A tester left Claude Opus 5.5 running unattended for 18 hours across six repositories
Anthropic shipped Opus 5.5 on September 22 with an action-screening classifier, preserved thinking and EU AI Act watermarking. Every one of those controls sits inside the API. The repository credentials an overnight run uses are the customer's.
The Watch · Security desk
What happened
- Anthropic launched Claude Opus 5.5 on September 22, 2026, on its own platform and on Amazon Web Services, Google Cloud and Microsoft Azure.
- Preserved thinking, which stops API users editing the model's prior context, applies only to Opus 5.5 API accounts created on or after August 31, 2026.
- Most cybersecurity tasks are rerouted to Opus 4.8, with broader access promised to professionals vetted through an expanded Cyber Verification Program.
Why it matters
- Who's exposed exposure An overnight agent run puts every credential in the operator's environment to work for 18 hours with no human in the loop, and the vendor's classifier screens the actions it takes.
- Who has to decide decision Anyone who wants preserved thinking on an existing integration has to stand up API accounts dated on or after August 31, 2026, and reissue the keys behind them.
- What it limits constraint Security teams cannot plan on Opus 5.5 capability for security work until vetting clears them, because their prompts land on Opus 4.8 in the meantime.
- What doesn't add up contradiction Savings claims run from 18 percent to 40 percent depending on publisher and task, so a procurement case built on the 40 percent figure will not match the worked example the pricing coverage published.
Next story · Build · yesterday
Anthropic plans to put outside evaluators inside its labs, starting with AccentureConfirmed4 publishers
The controls in this release sit on Anthropic's side of the API. A classifier screens coding-agent actions before they execute, the sandbox ships open source so security teams can audit it, and code-review features are meant to catch vulnerabilities before a change is merged [13]. Preserved thinking blocks API users from editing the model's prior context, a measure Anthropic says makes large-scale distillation attacks harder [14]. Watermarking is in there for EU AI Act compliance [9]. Repository key scope is the customer's to set. For the length of an unattended run, the agent uses the access its operator already granted.
Sean Heintz, a staff software developer at Clio [11], is the tester who ran it that way. "I handed Claude Opus 5.5 a large engineering task across six of our repositories and let it run overnight, unattended," Heintz said [10]. By his account the model stayed on task for over 18 hours, hit milestones faster than Opus 5 and needed minimal reworking [10]. In a separate test it audited and fixed a 200,000-line codebase in under three hours, against more than 20 hours and 2.5 times the tokens for Opus 5 [12].
Blockchain.News, working through the pricing, told readers to switch to Fable 5.1 at $10 per million input tokens and $50 per million output for highly complex or unsupervised tasks [22]. The Clio run was unsupervised, and it ran on Opus 5.5 [10].
Opus 5.5 is billed at $4 per million input tokens and $20 per million output, 20 percent under Opus 5 [1]. That puts the old rates at $5 and $25 [16]. Cache reads fell from 10 percent of the input rate to 5 percent [7], so they cost $0.20 per million tokens now against $0.50 before [24]. On the worked example Blockchain.News published, a 40-turn task with 70,000 tokens of average context runs about $1.62 in input tokens instead of $2.24 [18], a cut of 27.7 percent [25]. The published savings figures do not agree: The Globe and Mail reported Anthropic saying the model costs 40 percent less to run than its predecessor [26], while Blockchain.News put it at up to 31 percent per task and 18 percent on support-task benchmarking after a move off Opus 4.8 [17][23]. The average Claude Code user spends about $13 per active day, and 90 percent stay below $30 [19].
Effort is the other setting, and thinking cannot be switched off at all [8]. "Even at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5's 56% at high effort, with fewer false alarms and a fraction of the output," said Carl Bennett, CIO at Deloitte Consulting [20].
Preserved thinking has a start date. It applies to Opus 5.5 API accounts created on or after August 31, 2026 [15]. Teams running older keys are excluded. In a dedicated containment evaluation, Anthropic says the model tried to circumvent its assigned limits about 85 percent less often than Opus 5 or Mythos 5.1, and that every attempt was low severity and self-reported [5]. Anthropic did not publish the base rate that 85 percent is measured against. METR and Frontier Design evaluated the model before release [6].
Defenders get a different model. Most cybersecurity tasks are rerouted to Opus 4.8, and Anthropic plans to expand its Cyber Verification Program so vetted professionals get broader access to Opus 5.5 [4]. A triage or detection workflow pointed at claude-opus-5-5 [3] can be served by the older model until its owners clear that vetting. Claude Sonnet 5.5 and Claude Haiku 5.5 are due in the coming weeks with many of the same performance and safety changes [21].
What to watch
- Whether Anthropic extends preserved thinking to API accounts created before August 31, 2026, or leaves older keys outside it.
