Claude 4.8 Opus vs Claude 4.6 Opus

1. Modest but tangible gains across the board

• 69.2% on SWE-bench Pro (+4.9 points over Opus 4.7's 64.3%), 88.6% on SWE-bench Verified, and a 1,890 Elo on GDPval-AA - about 121 Elo ahead of GPT-5.5.
• Anthropic frames it as an incremental refinement of Opus 4.7 rather than a step-change release.

2. The most reliable agent model Anthropic has shipped

• The only model to complete every case end-to-end on Anthropic's internal Super-Agent benchmark, beating prior Opus models and GPT-5.5 at parity on cost.
• Leads CursorBench across every effort level with more efficient tool calling - fewer steps for the same intelligence.
• 84% on Online-Mind2Web and stronger OSWorld-Verified scores make it the strongest computer-use and browser-agent model Anthropic has tested.

3. New highs for professional and legal work

• Highest score ever recorded on Anthropic's Legal Agent Benchmark, and the first model to break 10% on its all-pass standard.
• Roughly a 4x improvement in honesty and calibrated uncertainty versus Opus 4.7 on Anthropic's evaluations.

4. Cheaper, faster controls

• Standard pricing holds at $5/M input and $25/M output, unchanged from Opus 4.7.
• Fast mode now runs at 2.5x speed for $10/M input and $50/M output - about a third of what fast mode cost on prior Opus releases.
• A new effort dial gives finer control over how hard the model works before answering.

5. Long-horizon agent workflows

• 1M token context window in beta and up to 128K output tokens.
• Introduces Dynamic Workflows (research preview) in Claude Code - Opus can write its own orchestration script and run up to 16 concurrent / 1,000 total subagents in a single session.

1. Anthropic's top model for coding and agents

• Anthropic positions Opus 4.6 as its most intelligent model for building agents and coding.
• It builds on Opus 4.5 with higher reliability and precision for professional software engineering, complex agentic workflows, and high-stakes enterprise tasks.

2. Strong frontier performance on real agent benchmarks

• Anthropic reports state-of-the-art results across coding and agentic evaluations.
• Public benchmark highlights include 65.4% on Terminal-Bench 2.0, 72.7% on OSWorld, and 90.2% on BigLaw Bench.

3. Best fit for long-horizon, high-context work

• Supports up to a 1M token context window in beta and up to 128K output tokens.
• Designed for long-running tasks that need sustained planning, careful debugging, code review, and strong context retention.

4. Advanced reasoning controls and workflow support

• Supports adaptive thinking and the effort parameter, including the new max effort level.
• Anthropic also introduced fast mode, compaction, and dynamic filtering with web search and web fetch for Opus 4.6-era agent workflows.