Anthropic Releases Claude Opus 4.8: A More Honest, Capable, and Cost-Effective Flagship

On Terminal-Bench 2.1, which tests command-line agentic coding, Opus 4.8 scores 74.6%. It still trails GPT-5.5's 78.2% but jumps substantially ahead of Opus 4.7's 66.1% on the same evaluation . For agentic computer use, Opus 4.8 reaches 83.4% on OSWorld-Verified, edging past Opus 4.7 (82.8%) and GPT-5.5 (78.7%) .

On knowledge work measured by GDPval-AA, Opus 4.8 scores 1890 Elo, well ahead of GPT-5.5 (1769) and a wide gap over Gemini (1314) . On Humanity's Last Exam for multidisciplinary reasoning, it scores 57.9% with tools—Anthropic's highest general-access result—against 49.8% without tools .

Honesty and alignment: a safer model

Anthropic positions Opus 4.8 as a meaningfully more honest model. In the company's own evaluations, it is roughly four times less likely than Opus 4.7 to let coding flaws pass unremarked in its self-assessments .

Rates of misaligned behavior—including deception or cooperation with misuse—are substantially lower than in Opus 4.7 and comparable to Claude Mythos Preview, which Anthropic considers its best-aligned model . This matters for developers who rely on AI to review or generate production code and need a model that flags its own blind spots rather than confidently delivering flawed output.

Effort control arrives on claude.ai and Claude Code

The most visible user-facing change is a new effort control dial now available on claude.ai and the Cowork interface . Users can select how much computational effort Claude applies to a response across several tiers:

• Low: Faster responses that consume rate limits more slowly.
• High: The new default for Opus 4.8, providing a balance between quality and speed.
• Extra (xhigh in Claude Code settings): More thorough reasoning recommended for difficult tasks and long-running workflows.
• Max: Maximum depth for the hardest problems .

In Claude Code, Anthropic has increased rate limits to accommodate the higher token usage that comes with elevated effort levels . This gives developers finer-grained tradeoffs between latency, cost, and reasoning depth on complex coding and agentic tasks.

Dynamic workflows in Claude Code

For developers tackling very large-scale problems, Anthropic is rolling out dynamic workflows as a research preview within Claude Code for Enterprise, Team, and Max plan subscribers .

The feature lets Claude plan a task, then spawn and run hundreds of parallel subagents in a single session. Outputs are verified before being reported back, which makes the system suitable for codebase-scale migrations across hundreds of thousands of lines of code .

Pricing and fast mode

Standard pricing for Opus 4.8 remains exactly what it was for Opus 4.7: $5 per million input tokens and $25 per million output tokens . Prompt caching write and refresh rates stay consistent with the premium Opus tier .

The more significant pricing shift is on the speed side. Fast mode for Opus 4.8 delivers up to 2.5x faster output token generation and now costs $10 per million input tokens and $50 per million output tokens . That is three times cheaper than fast mode was for Opus 4.6 and Opus 4.7, where it cost $30/$150 . Anthropic has deprecated fast mode for Opus 4.6 and is directing users to migrate to fast mode for Opus 4.8 or 4.7 .

To use fast mode via the API, developers set

speed: "fast"

API availability and model identifier

The model is available today through the Claude API using the alias claude-opus-4-8, and it is supported in fast mode, prompt caching, and batch processing configurations . Anthropic's API documentation and platform release notes confirm that customers on Claude for Pro, Max, Team, and Enterprise plans can access Opus 4.8 immediately .

Project Glasswing and the approaching Mythos-class release

Alongside Opus 4.8, Anthropic sharpened its language on the timeline for making Mythos-class models generally available. Since April 7, 2026, Claude Mythos Preview has been restricted to roughly 50 defensive-security partners through Project Glasswing . The model's offensive cybersecurity capability was strong enough that Anthropic withheld it from public release .

On May 28, Anthropic updated its public framing to say it plans to release Mythos-class models to all customers in the coming weeks . This is the most explicit timeline the company has given since it launched Glasswing. It also disclosed a $65 billion Series H round at a $965 billion post-money valuation, reinforcing that the commercial path for Mythos-class models is moving from research preview toward broad availability .

What this means for developers and teams

Claude Opus 4.8 is an incremental model upgrade that makes three practical differences for builders:

• If you want to move from Opus 4.7, the transition is a drop-in replacement at identical pricing, with better coding benchmarks and a dramatically more honest self-assessment of code quality.
• If you regularly need high-throughput output and were priced out of fast mode at the old $30/$150 rates, the new $10/$50 pricing makes it viable for a wider range of production workloads.
• If you are planning for an architecture that might include Mythos-class models, Anthropic's new language suggests that the restricted preview phase is winding down and that general API access should arrive soon, though exact dates and safety controls remain unconfirmed.

For teams evaluating whether to switch, the strongest signal is to test Opus 4.8 directly on your own coding, agentic, and knowledge-work tasks—particularly for long-running sessions where self-correction, uncertainty flagging, and parallel subagent orchestration matter most.