A Slack message landed in an ML team's channel on July 25, the morning after Claude Opus 5 shipped: "Our summarisation agent is returning half a sentence. Nothing in the code changed." Nothing in the code had changed. But one thing underneath it had, quietly and without fanfare in the changelog headline: on Opus 4.8, leaving the thinking parameter out of a request meant no thinking. The same request on Opus 5 thinks.
That is not a minor tuning difference.
The max_tokens parameter now caps thinking plus response text together, out of one shared budget, so a call tuned tightly last week can now run out of room mid-answer.
If you have been seeing truncated outputs since late July and cannot work out why, start there.
What actually changed in Claude Opus 5
Claude Opus 5 has a 1M token context window, 128k max output tokens, and thinking on by default. Those three facts appeared together in Anthropic's release notes on July 24, 2026, and most of the coverage focused on the first two. The third one is the one that breaks things.
This is generally available behaviour on Claude Opus 5 onward, enforced on each request, and it is a breaking change from Claude Opus 4.8, where disabling thinking was independent of the effort level. The practical consequence: every API call that previously carried no thinking parameter now silently allocates a reasoning budget before it writes a single word of the actual response.
One max_tokens budget, one pool. On Opus 5, thinking takes its slice first, and the answer is whatever fits in the rest.
A summarisation prompt that reliably returned 400 tokens of structured output on Opus 4.8 can return a fragment on Opus 5 with no error, no warning, and no indication in the API response that the answer was cut short.
The ceiling you can't override
Extended thinking cannot be turned off in Claude when using Claude Opus 5 via the consumer UI. On the Claude API, thinking can be turned off at effort levels high and below, but attempting to disable thinking at xhigh or max effort returns an error.
That matters for teams running agents at maximum compute. If your pipeline sets effort to xhigh and tries to suppress thinking - either because you do not want the latency or because you pre-compute the reasoning yourself - you will get a 400 back.
Send thinking: {type: "disabled"} with effort set to xhigh or max and you get a 400 back. You can disable thinking up to high effort and no further.
The workaround Anthropic recommends: where possible, keep thinking enabled and control token cost with lower effort levels; for integrations that must keep thinking disabled, see the prompting mitigations for running with thinking disabled.
What Opus 5 is actually built for
It is worth separating the migration headache from the genuine capability step. Claude Opus 5 is Anthropic's July 24, 2026 model for complex agentic coding and enterprise work. It has a 1M-token context window, thinking on by default, and a five-level effort dial.
Anthropic's own framing is that Opus 5 comes close to the frontier intelligence of Claude Fable 5 at half the price. On a few benchmarks it doesn't come close: it goes straight past.
Pricing is $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8 and half of Fable 5's $10/$50.
The effort dial is the genuinely new control.
On Opus 5 the model decides how much to think, per turn, and effort is the dial that governs depth.
That is a meaningful architectural shift: prior models treated reasoning as a binary - you passed a thinking block or you did not. Opus 5 exposes a spectrum, which means you can tune compute cost per call rather than per deployment.
There is also a new mid-conversation capability worth noting for agent builders: you can add or remove tools between turns of a conversation while preserving the prompt cache, instead of resending a fixed tool list for the life of a session. Mid-conversation tool changes are in beta. For long agentic tasks where tool availability changes - a read-only phase followed by a write phase - this removes the need to expire and restart the cached context.
Beagle in action#eng-platform, 9:47amThe non-obvious second-order problem
Here is what most teams will miss: the truncation is silent. Opus 5 does not return an error when the answer runs out of room. It returns a partial response with a stop_reason of max_tokens, exactly as it would if you had set the limit intentionally. If you are not logging stop_reason on every call - and most teams are not - you will not catch it in monitoring. The agent will just start behaving oddly, and your first indication will be a Slack message from someone on the business side saying the summary looks off.
The deeper consequence is that "thinking on by default" inverts the normal intuition about which calls need the most headroom. On Opus 4.8 you set max_tokens to cover the output you expected. On Opus 5 you have to cover the output you expect plus however many tokens the model spends reasoning - and that reasoning budget varies by problem difficulty. A simple extraction task might consume 200 thinking tokens; a multi-step synthesis might consume 4,000. You cannot predict it reliably from the prompt length alone.
The practical fix is to treat your old max_tokens settings as lower bounds and audit them upward. A tool like a teammate such as Beagle can surface calls in a channel thread where the stop reason was max_tokens, flagging them before a user does.
Claude Opus 5 extended thinking: common questions
What does "thinking on by default" mean in Claude Opus 5?
Thinking on by default means every API call to Claude Opus 5 allocates an internal reasoning budget before writing the visible response. On Opus 4.8, omitting the thinking parameter meant no reasoning tokens were spent. On Opus 5, the same omission starts a thinking pass automatically, drawing from the same max_tokens pool as the answer.
Why is my Claude Opus 5 response being cut off mid-sentence?
The most common cause is a max_tokens value tuned for Opus 4.8, which did not think by default. On Opus 5,
the model's thinking and the response text share one token budget, so a call tuned tightly last week can now run out of room mid-answer.
Raise your max_tokens limit or explicitly lower the effort level.
Can you turn off extended thinking in Claude Opus 5?
Yes, but only up to a point.
On the Claude API, thinking can be turned off at effort levels high and below, but attempting to disable thinking at xhigh or max effort returns an error.
In the consumer Claude interface, thinking cannot be toggled off at all on Opus 5.
How does the Claude Opus 5 effort dial work?
The effort dial runs from low to max and controls how many thinking tokens the model spends per turn. Lower effort = less reasoning, lower latency, lower cost.
The model decides how much to think per turn, and effort is the dial that governs depth.
Setting effort to low on a simple extraction task is a reasonable cost-control move; setting it to max on a multi-step code review is where the model earns its price.
Is Claude Opus 5 more expensive than Opus 4.8?
The per-token rate is identical: $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. However, thinking tokens count as output, so a typical agentic call will consume more output tokens than it did on Opus 4.8, making the effective cost per task higher even if the rate card did not change.