Claude Opus 4.8 launches at the same price as Opus 4.7, and the week in Claude

PLUS: Anthropic raises $65 billion at a $965 billion valuation, Claude Code can run hundreds of agents at once in research preview, and you can now set Claude's effort in claude.ai

By SamMay 31, 2026

Hello there. Here's your week in Claude, from May 24 to May 30.

This week, Anthropic launched Claude Opus 4.8 at the same price. Anthropic raised $65 billion. And Claude Code can now run hundreds of agents at once, in research preview.

Plus 13 quick hits. Every claim has a link.

Claude Opus 4.8 is out, at the same price as Opus 4.7

Anthropic launched Claude Opus 4.8 on May 28. Regular use still costs $5 per million input tokens. Output stays at $25 per million.

Per Anthropic's launch post, Opus 4.8 builds on Opus 4.7 with gains across benchmarks. Anthropic calls it "a modest but tangible improvement."

Anthropic's tests compare it with Opus 4.7. Opus 4.8 is around four times less likely to let flaws in its own code pass unremarked. Its rates of misaligned behavior are substantially lower, per Anthropic's alignment review.

Opus 4.8 has a 1M token context window by default, per the release notes. That's on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Max output is 128k tokens.

Anthropic also said it expects to be able to bring Mythos-class models to all customers in the coming weeks. It says those models need stronger cyber safeguards first.

Anthropic raises $65 billion and is now valued at $965 billion

Anthropic announced $65 billion in Series H funding on May 28. The round values the company at $965 billion post-money.

Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital led it, per Anthropic's announcement. It includes $15 billion that cloud providers had already committed. Of that, $5 billion is from Amazon.

Anthropic says its run-rate revenue crossed $47 billion earlier in May. It says the money is expected to go to safety and interpretability research, compute, and products.

Micron, Samsung and SK hynix also joined, as infrastructure partners.

Claude Code's new research preview runs hundreds of agents at once

A new Claude Code feature, dynamic workflows, splits a big job across many agents.

Claude plans the work, per Anthropic's Opus 4.8 post. Then it runs hundreds of parallel subagents in a single session. It checks its outputs before it reports back.

With Opus 4.8, Anthropic says Claude Code can now migrate hundreds of thousands of lines of code. It goes from kickoff to merge. The existing test suite is its bar.

Dynamic workflows are in research preview for Enterprise, Team and Max plans. You ask Claude to create one. Then /workflows shows your runs, per the changelog (2.1.154).

You can now set how hard Claude works on claude.ai

A new control next to the model picker sets how much effort Claude puts into a reply. It's in claude.ai and Cowork, on all plans, per Anthropic's May 28 post.

On higher settings, Claude thinks more often and more deeply. On lower settings, it answers faster and uses up your rate limits more slowly.

Opus 4.8 defaults to high effort. You can pick "extra" ("xhigh" in Claude Code) or "max." Then it spends more tokens for better results.

Anthropic says it raised Claude Code rate limits. That covers the extra tokens of higher effort levels.

Fast mode on Opus 4.8 is 3x cheaper than on older models

In fast mode, Opus 4.8 can work at 2.5x the speed. Per Anthropic, it's now three times cheaper than fast mode was for previous models.

It costs $10 per million input tokens and $50 per million output, per the launch post. Claude Code's weekly digest says that's 2x the standard rate. Fast mode on Opus 4.7 and 4.6 stays at $30/$150.

On the platform, fast mode for Opus 4.8 is a research preview on the Claude API only. That's per the release notes.

Fast mode for Opus 4.6 is deprecated. The release notes put its removal approximately 30 days after launch.

Anthropic has disclosed 1,596 open-source bugs and knows of 97 fixes

As of May 22, Anthropic had disclosed 1,596 vulnerabilities from its scans of open-source software. To its knowledge, 97 of them have been patched. That's per the May 27 guide on using Claude Opus to secure code.

The guide says finding bugs is now straightforward to run in parallel. It says the bottleneck has shifted to verifying, triaging and patching.

