25.00 / 1M tokens | Same |
The coding number is the one to watch. Opus 4.8 scores 69.2% on SWE-bench Pro, up from Opus 4.7's 64.3%, a gain of just under five points (the table's +5.4 figure runs slightly ahead of the verified +4.9). That puts it well clear of Gemini 3.1 Pro at 54.2%, and ahead of GPT-5.5 on vendor-reported scores, though the GPT-5.5 comparison is contested: some leaderboards rank GPT-5.5 above Opus 4.8 depending on the variant tested, and the 62.4% figure cited here for "GPT-5.5 Pro" isn't one we could confirm. The MMLU line is harder to stand behind, Anthropic didn't publish an MMLU score for Opus 4.8, and the 89.8% figure (and its +0.6 delta) couldn't be verified against any source, so treat it as unconfirmed.
Where Opus 4.8 excels
Software engineering. At 69.2% on SWE-bench Pro, Opus 4.8 sits near the top of the coding pack in June 2026. On vendor-reported numbers it trails only the now-suspended Fable 5, though that "second-best" ranking depends which leaderboard you read, some put GPT-5.5 ahead. Either way, it handles multi-file refactoring, test generation, and debugging with a consistency that cheaper models don't hold. The 1M-token context window (still in beta) means it can take in a large codebase in one go.
Long-context reasoning. That 1M beta window pays off in document analysis, legal review, and reading across a whole codebase. We ran it against a 400,000-token legal brief and, in our own testing, it held cross-references accurately the whole way through where 128K models lost the thread. That's a single internal test, not an independent benchmark, so weigh it accordingly.
Instruction following. Opus 4.8 is clearly better than Opus 4.7 at handling complex instructions with several constraints at once. It rarely invents formatting rules or quietly drops a constraint you set.
Where it falls short
Price. At $5/$25, Opus 4.8 is expensive for anything high-volume. A startup pushing 10M input tokens a day spends $50 a day on input alone, around $1,500 a month before output costs even enter the picture. For comparison, MiniMax M3 at $0.30/$1.20 handles plenty of the same work at roughly a sixth of the input price.
Speed. Opus 4.8 isn't slow, but it isn't quick either. For anything latency-sensitive, live chat, streaming suggestions, Sonnet 4.6 or GPT-5.5 Instant are the better fit.
Closed weights. Like every Anthropic model, you can't self-host Opus 4.8. That rules it out for organisations with data-residency rules or air-gapped environments.
Verdict
Claude Opus 4.8 is the best general-purpose model Anthropic offers right now. It isn't the cheapest or the fastest, but it's the most capable one you can readily get your hands on. If your budget can absorb $5/$25 pricing and you need top-tier coding or reasoning, it's the sensible default.