Claude Opus 5 is Anthropic's flagship model in the Claude 5 family, built for the long-horizon agentic coding and deep-reasoning work that Sonnet 5 is not quite tuned for. It is a drop-in upgrade from Opus 4.8 at the same $5 input / $25 output per-million-token pricing, with the same 1M-token context window and 128K max output. The meaningful change is behavioral rather than a pricing shift: thinking now runs by default instead of requiring an explicit opt-in, the full effort ladder is available through xhigh and max, and a new fast mode trades a pricing premium for noticeably quicker output on demand. The practical question most teams should ask isn't "should I use Opus," it's "does this specific task actually need it" — because Sonnet 5 now absorbs a large share of what used to require paying Opus prices.
What changed in 2026
- Thinking on by default. Opus 4.8 and 4.7 ran without thinking unless you set it explicitly. Opus 5 flips that default, so an unmodified request now reasons before answering — which also means your output token ceiling needs headroom for thinking plus response text.
- Disabling thinking is capped. You can still turn thinking off, but only at high effort or below; pairing a disabled-thinking request with xhigh or max effort is rejected outright.
- Fast mode arrives. A research-preview mode trades roughly 2.5x higher output throughput for premium pricing — $10 in / $50 out per million tokens — available on the Claude API only, not on Bedrock, Google Cloud, or Microsoft Foundry.
- Cheaper prompt caching. The minimum cacheable prompt drops to 512 tokens, half of Opus 4.8's threshold, so shorter prompts that previously missed the cache now qualify for the discount.
- Its own rate-limit pool. Opus 5 does not share headroom with the combined Opus 4.6/4.7/4.8 bucket — moving traffic over doesn't automatically inherit whatever limit increases you'd negotiated for the older models.
Opus 5 at a glance
| Spec |
Claude Opus 5 |
| Model ID |
claude-opus-5 |
| Context window |
1,000,000 tokens (also the maximum) |
| Max output |
128,000 tokens |
| Input price |
$5 / million tokens |
| Output price |
$25 / million tokens |
| Fast mode |
$10 / $50 per million tokens, Claude API only |
| Effort levels |
low, medium, high, xhigh, max |
Where Opus 5 actually earns its price
Opus 5 is strongest on the tasks where Sonnet 5 is almost right: multi-file refactors, repository-scale reviews, long autonomous coding sessions, and problems where a wrong first attempt is expensive to catch and redo. Anthropic's own guidance leans toward xhigh effort for coding and agentic work and a high baseline elsewhere, with max reserved for the hardest, most latency-tolerant cases. Counterintuitively, low and medium effort also perform unusually well on this model — worth testing before assuming a task needs the top of the ladder.
Two things worth flagging honestly. Opus 5 ships with more assertive cybersecurity safety classifiers than the prior generation, so legitimate security-tooling or research prompts can occasionally trigger a decline, returned as a normal response rather than an error — build a fallback path if that risk matters to your product. It also tends to delegate to sub-agents more readily than Opus 4.8 did, which is useful for parallel work but worth capping explicitly if you're watching token spend.
Common mistakes
- Assuming the output limit covers only the answer. Since thinking runs by default, a tight output ceiling tuned for a thinking-off Opus 4.8 route can now truncate mid-response.
- Reaching for Opus 5 by default. It's the flagship, not the everyday model. Route routine coding, extraction, and chat traffic to Sonnet 5 first and reserve Opus 5 for tasks it demonstrably wins on.
- Ignoring the separate rate-limit bucket. Teams that migrate volume from Opus 4.8 sometimes hit limits they didn't expect, because Opus 5 quota isn't inherited automatically from the old pool.
- Skipping a refusal fallback. Code that reads the first content block unconditionally breaks on a decline. Check the stop reason first, especially on security- or research-adjacent workloads.
FAQ
Is Claude Opus 5 more expensive than Opus 4.8?
No — it's priced identically at $5 input / $25 output per million tokens. Fast mode is the only add-on, at a separate $10/$50 rate.
Do I need to change anything to get the old Opus 4.8 behavior back?
Only if you relied on thinking being off by default. Add an explicit disabled-thinking setting, capped at high effort or below, to reproduce the old default.
Is fast mode worth it?
For latency-sensitive interactive use, often yes — it substantially increases output speed. For batch or background jobs where latency doesn't matter, standard pricing is the better deal.
When should I pick Sonnet 5 instead?
For the large majority of coding, extraction, and agent work under a few dozen steps. See the direct comparison linked below for a fuller, workload-by-workload breakdown.
Where to go next
For the model most teams should try first, read our Claude Sonnet 5 review, then the direct Sonnet 5 vs Opus 5 comparison for workload-by-workload guidance. If coding is your main use case, see the best Claude 5 model for coding.