The guide lays out six steps, from a threat model to patching. It comes with a repo of skills and a demo harness for autonomous scans.

Claude Code users said yes to about 93% of permission prompts, per Anthropic

Anthropic shared how it limits the damage Claude can do in its apps.

Its telemetry showed users approved roughly 93% of permission prompts. That's per a May 25 post. The more approvals people see, the less attention they pay to each.

Per Anthropic, auto mode catches roughly 83% of overeager behaviors before they run. It adds that any probabilistic defense has a non-zero miss rate.

Anthropic also cites a prompt injection test by Gray Swan. On single attempts, Opus 4.7 holds attack success to roughly 0.1%. That rises to around 5–6% after 100 adaptive attempts.

Survey: 20% of social scientists use coding agents regularly

Most of the social scientists in an Anthropic survey have tried AI chatbots for research. Far fewer use coding agents.

The May 27 report covers 1,260 respondents. Of them, 81% have tried AI chatbots in research. Only 20% use coding agents regularly, more than once a week.

Among coding agent users, 86% report using Claude Code.

Twice as many researchers with typically male names use coding agents as those with female names. Researchers at top universities are 40% more likely to use them.

The sample wasn't representative. People were recruited for a study that offered Claude Max accounts. So it could tilt toward those curious about AI.

Quick hits

Opus 4.8 is in GitHub Copilot for Pro+, Business and Enterprise users (May 28). It launched with a 15X premium request multiplier until usage-based billing starts Jun 1. Rollout is gradual, and Business and Enterprise admins must turn it on.
Claude Code's auto mode no longer needs opt-in consent (2.1.152, May 27).
Claude Code's auto mode now works on Bedrock, Vertex and Foundry (2.1.158, May 30). It's for Opus 4.7 and 4.8. You opt in with an environment variable, per the changelog.
Claude Code's security-guidance plugin now reviews Claude's code changes for vulnerabilities (week of May 25). It fixes them in the same session. You install it from Anthropic's official marketplace, per the weekly digest.
Plugins in .claude/skills folders now load on their own (2.1.157, May 29). No marketplace is needed.
Claude Code's new /reload-skills re-scans skill folders without a restart (2.1.152, May 27).
Messages API responses now show how many billed output tokens were extended thinking (May 27). When streaming, it shows only on the final event, per the release notes.
On Opus 4.8, you can send system messages mid-conversation (May 28). Placement rules apply. Instructions can change without losing prompt cache hits, per the release notes.
The stop_details field on API refusals is now publicly documented (May 28). It gives a category (cyber, bio or null) and an explanation, per the release notes.
Claude Platform on AWS got more Managed Agents features (May 29). They are webhooks, multiagent orchestration and self-hosted sandboxes.
Enterprise admins can now control connector access with custom roles (May 28). They pick which connectors and tools each role gets, per the release notes.
Anthropic will open a Milan office (May 27), its sixth in Europe.
KiYoung Choi is joining Anthropic as Representative Director of Korea (May 26). That's ahead of its Seoul office opening. Per Anthropic's Economic Index, Koreans use Claude at more than 3.5 times the rate expected for the population size.

Worth your time

Chris Olah at the Vatican (May 25): the Anthropic co-founder spoke at the presentation of Pope Leo XIV's encyclical on AI. He names three questions where he thinks the Church's voice is most needed.
Zero Trust for AI agents (May 27): Anthropic's security framework for deploying autonomous agents in the enterprise. It has three tiers.
How CodeRabbit plans before code (May 27): CodeRabbit reviews 2 million PRs per week. It used Claude to build a planning layer that runs before any code is written.
Simon Willison on product-market fit (May 27): he pays $100 a month each for Anthropic's Max plan and OpenAI's Pro plan. He estimates paying API prices for his past 30 days would have cost $2,180.16.

That's the week. The next issue lands Sunday, Jun 7.

Reply and tell me what you're building. I read every reply.

— Sam