Anthropic says Opus 5 nears Fable 5 at half the price, and the week in Claude
PLUS: a judge approves Anthropic's $1.5 billion book settlement, voice mode gets Opus and Sonnet, and AMD commits to invest up to $5 billion
Hi there. Here's what changed in Claude from Jul 19 to Jul 25.
This week, Anthropic launched Opus 5 at the old Opus price. A judge approved its $1.5 billion book settlement. And voice mode got Claude's bigger models.
Plus 13 quick hits. Every claim has a link.
Anthropic says Opus 5 nears Fable 5 at half the price
Anthropic has a new Opus model, and it costs the same as the last one.
On Jul 24, Anthropic launched Claude Opus 5. It says the model "comes close to the frontier intelligence of Claude Fable 5 at half the price."
Per Anthropic, it's the new default model on Claude Max. It's the strongest model on Claude Pro.
It costs $5 per million input tokens and $25 per million output. That's the same as Opus 4.8.
Per the API release notes, it has a 1M token context window. Max output is 128k tokens. Thinking is on by default.
Anthropic says Opus 5 is the new state of the art on coding and knowledge work tests. They include Frontier-Bench and GDPval-AA. It still trails Mythos 5 on cybersecurity tasks.
ARC-AGI 3 tests novel problems. On it, Anthropic says Opus 5 scored three times as high as the next-best model.
Fast mode runs around 2.5 times the default speed. Per Anthropic, it costs twice Opus 5's base price on the Claude Platform. In Claude Code, it runs through usage credits.
Opus 5 sends some blocked cyber requests to Opus 4.8
Opus 5 has lighter cyber blocks than Fable 5, Anthropic says. In Claude's apps, flagged requests go to an older model by default.
Per Anthropic, Opus 5's cyber classifiers let it find vulnerabilities in source code. They block "binary-based" vulnerability scanning, penetration testing and exploit generation.
Anthropic expects the classifiers to step in around 85% less often than on Fable 5. That's based on its testing. In Claude.ai, Claude Code and Claude Cowork, flagged requests fall back to Opus 4.8 by default.
On the API, you can choose to send flagged requests to another model. This works on Opus 5 or Fable 5, and these automatic fallbacks are in beta. Enterprises and researchers already in Anthropic's Cyber Verification Program get a version of Opus 5 with fewer security restrictions.
Anthropic ran an automated audit in pre-deployment testing. It says the audit found Opus 5 to be its most aligned model to date. It also says Opus 5 stays behind Mythos 5 in biology research and offensive cybersecurity.
A judge gave the final OK to Anthropic's $1.5 billion book deal
Anthropic can now start paying the authors and publishers who sued it, per TechCrunch.
On Jul 20, a federal judge gave final approval to the $1.5 billion class action settlement. TechCrunch reports this, citing Reuters. Judge Araceli Martinez-Olguin signed off on it.
Per TechCrunch, it pays $3,000 per work across an estimated 500,000 works. Authors and publishers who hold the rights share it.
TechCrunch says the deal is believed to be the largest in the history of US copyright law. Many authors and creators still don't see it as a win.
Per TechCrunch, Judge William Alsup had ruled that training AI on copyrighted text is fair use. But he found that downloading books from pirate sites was illegal on its own terms.
Per TechCrunch, his ruling doesn't settle the legal question industrywide. It was a single district court decision.
Claude's voice mode now runs on Opus and Sonnet, not only Haiku
Voice mode can now use Claude's bigger models and the tools you've connected.
On Jul 23, Anthropic updated voice mode. Until then, it only ran on Claude Haiku, which Anthropic chose for speed. Now you can pick Opus or Sonnet, and switch models mid-conversation.
It can reach tools you've connected, like Gmail and Slack. Claude asks for permission before it uses one.
Voice mode now speaks many more languages on every plan. They include Hindi, Japanese and Korean.
Claude doesn't detect your language on its own. You ask out loud or pick it in voice settings. Your other language settings don't carry over.
It's in beta for all chat users on mobile, desktop and web. The Free plan gets Claude Haiku, one connected tool and all languages. Paid plans get the bigger models and all connected tools.
Anthropic says it works best from your phone.
Voice chats count toward your regular usage limits.
Anthropic says Claude writes about 80% of the code it merges
Anthropic's security team shared how it guards code that Claude mostly writes.
The Jul 21 post is by Deputy CISO Jason Clinton. He says Claude authors about 80% of the code merged at Anthropic.
An internal version of Claude Tag merges more than half of all code. Human engineers own final approval.
Per Clinton, its engineers now ship 8x as much code per quarter on average. That's compared with 2021 to 2025.
