Product Announcements introduced Claude Opus 4.8, an upgraded version of Claude Opus building on Opus 4.7, offering improvements across benchmarks and enhanced collaborative effectiveness, available at the same price as its predecessor.

Opus 4.8 demonstrates significant advancements in various capabilities:

• Coding: It exhibits improved judgment, asking pertinent questions, self-correcting, challenging unsound plans, and building confidence for complex, multi-service explorations before significant changes. On CursorBench, it surpasses previous Opus models across all effort levels. Testers reported that it addresses comment-verbosity and tool-calling issues observed in Opus 4.7, directly translating to faster capability gains for engineers. It is approximately four times less likely than its predecessor to allow flaws in its written code to pass unremarked.
• Agentic Skills: On the Super-Agent benchmark, Opus 4.8 is the only model to complete every case end-to-end, outperforming prior Opus models and GPT-5.5 at cost parity. Tool calling is notably more efficient, requiring fewer steps for equivalent intelligence, and it successfully carries end-to-end tasks. It scores 84% on Online-Mind2Web, a substantial improvement over both Opus 4.7 and GPT-5.5, demonstrating reflective and on-task behavior essential for reliable agent workloads. It follows instructions cleanly and consistently, crucial for autonomous engineering workloads.
• Reasoning: It achieved the highest score on the Legal Agent Benchmark, being the first model to surpass 10% overall on the all-pass standard, indicating an accuracy lift for substantive legal work. In CoCounsel Legal, it delivers meaningful improvements in consistency and reasoning quality. For Databricks’ Genie AI agent, Opus 4.8 unlocks a step change in agentic reasoning, tackling deeper, multistep questions faster than any prior Opus model.
• Practical Knowledge Work: Analysis quality is consistently higher, producing faster, richer, and more information-dense outputs with a better signal-to-noise ratio. A key differentiator is its tendency to proactively flag issues with inputs and outputs, a behavior other models often miss. For financial-document workflows in Hebbia’s orchestrator, it maintains strong quality while offering noticeably better citation precision and token efficiency on retrieval.
• Honesty and Alignment: Opus 4.8 is explicitly designed to be more "honest," exhibiting a reduced tendency to jump to conclusions or make unsupported claims. Early testers report it is more likely to flag uncertainties. Its alignment assessment shows it reaches new highs on prosocial traits (supporting user autonomy, acting in user's best interest) and has substantially lower rates of misaligned behavior (deception, cooperation with misuse) compared to Opus 4.7, similar to the best-aligned Claude Mythos Preview model.
• Efficiency and User Experience: Its multimodal strength allows Genie to reason directly over PDFs and diagrams at 61% cheaper token cost than Opus 4.7. Overall, it provides a quality-of-life update, being faster, easier to collaborate with, and better at carrying context and style direction across long sessions.

Several new features are launched concurrently:

• Dynamic Workflows: Available in research preview within Claude Code (for Enterprise, Team, and Max plans), this feature enables Claude to undertake larger tasks. It plans work and can execute hundreds of parallel subagents in a single session, with Opus 4.8 allowing longer agent runs. It verifies outputs before reporting to the user, capable of handling codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge, using existing test suites.
• Effort Control: A new control in claude.ai and Cowork allows users to specify the computational effort Claude dedicates to a response. Higher effort settings lead to more frequent and deeper thought for improved responses, while lower effort settings result in faster responses and slower rate limit consumption. The default for Opus 4.8 is "high effort," which provides a balance of quality and user experience, consuming similar token amounts to Opus 4.7's default but yielding better performance. "Extra" (xhigh in Claude Code) or "max" settings spend more tokens for superior results, recommended for difficult or long-running asynchronous workflows. Rate limits in Claude Code have been increased to accommodate higher token usage.
• Messages API Enhancement: The Messages API now supports system entries within the messages array. This allows developers to update Claude’s instructions mid-task without disrupting the prompt cache or requiring a user turn, useful for dynamically adjusting permissions, token budgets, or environmental context as an agent operates.

Claude Opus 4.8 is immediately available. Pricing for regular usage remains $5 per million input tokens and$25 per million output tokens, consistent with Opus 4.7. A fast mode, operating at 2.5x the speed and three times cheaper than for previous models, is priced at $10 per million input tokens and$50 per million output tokens. Developers can access the model via claude-opus-4-8 through the Claude API. Future development includes models with similar capabilities at lower costs and a new class of models, such as Mythos-class models currently in limited use for cybersecurity, expected for broader release following the implementation of stronger cyber safeguards.