QWE AI Academy Here’s the hot take everyone launching a takedown of Opus 5 today is going to miss: the interesting thing about Claude Opus 5 isn’t that it’s smarter. It’s that Anthropic finally admitted most of us don’t need smarter – we need cheaper thinking on demand.
The company said it directly. In their launch post, Opus 5 comes close to the frontier intelligence of Claude Fable 5 at half the price, and is the new state-of-the-art on evals like Frontier-Bench and GDPval-AA, though it remains behind Mythos 5 on cybersecurity tasks. Translation: it’s not the smartest model in the lineup. It’s the one you’re supposed to actually run in production.
Quick context: what actually shipped on July 24
Opus 5 dropped Friday alongside two beta features. Developers can now change which tools Claude can use mid-conversation without invalidating the prompt cache, and API requests flagged by safety classifiers on Opus 5 (or Fable 5) can automatically route to another model instead of being blocked.
Pricing didn’t move. Opus 5 is $5 per million input tokens and $25 per million output tokens, matching Opus 4.8 pricing. It’s the new default model on Claude Max, Anthropic’s premium consumer tier, and the strongest model available on Claude Pro.
Community reaction has been calmer than past releases. There’s no “nerfed” outrage yet, mostly because the price stayed flat and the effort dial gives people a way to opt into the old speed if they don’t want the new depth.
Hands-on: your first Opus 5 call (and the setting you must change)
If you’re calling the API, the model ID is claude-opus-5. On AWS it’s anthropic.claude-opus-5 through Amazon Bedrock, on Google Cloud it’s claude-opus-5, and it’s also available through Microsoft Foundry.
curl https://api.anthropic.com/v1/messages
-H "x-api-key: $ANTHROPIC_API_KEY"
-H "anthropic-version: 2023-06-01"
-H "content-type: application/json"
-d '{
"model": "claude-opus-5",
"max_tokens": 1024,
"effort": "medium",
"messages": [
{"role": "user", "content": "Refactor this Python function for readability: def f(x): return [i*2 for i in x if i%2==0]"}
]
}'Notice the effort field. This is the thing to understand before you do anything else.
On Claude Opus 5, requests run with thinking on by default – the model decides when and how much to think on each turn, and the effort parameter is the control for thinking depth. That’s a behavior change from Opus 4.8, which sat idle unless you explicitly turned thinking on. If you swap the model string in an old codebase and change nothing else, your token bill will go up. This is not a bug. This is the default.
Picking your effort level
- low – quick classification, formatting, short answers. Skips thinking on simple problems.
- medium – the honest sweet spot for most work: refactors, drafting, ordinary Q&A.
- high – the default. Deep reasoning, multi-step planning, hard debugging.
Harvey, the legal AI company, said Opus 5 achieved similar performance to Opus 4.8’s maximum-reasoning mode while generating 26% fewer tokens on average. That number is the real reason to pay attention – same answer, fewer output tokens, and output tokens are the ones that hurt on the bill.
Common pitfalls (the ones people will hit this weekend)
The docs bury a few things. Here are the ones that will bite you first.
Pro tip: If you want Fast mode’s 2.5x speedup, turn it on at the start of a conversation, not halfway through. The first time you enable fast mode in a conversation, you pay the full fast mode uncached input token price for the entire conversation context – the deeper you are, the more this costs. The cost applies once per conversation, so toggling off and on again later doesn’t repeat it. A 50K-token chat you Fast-toggled at message 20 will cost you more than the same chat started in Fast mode from message 1.
The Fast mode geography trap.Fast mode for Claude Opus 5 is available on the Claude API only; it is not currently available on Amazon Bedrock, Google Cloud, or Microsoft Foundry. If your company standardized on Bedrock for compliance reasons, the 2.5x speedup is off the table. Full stop.
The Fast mode price shock.Fast mode for Claude Opus 5 is priced at $10 per million input tokens and $50 per million output tokens. That’s exactly double the base rate. Combine that with thinking-on-by-default and a chatty effort: high setting, and a live coding session can burn budget in ways that feel disconnected from what you actually asked for.
Performance: what to actually expect
The vendor-quoted numbers are what they are – near-frontier, half the price, blah blah. What matters in practice is a specific pattern I’ve seen echoed in the launch posts.
| Task type | Recommended setting | Why |
|---|---|---|
| Interactive coding in an editor | effort: medium |
Balances thinking overhead against latency |
| Long-running agent (hours to days) | Consider Fable 5 instead | Anthropic still recommends Fable 5 for advanced projects the model may handle autonomously for days on end |
| Bulk classification / straightforward transforms | effort: low |
Kills the thinking budget you don’t need |
| Complex debugging with tool use | effort: high + fast mode if on API |
Latency matters, quality matters, budget doesn’t |
The Harvey token-reduction number is the honest headline here. If you’re already on Opus 4.8 at max reasoning, expect roughly a quarter fewer output tokens for equivalent work. Whether that offsets the thinking-on default depends on how you set effort.
When NOT to use Claude Opus 5
Every tutorial about a new model pretends it’s the answer to every prompt. It isn’t.
- Bulk data pipelines where Sonnet is fine. If Sonnet 4.5 handles your task at $3/$15, moving to Opus 5 at $5/$25 for marginally better output is just lighting money on fire.
- Genuinely frontier work.Anthropic is not claiming Opus 5 is its smartest model – that distinction still belongs to Fable 5, and rival systems retain an edge in certain domains. If you’re doing autonomous research runs or novel scientific reasoning, the cheaper model isn’t the point.
- Cybersecurity red-team work.The company said Opus 5 remains behind Mythos 5 in offensive cybersecurity and biology research. There are specialized routes for that.
- Anything already working on Opus 4.8 that you can’t budget-audit right now. The thinking-on default will change your bill. If you can’t watch that number this week, don’t migrate this week.
That last one is the boring, correct answer nobody writing an Opus 5 hot take will say out loud. A model release is not a mandate. Your old code still works.
FAQ
Do I need to change anything to use Opus 5 on Claude.ai?
No. If you’re on Max, it’s already your default.
What actually happens if I just swap claude-opus-4-8 for claude-opus-5 in my API calls?
Your requests will now use thinking by default at high effort, which means longer response times and higher token counts than the same call on Opus 4.8. The workaround is a one-line change: add "effort": "medium" (or "low") to match your old behavior more closely. Run a diff on token usage across a sample of your traffic before you flip the switch in production – the 26% reduction Harvey reported is at max-reasoning parity, not at your old settings.
Is Fast mode worth it?
Only when latency is the actual bottleneck and you’re on the Anthropic API directly. At double the base rate and no cloud-marketplace availability, Fast mode is a niche tool – great for live pair-programming sessions where you’d rather burn $50/million output tokens than watch a spinner, useless for batch work where you’d rather sleep on it and pay less.
Next step: open your existing Opus 4.8 code, add "effort": "medium" to a single test call against claude-opus-5, and diff the output token count against the same prompt on 4.8. That single measurement will tell you more about whether Opus 5 is worth migrating than any launch-day benchmark chart.