Claude Opus 4.8 launches as Anthropic raises $65 billion

Anthropic released Claude Opus 4.8 on May 28, 2026. The model is available through the Claude API and on Amazon Bedrock, among other platforms.¹² The company also introduced “dynamic workflows” for Claude Code, designed to handle large tasks with multiple subagents working in parallel.¹ Anthropic says a faster mode increases speed by a factor of 2.5.¹

Regular API pricing remains unchanged from Opus 4.7, while the faster mode has separate, higher token rates.¹ Anthropic reports improvements in coding, agentic tasks, and knowledge work.¹ Those performance claims come from the company’s own tests; they do not establish how well parallel workflows will perform on a particular software migration. Teams will need to test them against their own codebases.

Anthropic announced a $65 billion Series H funding round on May 28, 2026. The company put its post-money valuation at $965 billion. It plans to use the money to expand compute capacity, support safety and interpretability research, and scale Claude products and partnerships. Anthropic did not give launch dates for new products in the announcement.¹

The round highlights the capital needed to expand advanced AI systems. Anthropic’s new valuation exceeds OpenAI’s most recently reported valuation, though that comparison concerns funding valuations rather than revenue or market share.²³ For companies using Claude, the practical question is whether the planned investment eventually improves capacity and product options. The funding announcement alone does not establish when or whether those improvements will reach customers.

On June 2, Anthropic expanded access to Claude Mythos Preview through Project Glasswing.¹ Selected organizations can use the model to scan codebases for vulnerabilities.¹ This is not a general customer release.¹² Anthropic said it expected to make Mythos-class models available to all customers in the coming weeks, but gave no firm date.²

Anthropic describes Mythos as its strongest model for cybersecurity work and says its use comes with safeguards.¹ The phased access includes security defenders and critical infrastructure operators.¹³ That approach keeps access controls in focus: vulnerability-finding capabilities can help defenders but also require measures against misuse. For companies, the immediate question is whether they qualify for access and what conditions would govern their use of the model.

OpenAI has given some Japanese financial institutions access to GPT-5.5 to help prevent cyberattacks, Japan’s finance minister said, according to Reuters on May 29. On June 1, Japanese media reported on an OpenAI action plan that starts in finance and aims to expand access to other parts of critical infrastructure.¹²

The access could matter to banks evaluating AI for sensitive security workflows. The available information does not establish which institutions are already using the model or how they are deploying it. OpenAI help pages published later identify a model called GPT-5.5-Cyber within its Trusted Access for Cyber program. They do not establish that the Japanese banks received access through that program. Nor does the available reporting confirm a formal partnership with Japan’s government.³¹

OpenAI expanded Codex’s Computer Use feature to Windows on May 29, 2026.¹ Codex can now operate applications on a Windows PC. Users can start, review, and steer tasks from the ChatGPT mobile app while work continues on the computer.¹ This extends Codex’s existing workflow rather than introducing a separate Windows remote-control product.¹²

The update connects desktop app interaction with the ability to follow ongoing tasks away from the PC.¹ That may be useful for teams that cannot monitor Codex work at a Windows desk throughout a task. OpenAI describes the experience as early.¹ Its announcement does not establish how reliably the feature handles particular applications or workflows, so companies will need to assess those uses for themselves.

Researchers examined whether language models absorb fabricated claims during fine-tuning when the training documents explicitly identify those claims as false. They used synthetic texts and then tested the models’ responses. Across six claims, the reported average agreement rate after fine-tuning on these documents was 88.6%. The research drew attention in late May 2026.¹²

The finding matters for teams working with contradictory or unreliable training material: Explicitly denying a claim in the text did not reliably prevent models from adopting it in this experiment. It does not establish that every language model generally accepts false claims. The results come from a limited test using synthetic documents, and outcomes differed across the models and interventions examined.¹²