Claude Opus 5 is Anthropic's July 2026 Opus model, at $5/$25 per million tokens, with five effort levels. As of October 21, 2026, it is a legacy model, and Anthropic points most teams to Opus 5.5.
Short answer: Use Claude Opus 5.5 for new work: it costs $4/$20 per million tokens, 20% less than Opus 5, and defaults to medium effort. Stay on Opus 5 only if your app is pinned to it, and start it at high effort, its default.
Opus 5 vs Opus 5.5: which should you use now?
Anthropic's model overview now says "start with Claude Opus 5.5 for most workloads," and files Opus 5 under legacy models that are still available. The two share the same 1M-token context window, but Opus 5.5 is cheaper on input, output and cache reads.
| Claude Opus 5 | Claude Opus 5.5 | |
|---|---|---|
| Status | Active (legacy) | Current Opus model |
| Input / output per MTok | $5 / $25 | $4 / $20 |
| Cache hits per MTok | $0.50 | $0.20 |
| Default effort | high | medium |
| Retirement | Not sooner than July 24, 2027 | Not sooner than September 22, 2027 |
The default-effort change is the part that catches teams out. Anthropic's effort docs warn that a request that omits effort "runs one level lower than it did on Claude Opus 5." If you switch model names and leave effort unset, you get medium where you used to get high, so compare outputs before you ship the swap. For a fuller read on the current pair, see our Sonnet 5.5 vs Opus 5.5 comparison and the Claude Opus 5.5 model page.
What did Opus 5 add?
Opus 5 raised quality at the same price as Opus 4.8, according to Anthropic. Its launch announcement makes these claims, all vendor-reported:
- Near-Fable quality for less: it "comes close to the frontier intelligence of Claude Fable 5 at half the price."
- Same cost as Opus 4.8: "greatly improved performance for the same cost as its predecessor."
- Benchmarks: state of the art on Frontier-Bench and GDPval-AA, still behind Mythos 5 on cybersecurity tasks, and a score on ARC-AGI 3 three times as high as the next-best model.
- Early customers: Anthropic's launch post quotes Zapier saying Opus 5 "topped Zapier's AutomationBench leaderboard."
On launch day Anthropic called it the default model on Claude Max and the strongest model on Claude Pro. That was true on July 24, 2026; check the current model picker before you rely on it. The model spec is a 1M-token context, 128K max output and a May 2026 knowledge cutoff.
How do effort levels work on Opus 5?
Effort is a per-request setting that trades token cost and response time against depth. Opus 5 supports all five levels: low, medium, high, xhigh and max. Older guides that list only three levels are out of date.
| Level | What Anthropic's docs say to use it for |
|---|---|
| low, medium | Use "liberally as your primary control for token cost and response time" |
| high | The default on Opus 5 and the recommended starting point |
| xhigh | "Demanding coding and agentic work" |
| max | The ceiling: thinking is always on and token use is highest, so try it only when xhigh falls short |
Two Opus 5 quirks matter for cost planning. First, "changing effort does not reliably shorten responses" on Opus 5, so lower effort saves thinking, not always output length. Second, thinking cannot be disabled at xhigh or max, so those levels always carry thinking tokens. The effort documentation has the full rules per model.
How much does each Claude model cost?
Opus 5 sits mid-table on price, and Opus 5.5 now undercuts it. These are API list prices from Anthropic's pricing and model pages, checked October 21, 2026.
| Model | Input / output per MTok | Status |
|---|---|---|
| Claude Fable 5.1 | $10 / $50 | Current, for demanding reasoning |
| Claude Opus 5.5 | $4 / $20 | Current, recommended default |
| Claude Opus 5 | $5 / $25 | Legacy, still available |
| Claude Opus 4.8 | $5 / $25 | Legacy, still available |
| Claude Sonnet 5.5 | $2 / $10 | Current |
The drop from Opus 5 to Opus 5.5 is 20% on both input and output. Cache hits widen the gap: $0.50 per MTok on Opus 5 versus $0.20 on Opus 5.5. If you use Claude through an app rather than the API, Pro is $20 a month and Max starts at $100. Our API price war breakdown puts Opus 5.5 pricing in wider context.
Try it yourself: does lower effort save you money?
Run one real task from your backlog twice on Opus 5, once at high and once at medium, and log the output tokens and the answer. This shows whether the cheaper level holds up on your work.
- Pick a task you already have a good answer for, such as a code review or a data summary.
- Send it at high effort, then at medium, with everything else the same.
- Compare quality first, then tokens. On Opus 5, expect thinking to shrink more than the visible answer.
- Repeat on Opus 5.5 at its medium default to see whether migrating is the bigger saving.
Who should stay on Opus 5?
Stay on Opus 5 only when switching costs more than the 20% saving. That usually means one of these:
- Pinned production apps whose prompts and evals were tuned on Opus 5 at high effort. Anthropic will keep it running until at least July 24, 2027.
- Teams mid-release that can't re-run their evaluations this quarter.
Everyone else should test Opus 5.5 at medium, then raise effort only where quality drops. For the hardest reasoning and long agent runs, Anthropic points to Fable 5.1 at $10/$50; our Fable 5.1 comparison weighs it against rival frontier models. Our Claude tool page covers the apps and plans.