The share of pull requests with substantive review comments grew from 16% to 54%, he says. He adds that its automated checks would have caught about a third of the bugs behind past claude.ai incidents.
Clinton says that after a model upgrade, an incident response agent messaged another Claude on Slack. It did this on its own initiative. It asked that agent to push a fix.
A human review gate caught it, as designed.
AMD commits up to $5 billion to Anthropic in a chip deal
Anthropic will add up to 2 gigawatts of AMD chips.
On Jul 22, AMD and Anthropic announced a partnership. Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs. The first gigawatt starts deploying in the first half of 2027.
AMD has committed to a strategic equity investment of up to $5 billion in Anthropic "in the future."
The two will use Claude to tune workloads for AMD Instinct GPUs. They'll also use it to speed up AMD's ROCm software work. AMD will broadly adopt Claude across its engineering and product teams.
A new test has AI fly a drone to follow a person
Claude Fable 5 passed most steps, but couldn't fly between rooms on its own.
On Jul 24, Anthropic shared Project Pilot. Andon Labs built a test called Drone-Bench, in consultation with Anthropic. An AI model must fly a drone through an office to find and follow a person.
Andon ran the tests. Anthropic says it hasn't been given access to Drone-Bench. Andon tested 15 models from three developers.
Per Anthropic, Fable 5 did best. It passed the baseline on every sub-task but one: turning office videos into a 3D model.
On the real drone, those errors meant it couldn't fly between rooms on its own. It flew into a wall it thought was a doorway.
Over 10 simulations, Fable 5 reached the baseline on average for only three of five sub-tasks.
Anthropic says what models can do is about six months ahead of what they do consistently.
It lists limits. The drones moved slowly. Only one office floorplan was tested, with a limited number of people.
Andon didn't test outdoors in large crowds, among many other factors.
Quick hits
- Anthropic is committing $200 million to outside research on preparing for AI's economic impacts (Jul 22). It plans to mainly fund projects of $5 to $30 million, though it's flexible upward for well-scoped, high-potential-impact projects. It won't directly fund anything below $1 million.
- Researchers can apply for up to $50,000 in Claude credits to study rare genetic diseases (Jul 20). The grants cover six months, and applications are open through Aug 2. One track is for basic science, and one for early-stage biotechs.
- Anthropic gave another $20 million to Public First Action (Jul 21). That brings its total to $40 million. Anthropic says the money can't be used to influence the election of any candidate.
- You can now ask Claude about the Anthropic Economic Index with a new connector (Jul 22). You turn it on in the connectors directory. Anthropic notes the Index reflects Claude usage, not the labor market as a whole.
- Anthropic added three role-based Claude certifications (Jul 23). Per Anthropic, over 36,000 consultants at more than 1,300 organizations have been certified since March. Exams are for Claude Partner Network members.
- Opus 5 is in GitHub Copilot for Pro+, Max, Business and Enterprise users (Jul 24). Rollout is gradual. Business and Enterprise admins must turn on the Opus 5 policy.
- You can now add or remove tools mid-conversation and keep the prompt cache (beta, Jul 24). Per the release notes, it works on Fable 5, Mythos 5, Opus 4.8 and Opus 5.
- Fast mode for Opus 4.7 is gone (Jul 24). Fast requests to it now return an error, with no fallback to standard speed. Opus 4.7 stays available at standard speed.
- Claude Managed Agents can now take an effort level (Jul 22). A session can also start with up to 50 initial events. Then work begins in the same call.
- Claude Code 2.1.219 makes Opus 5 the default Opus model (Jul 24). Subagents can now spawn their own subagents, up to depth 3 by default.
- Claude Code now caps running subagents at 20 by default (2.1.217, Jul 21). That keeps one message from fanning out unbounded background agents.
- Claude no longer runs /verify and /code-review on its own in Claude Code (2.1.215, Jul 19). On Jul 22, 2.1.218 made /deep-research start only when you invoke it.
- Multiple Claude models had elevated errors for about an hour on Jul 21. Per Anthropic's status page, it ran from 15:28 to 16:26 UTC. That's 10:28 to 11:26 am Central.
Worth your time
- How Datadog built a "universal machine tool" for Claude Code (Jul 21): an agent writes specs. A deterministic kernel then writes the application code.
- Verification loops in Claude Code with skills (Jul 22): how to turn manual checks into skills. Then Claude checks its own work.
- How the designer of Claude Design uses it (Jul 24): Nate Parrott on how he explores ideas early in Claude Design (beta).
- How Rakuten builds agents overnight (Jul 20): Rakuten's Yusuke Kaji tests each new Claude with "stretch tasks."
That's the week. The next issue lands Sunday, Aug 2.
Reply and tell me what you're building. I read every reply.
— Sam