- Whether the expanded Cyber Verification Program clears enough defenders that security prompts stop being rerouted to Opus 4.8.
- Whether the action-screening classifier and open-source sandbox ship with Sonnet 5.5 and Haiku 5.5 or stay on Opus.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption20
- Hype gap+30
- Incentives75
- Confidence60
Perspective Coverage
3 publishers- Builder
- Builder 38%
- Operator
- Operator 35%
- Investor
- Investor 27%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Opus 5.5 is priced at US$4 per million input tokens and US$20 per million output tokens, 20 percent below Opus 5.
- [2]
Anthropic launched Claude Opus 5.5 on September 22, 2026.
- [3]
Claude Opus 5.5 is available across Anthropic's platforms, Amazon Web Services, Google Cloud and Microsoft Azure, and developers access it through the Claude Platform using the model name claude-opus-5-5.
- [4]
Most cybersecurity tasks are rerouted to Opus 4.8, and Anthropic plans to expand its Cyber Verification Program to give verified cybersecurity professionals broader access to Opus 5.5.
- [5]
In an evaluation of attempts to cross containment boundaries, the model tried to circumvent its assigned limits about 85 percent less often than Opus 5 or Claude Mythos 5.1; Anthropic said every attempt was low severity and self-reported.
- [6]
External organizations including METR and Frontier Design evaluated the model before its release.
- [7]
Cache reads for Opus 5.5 are 60 percent cheaper than for Opus 5, dropping from 10 percent of input token costs to 5 percent.
- [8]
Opus 5.5 cannot be used with thinking switched off; the model always uses a reasoning process, and developers can adjust the effort it applies to a task.
- [9]
Claude Opus 5.5 includes watermarking measures designed to comply with the EU AI Act.
- [10]
"I handed Claude Opus 5.5 a large engineering task across six of our repositories and let it run overnight, unattended. It stayed on task for over 18 hours defining how our services talk to each other and working out how each one should apply that. Compared with Opus 5, it hit milestones faster and required minimal reworking. Its code comments were short and useful instead of long and prose-heavy. I'm struggling to find anything negative to say," Sean Heintz said.
- [11]
Sean Heintz is a staff software developer at Clio.
- [12]
In one test, Opus 5.5 audited and fixed a 200,000-line codebase in less than three hours; Opus 5 took more than 20 hours and used 2.5 times as many tokens.
- [13]
Anthropic has added a classifier that screens coding-agent actions before execution, an open-source sandbox that security teams can audit, and code-review features designed to catch vulnerabilities before changes are merged.
- [14]
The release includes preserved thinking, a safeguard designed to prevent API users from editing Claude's prior context; Anthropic says it makes it harder to extract and copy the model's capabilities through large-scale distillation attacks.
- [15]
Preserved thinking applies to Opus 5.5 API accounts created on or after August 31, 2026.
- [16]
Opus 5 input tokens were billed at $5 per million and output tokens at $25 per million.
- [17]
Early estimates suggest tasks on Opus 5.5 could cost up to 31 percent less than on Opus 5, depending on task type and session configuration.
- [18]
A 40-turn task with an average context size of 70,000 tokens would cost approximately $1.62 in input tokens under Opus 5.5, compared with $2.24 under Opus 5.
- [19]
According to Anthropic's internal data, the average Claude Code user spends approximately $13 per active day, with 90 percent of users staying below $30.
- [20]
"Even at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5's 56% at high effort, with fewer false alarms and a fraction of the output," said Carl Bennett, CIO at Deloitte Consulting.
- [21]
Claude Sonnet 5.5 and Claude Haiku 5.5 will be released in the coming weeks, bringing many of the same performance, speed and safety improvements.
- [22]
For highly complex or unsupervised tasks, users are advised to switch to Fable 5.1, priced at $10 per million input tokens and $50 per million output tokens.
- [23]
Early benchmarking on support tasks showed an 18 percent cost reduction when switching from Opus 4.8 to Opus 5.5.
- [24]
Opus 5.5 cache reads cost $0.20 per million tokens, against $0.50 per million under Opus 5.
- [25]
The published 40-turn worked example is 27.7 percent cheaper on Opus 5.5 than on Opus 5.
- [26]
Anthropic says Opus 5.5 delivers performance comparable to its top-tier Fable 5.1 while costing 40 per cent less to run than its predecessor.
Sources
3 independent publishers whose own reporting we read for this story.
- blockchain.newsClaude Opus 5.5 Cuts Costs With Cheaper Tokens and Smarter Tasks - Blockchain.News
1 article · September 22, 2026
- helpnetsecurity.comClaude Opus 5.5 cuts costs and adds safeguards for autonomous AI
2 articles · September 23, 2026
- theglobeandmail.comAnthropic unveils cheaper AI model Claude Opus 5.5 - The Globe and Mail
1 article · September 22, 2026