Anthropic says Opus 5.5 costs 40% less than Opus 5 on typical workloads, and the week in Claude
PLUS: Claude spots a new enzyme system in 21 hours, AI bosses brief the UN, and Claude agents trade books
Sunday hello. Here's your week in Claude, from Sep 20 to Sep 26.
This week, Anthropic launched Opus 5.5. Two AI bosses spoke to the UN. And Claude found a new enzyme system.
Plus 13 quick hits. Every claim has a link.
Anthropic says Opus 5.5 costs 40% less than Opus 5 on typical workloads
The new Opus costs less per token than Opus 5. Anthropic says it works at the level of Fable 5.1 on most work.
Opus 5.5 launched on Sep 22. It's the first model in the Claude 5.5 family. Anthropic says its tests show it "will cost 40% less than Opus 5 on typical workloads" at default settings.
Input costs $4 per million tokens and output costs $20. Per Anthropic, that's 20% less than Opus 5. Cache reads cost $0.20 per million, 60% less.
Anthropic says it writes output more than 30% faster than Opus 5.
Anthropic also raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. And subscription users got a rate limit reset they can save and use when they choose.
Per the release notes, it has a 1M token context window by default. Max output is 128k tokens. Thinking can't be turned off.
Opus 5.5 sends most cyber tasks to an older model
Anthropic says Opus 5.5 has "extremely strong cyber capabilities." So most of that work goes to another model.
Per Anthropic, users can still find and fix bugs in their code as part of routine software work. But "most cybersecurity tasks will be re-routed to Opus 4.8."
Opus 5.5 uses the same biology safeguards as Fable 5.1. Vetted groups can apply to a separate program to use it for biology research.
Anthropic ran a new test of whether models try to get around containment boundaries. It says Opus 5.5 tried about 85% less often than Opus 5 or Mythos 5.1. It says every try was low severity and self-reported.
Anthropic also says it sees signs that Opus 5.5 "often suspects it is being evaluated." It says that makes it harder to judge how the model will act in the real world.
METR says Opus 5.5 is likely a small step past Fable 5.1
METR is a nonprofit that studies AI agents. It's "reasonably confident" Opus 5.5 isn't a huge leap in AI R&D skill over Fable 5.1.
METR's Sep 22 summary of its preliminary evaluation says Opus 5.5 "likely represents a modest improvement upon Fable 5.1." It says the model is "unlikely to be able to fully automate AI R&D."
METR tested it through the API for 10 business days on five tasks. It drafted the summary, then Anthropic could review and edit it. The work ran under an unpaid agreement.
A separate METR team wrote a "highly experimental and preliminary" report on AI R&D inside Anthropic. It estimated "~1.5X overall acceleration in capabilities due to AI." It gave "perhaps 30% chance of 2X."
That team shared its conclusions but not its evidence. It didn't name a time period. So it's unclear if the estimate covers Opus 5.5's development.
Amodei and Altman ask the UN for world AI standards
Anthropic's Dario Amodei and OpenAI's Sam Altman spoke to the UN Security Council on Sep 23. Per CNN, both said AI needs international standards and cooperation.
CNN reports that Amodei spoke by video. He said, "I believe that this is the most important global security issue facing the world today."
CNN says he offered three ideas. One was narrow global deals, like a ban on using AI to make biological weapons. Another was ways for countries to verify each other's promises.
The last was shared AI testing standards, plus a way to report AI security incidents.
Altman spoke in person. CNN says no written agreements were expected from the meeting.
Claude found a new enzyme system with CRISPR-like repeats
Anthropic says Claude found a new enzyme system with only high-level direction from its scientists. Anthropic calls these early results.
On Sep 23, Anthropic announced a new life sciences research group and lab. Its scientists gave Claude one prompt to search a huge DNA database. They say their role was limited to that prompt and the lab work.
About 950 Claude agents searched for 21 hours, using 210 million tokens. One agent spotted a repeating DNA pattern next to the gene for an odd-looking enzyme. The team calls the system ART.
The enzyme at its core was known before. Anthropic says Claude "appears to be the first" to notice the system's defining features.
Anthropic says it doesn't yet know what ART does. Work on its main function is ongoing. All lab work is done by human scientists.
Anthropic let Claude agents trade books for 201 staff
In an Anthropic test, the model an agent ran on mattered more to its deals than its instructions did.
In Project Swap (Sep 24), 201 Anthropic staff each brought a book to give away. Each one chatted with Claude about what they like to read. Then a Claude agent traded books for them on a shared "trading floor."
From a five-minute chat, an agent's book ranking matched its person's on 61% of pairs. Anthropic says markets with stronger models were more efficient.
Most people who read their new book liked it, per Anthropic. The average person said they'd give Claude about a third of their yearly book budget to spend.
Quick hits
- Claude Marketplace puts plugins, agents and service partners in one place (Sep 23). It lists more than 2,000 connectors and plugins. Teams can use part of their committed Anthropic spend on Claude-powered software.
- Developers on paid Claude plans can now submit plugins to the Claude directory (Sep 25). Each one is checked and safety-scanned on submission. Once live, developers see usage analytics.
- Claude Tag (beta) can now use your own connectors for your requests in a Slack channel (Sep 24). Nobody else can use them. It's rolling out on Team plans, with Enterprise to follow.
- Opus 5.5 is in GitHub Copilot (Sep 22) for Pro+, Max, Business and Enterprise users. GitHub says the rollout "will be gradual."
- Opus is now Claude Code's default model on Pro and Team Standard plans, up from Sonnet (2.1.280, Sep 22). That matches Max, Team Premium and Enterprise.
- Fast mode for Opus 5.5 is in research preview on the Claude API (Sep 22), per the release notes. Anthropic says it runs at up to 2.5x speed. It costs $8 per million input tokens and $40 per million output.
- You can now add a tool partway through a chat without breaking the prompt cache (Claude API beta, Sep 22). You define it inside a mid-conversation system message, per the release notes.
- Anthropic is billing again for some refusals that come before any output (Sep 24). It covers the "bio", "frontier_llm" and "reasoning_extraction" categories, per the release notes. Early refusals in other categories still aren't billed.
- Cache diagnostics is out of beta on the Claude API (Sep 23). You opt in per request.
- The Compliance API Activity Feed no longer returns file names, project document names or artifact titles (Sep 24). That includes older activities, per the release notes.
- Claude Code admins can now block specific models with a new deniedModels setting (2.1.283, Sep 25). It works even when availableModels allows them.
- Claude Code 2.1.283 also fixes a Windows bug (Sep 25). The PowerShell tool could let cmd /c delete folders that Remove-Item refuses to, like drive roots and the home folder.
- Claude had about 80 minutes of elevated errors on Sep 21, from 7:50 to 9:10 p.m. Central. Per the status page, it hit requests to Mythos 5.1, Fable 5.1 and Opus 5.
Worth your time
- The Situation Report (Sep 21): how WHO's Africa office and others use Claude against a rare Ebola strain in the DRC. Per Anthropic, a report that took all day can now take under an hour.
- Coding sessions are longer and use more context (Sep 24): Anthropic's Claude Code data from March to September 2026. Claude works 3.3x longer on each prompt, and context per request grew 2.6x.
That's the week. The next issue lands Sunday, Oct 4.
Reply and tell me what you're building. I read every reply.
— Sam