Claude Opus 5 review covering real pricing, effort settings, and benchmarks against Opus 4.8. See who should switch and who should skip it.
Opus 5 costs exactly what Opus 4.8 cost. Same $5 per million input tokens, same $25 per million output tokens, not a cent moved. That’s the headline, and it’s also the least interesting part of this release. What actually changed is what you get for that price, and how unpredictable your bill becomes once you start touching the new effort dial.
Claude Opus 5 launched July 24, 2026 as Anthropic’s new default Opus-tier model. It performs close to Claude Fable 5 on several benchmarks while staying at half Fable’s price, ships with a 1 million token context window, and introduces a low, medium, high effort setting that trades capability for cost on a per-request basis. If you’re deciding whether to switch, the short version is this: the model is genuinely better, the pricing is flat, and the effort setting is the thing that will actually decide what you pay.
What changed from Opus 4.8 to Opus 5
The jump is bigger than the identical price tag suggests. On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8’s score. On ARC-AGI-3, independent testing from the ARC Prize Foundation put Opus 5 at 30.16% at high effort, roughly four times the best previously reported score on that benchmark.
None of that translates cleanly into a Tuesday afternoon spent drafting content. What matters more day to day is the 1 million token context window and a May 2026 knowledge cutoff, the most current of any Claude model at launch, per Anthropic’s Opus product page.
| Opus 4.8 | Opus 5 | |
|---|---|---|
| Input / output price | $5 / $25 per MTok | $5 / $25 per MTok |
| Context window | Smaller | 1M tokens |
| Effort control | No | Low / medium / high |
| Data retention | Standard | No 30-day retention requirement |
The takeaway: this is a capability upgrade at an unchanged price, which almost never happens in this market.
What it actually costs to run
Here’s where the flat pricing gets misleading. Effort settings, Fast mode, and routing mean the same workload can now bill roughly four times apart depending on how it’s configured, according to Anthropic’s pricing documentation. Two teams paying identical rates can walk away with very different invoices for the same task.
Fast mode adds another variable: about 2.5 times faster, at twice the price. Useful mid-session in Claude Code. Expensive if you leave it on by default.
- Low effort: cheapest, fastest, noticeably shallower output
- Medium effort: the practical default for most day-to-day work
- High effort: slowest and most expensive, reserved for the tasks that actually need it
The takeaway: the sticker price tells you almost nothing until you know which effort tier your workflow actually runs at.
The effort setting nobody’s pricing page warns you about
We’ve been running Opus 5 at high effort for detailed research and long-form drafting. Two things stood out immediately. It’s slow, meaningfully slower than Sonnet 5 and slower than Anthropic’s “everyday model” framing suggests. It also burns through tokens and usage limits fast on anything that runs long.
Running a full technical document at high effort took over 15 minutes to complete. That’s not a figure Anthropic puts in any marketing copy.
The output quality is why you tolerate it. Detailed and rarely needing a second pass.
That trade-off changes how you should use it. Front-load the request: full context, constraints, and desired format in one prompt, rather than refining across several. Opus 5 handles one dense prompt more efficiently than three small follow-ups, and at high effort, those follow-ups are where your token budget quietly disappears.
The task was a complete .docx document covering a peer-to-peer feature inside an antivirus product, including POC planning. One prompt, full context loaded upfront. That’s exactly the kind of structured, multi-section technical document Opus 5 is marketed for, and the wait is real. Budget accordingly.
The takeaway: batch your context into a single, complete prompt at high effort, or the effort setting will cost you more than the sticker price ever suggested.
Is Opus 5 better than Fable 5
Close, not equal. Opus 5 reaches near Fable 5 intelligence at roughly half the price, which is the entire pitch of this release. On CursorBench, it lands within half a percentage point of Fable 5’s peak, and the gap on most coding and agentic tasks is small enough that few workflows will notice it.
Where it doesn’t compete is dual-use security work. Anthropic deliberately did not train Opus 5 on offensive cybersecurity tasks, and flagged requests fall back automatically to Opus 4.8. Worth knowing before you build a pipeline around it.
Who should actually switch
Not everyone benefits equally from this release, which is easy to lose in a launch that’s mostly framed around benchmark wins.
- Existing Pro or Max subscribers: free upgrade, since Opus 5 is now the Max default. Take it.
- Heavy coding and agentic workflows: built for exactly this. Worth switching.
- Casual, low-volume users: you won’t notice the difference against Sonnet 5, which is cheaper.
- Fixed AI budgets: default to medium effort and watch actual spend for two weeks. The four-times swing is real.
Verdict
Switch if you’re already inside the Claude ecosystem. Don’t expect it to feel dramatically different from Opus 4.8 at a glance; the gain concentrates in long, complex tasks, not everyday requests.
We’d default to medium effort for most content and research work, saving high effort for tasks that genuinely need it. Build prompts to be complete on the first attempt if you’re running high effort regularly. That single habit did more for our token usage than any setting we changed.
What would change this advice: a competitor shipping comparable quality at a flat rate with no hidden multiplier. Until then, Opus 5 is a genuine upgrade with a genuinely confusing bill attached.
Frequently asked questions
Is Claude Opus 5 free to use?
No. It’s available on Claude Pro ($20/month) and Claude Max ($200/month), where it’s now the default model, and through the API at $5 per million input tokens and $25 per million output tokens.
What’s the difference between Claude Opus 5 and Sonnet 5?
Sonnet 5 is cheaper and faster for everyday tasks. Opus 5 is built for long, complex, multi-step work like large codebases or extended research, and its 1 million token context window is the main practical difference for most users.
Does Claude Opus 5 replace Claude Fable 5?
No. Fable 5 remains Anthropic’s higher tier for the most demanding frontier work. Opus 5 sits just below it, reaching near-Fable performance on several benchmarks at roughly half the cost.
How much does the effort setting change my bill?
Significantly. The same task can bill up to four times differently depending on whether it runs at low, medium, or high effort, so the setting matters more to your actual cost than the headline per-token price.
