1M
6.5
32.5
Chat
Active
Claude Opus 4.8
Sharper reasoning, more honest output, and the best agentic performance the company has shipped.
Across the tested dimensions — coding, agentic task completion, knowledge work, reasoning, and computer use — Opus 4.8 either matches or improves on its predecessor, and frequently outperforms competing frontier models.
Claude Opus 4.8 is the latest version of Anthropic's top-tier AI model, succeeding Claude Opus 4.7. Rather than a ground-up redesign, it represents a focused, meaningful upgrade — one that compounds on strong foundations with measurable improvements across coding, reasoning, agentic reliability, and what Anthropic calls honesty: the model's willingness to surface uncertainty rather than paper over gaps with confident-sounding approximations.
Multiple engineering teams report that Opus 4.8 is more reliable as an autonomous coding assistant. It asks sharper clarifying questions before making large changes, pushes back when plans seem flawed, and catches more of its own mistakes before they propagate. On CursorBench — a rigorous evaluation from the Cursor team covering full end-to-end development tasks, Opus 4.8 outperformed all prior Opus models at every effort level. Tool calling is more efficient too, completing the same work with fewer intermediate steps.
In complex, multi-step autonomous workflows, Opus 4.8 shows the reliability characteristics that production AI agent deployments depend on. On a Super-Agent benchmark developed by one external partner, it was the only model to complete every case end-to-end, outperforming prior Opus versions and GPT-5.5 at equivalent cost. It's consistently better at carrying context across long sessions and following stylistic or technical direction without drift.
Opus 4.8 is the first model to surpass 10% on the all-pass standard of the Legal Agent Benchmark — a significant threshold in an industry where accuracy errors carry real professional consequences. Multiple legal AI platforms report that the improvement in consistency and reasoning quality translates directly into confidence about which attorney tasks can be delegated to AI-assisted workflows.
With an 84% score on Online-Mind2Web, Opus 4.8 ranks as the strongest computer-use and browser-agent model tested by any external team at launch. It maintains focus across long, complex web-based tasks in ways that directly benefit real-world automation pipelines.
For financial document workflows, processing dense filings, earnings reports, and structured data, Opus 4.8 maintains the quality of Opus 4.7 while improving citation precision, reducing token consumption on retrieval tasks, and proactively flagging anomalies in inputs and outputs that other models left for human reviewers to catch.
The Messages API now accepts system entries inside the messages array, not just at the top level. Developers can update Claude's instructions mid-task — changing permissions, token budgets, or environmental context — without breaking the prompt cache or routing the update through a user turn. This makes it substantially easier to build sophisticated, adaptive agent harnesses.
Opus 4.8 is Anthropic's flagship model, positioned for work where quality is the primary constraint and cost is secondary. It's the right choice when you're building production-grade AI agents, handling high-stakes professional knowledge work, or need a model that can sustain coherent context and judgment across very long sessions.
On agentic benchmarks, Opus 4.8 outperforms GPT-5.5 in specific evaluations: one external partner's Super-Agent benchmark showed Opus 4.8 completing every case that GPT-5.5 could not, at cost parity. On computer use (Online-Mind2Web), Opus 4.8's 84% score beats GPT-5.5's reported result on the same evaluation. Comparative performance varies by task type; users with specific workloads should run their own evaluations.
The headline improvements are better honesty (Opus 4.8 flags uncertainties and code flaws at a significantly higher rate), improved judgment in autonomous tasks, more efficient tool calling, and better alignment scores. Verbose comment generation and tool-calling inconsistencies reported with Opus 4.7 are addressed in this release.
Dynamic workflows let Claude plan a large software task and then execute it by spinning up hundreds of parallel subagents within a single Claude Code session. It verifies its outputs before surfacing results. It's currently in research preview and available on Enterprise, Team, and Max plans.
