Opus 5 is now Max's default.
Claude Opus 5, released on July 24, has taken over as the default model on the Claude Max plan ($200/month), replacing Opus 4.8. Per-token pricing stays the same, two new reasoning depths ("xhigh" and "max") were added, and the context window is now fixed at one million tokens. Do nothing, and the next chat you open on Max is already running the new model.
Same price, new engine
under the default model
Until now, Claude Max's default model was Opus 4.8, which ran without thinking unless you explicitly set thinking: adaptive. On Opus 5 it's the opposite: thinking is on by default from the start. Anthropic's own docs flag this as a breaking change — the "What's new in Claude Opus 5" page warns that existing API users need to revisit max_tokens (a hard cap that now includes thinking tokens plus response text).
The other headline change is the context window. Opus 5 ships with 1 million tokens as both the default and the maximum — there is no smaller variant. Max output is capped at 128K tokens. According to TechCrunch's coverage, Opus 5 rolled out simultaneously across the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, claude.ai, Claude Code, and Cowork — positioned, as detailed below, as Fable-class intelligence at half the price.
| Opus 4.8 (before) | Opus 5 (now) |
|---|---|
| Thinking off by default | Thinking on by default |
| Thinking could be disabled at any effort level | Disabling thinking at xhigh/max returns a 400 error |
| Prompt-cache minimum: 1,024 tokens | Prompt-cache minimum: 512 tokens |
| SWE-bench Verified: lower | SWE-bench Verified: 96.0% |
Price held flat —
what moves is the "effort" dial
Per-token pricing hasn't budged from Opus 4.8. What's new is a set of levels for how much compute reasoning gets to spend.
The effort ladder is five steps: low / medium / high / xhigh / max. The default is high, but for coding and agentic work Anthropic still recommends xhigh as the starting point. At xhigh and max, though, thinking can no longer be turned off — request it disabled at those levels and the API returns a 400 error. That's the clearest behavioral break from Opus 4.8.
Five ways to dial it in
Same Opus 5, but higher effort spends more test-time compute; lower effort answers faster and cheaper.
low / medium
For short replies and routine tasks. Anthropic says these levels hold up quality reasonably well while cutting tokens and latency.
high (default)
Anthropic's recommended starting point — tune up or down from here based on your own evals.
xhigh / max
Built for multi-file refactors and long-horizon autonomous work spanning subagents. Needs a generously large max_tokens.
Keep the price the same. Raise only
the default's raw capability.
Who gets the upgrade
without lifting a finger
Max subscribers switch automatically. Pay-as-you-go API users have to change the model ID themselves.
Engineers
The xhigh recommendation and 1M-token window matter most for coding work in Claude Code or the API. But direct API users must edit model: "claude-opus-4-8" to "claude-opus-5" themselves — nothing switches on its own.
Business / non-technical Max users
The $200/month subscription doesn't change, but capability jumps at the next chat. No price hike and no approval process makes this an easy internal sell.
PMs / ops owners
Because thinking is now on by default, responses run longer and the model verifies its own work more often — Anthropic warns that leftover "add a final verification step" instructions now trigger over-verification. Worth auditing prompts before rollout.
Don't oversell it —
Opus 5 doesn't lead everywhere
Opus 5 isn't the top scorer across the board. It beats Opus 4.8 on SWE-bench Verified at 96.0%, and it edges out the higher-tier Fable 5 on Frontier-Bench v0.1 (43.3% vs. 33.7%). But on SWE-bench Pro — a benchmark for multi-day autonomous agent work — Fable 5 still leads at 80.3% versus Opus 5's 79.2%. Anthropic itself continues to recommend Fable 5 for the most demanding long-horizon autonomous tasks.
There's also real migration cost: with thinking on by default, output can get squeezed by thinking tokens unless max_tokens is adjusted, and any workflow that relied on disabling thinking at high effort levels breaks outright at xhigh/max. The sticker price didn't move, but the operational tuning isn't free.