Opus 4.8 can reason over PDFs, diagrams, charts, and other unstructured visual content. For document-heavy workflows, it delivers this at a 61% lower token cost compared to Opus 4.7, according to one enterprise data platform's internal benchmarks.
Claude Opus 4.8 is the latest version of Anthropic's top-tier AI model, succeeding Claude Opus 4.7. Rather than a ground-up redesign, it represents a focused, meaningful upgrade — one that compounds on strong foundations with measurable improvements across coding, reasoning, agentic reliability, and what Anthropic calls honesty: the model's willingness to surface uncertainty rather than paper over gaps with confident-sounding approximations.
Multiple engineering teams report that Opus 4.8 is more reliable as an autonomous coding assistant. It asks sharper clarifying questions before making large changes, pushes back when plans seem flawed, and catches more of its own mistakes before they propagate. On CursorBench — a rigorous evaluation from the Cursor team covering full end-to-end development tasks, Opus 4.8 outperformed all prior Opus models at every effort level. Tool calling is more efficient too, completing the same work with fewer intermediate steps.
In complex, multi-step autonomous workflows, Opus 4.8 shows the reliability characteristics that production AI agent deployments depend on. On a Super-Agent benchmark developed by one external partner, it was the only model to complete every case end-to-end, outperforming prior Opus versions and GPT-5.5 at equivalent cost. It's consistently better at carrying context across long sessions and following stylistic or technical direction without drift.
Opus 4.8 is the first model to surpass 10% on the all-pass standard of the Legal Agent Benchmark — a significant threshold in an industry where accuracy errors carry real professional consequences. Multiple legal AI platforms report that the improvement in consistency and reasoning quality translates directly into confidence about which attorney tasks can be delegated to AI-assisted workflows.
With an 84% score on Online-Mind2Web, Opus 4.8 ranks as the strongest computer-use and browser-agent model tested by any external team at launch. It maintains focus across long, complex web-based tasks in ways that directly benefit real-world automation pipelines.
For financial document workflows, processing dense filings, earnings reports, and structured data, Opus 4.8 maintains the quality of Opus 4.7 while improving citation precision, reducing token consumption on retrieval tasks, and proactively flagging anomalies in inputs and outputs that other models left for human reviewers to catch.
The Messages API now accepts system entries inside the messages array, not just at the top level. Developers can update Claude's instructions mid-task — changing permissions, token budgets, or environmental context — without breaking the prompt cache or routing the update through a user turn. This makes it substantially easier to build sophisticated, adaptive agent harnesses.
Opus 4.8 is Anthropic's flagship model, positioned for work where quality is the primary constraint and cost is secondary. It's the right choice when you're building production-grade AI agents, handling high-stakes professional knowledge work, or need a model that can sustain coherent context and judgment across very long sessions.
On agentic benchmarks, Opus 4.8 outperforms GPT-5.5 in specific evaluations: one external partner's Super-Agent benchmark showed Opus 4.8 completing every case that GPT-5.5 could not, at cost parity. On computer use (Online-Mind2Web), Opus 4.8's 84% score beats GPT-5.5's reported result on the same evaluation. Comparative performance varies by task type; users with specific workloads should run their own evaluations.
The headline improvements are better honesty (Opus 4.8 flags uncertainties and code flaws at a significantly higher rate), improved judgment in autonomous tasks, more efficient tool calling, and better alignment scores. Verbose comment generation and tool-calling inconsistencies reported with Opus 4.7 are addressed in this release.
Dynamic workflows let Claude plan a large software task and then execute it by spinning up hundreds of parallel subagents within a single Claude Code session. It verifies its outputs before surfacing results. It's currently in research preview and available on Enterprise, Team, and Max plans.
Opus 4.8 can reason over PDFs, diagrams, charts, and other unstructured visual content. For document-heavy workflows, it delivers this at a 61% lower token cost compared to Opus 4.7, according to one enterprise data platform's internal benchmarks.
-min-p-130x130q80.png)
.webp)
.svg)
.png){kind=small}
