Claude Opus 4.5: Launch Benchmarks, Pricing and Tool Use

Anthropic has released Claude Opus 4.5, shortly after other frontier model launches. In this November 2025 video, I review the company’s reported benchmarks, prices and developer features. The video also contains a paid sponsorship for Warp.

Anthropic’s launch chart reports 80.9% for Opus 4.5 on SWE-bench Verified, compared with 77.2% for Sonnet 4.5 and other displayed model scores. The vertical axis starts near 70 rather than zero, making the bars look farther apart. The difference from Sonnet is 3.7 percentage points, and the result applies to the reported 500-task setup, not all coding work.

Berman reviews other reported agent and computer-use results, including Terminal-Bench 2.0, τ²-bench and OSWorld. He also identifies tests where competitors scored higher, such as graduate-level reasoning and visual or multilingual tasks. Results reflect specific harnesses, configurations and datasets; missing competitor figures do not establish their performance.

He compares a vending-machine simulation leaderboard and ARC-AGI scores, noting that the displayed Opus result did not lead each chart. These are historical snapshots with their own scoring rules, and neither benchmark alone measures every aspect of intelligence or workplace performance.

Anthropic’s launch price was $5 per million input tokens and $25 per million output tokens. Berman compares it with the contemporary Gemini 3 Pro price tiers. A real bill depends on the input-output ratio, caching, prompt length and applicable prices, so one percentage difference is not universal.

Berman says Anthropic tested Opus 4.5 on a difficult performance-engineering hiring exercise and describes it as better than every hired candidate. In a later, more detailed account, Anthropic said a prerelease model matched the best human result under a two-hour limit; the human had used Claude with steering, and stronger human results were possible without the time limit. This is a task-specific comparison, not proof of superior engineering ability in general.

The video includes a paid promotion for Warp, an AI coding product. Berman describes its claimed benchmark placement, interface and support for parallel agents and tools. Those are promotional claims from the sponsor segment, not an independent evaluation in this video.

Berman describes Anthropic’s τ²-bench example: an airline-service task expected refusal of a basic-economy change, while the model proposed upgrading the cabin first and then changing the flight. Anthropic characterized this as a legitimate alternative that the benchmark marked wrong. Berman questions whether it was actually optimal; authorization, cost and the scenario’s rules matter.

Anthropic also introduced advanced tool-use features. Berman explains that loading definitions for many MCP tools up front can consume context before the user’s request. A tool-search mechanism can retrieve relevant definitions when needed, reducing upfront overhead; it does not give the model literally infinite tools or an unlimited context window.

He outlines three features: tool search, programmatic tool calling in a code-execution environment, and examples showing how to use a tool. Programmatic calling can keep intermediate results out of the model’s context, while examples can clarify usage that a schema alone may not convey.

Anthropic’s example estimates that definitions for several MCP servers can occupy tens of thousands of tokens. Berman contrasts a conventional approach with on-demand discovery and reports a large context reduction in the illustrated setup. Actual savings depend on the tools, schemas and requests.

Berman then discusses an Anthropic chart comparing SWE-bench results with output-token use. He observes that Opus 4.5 achieves a higher score at fewer output tokens than Sonnet 4.5 in that example. The plotted settings are specific, and output tokens alone are not the complete cost or a universal measure of intelligence.

Finally, he cites early reactions from Dan Shipper and Ethan Mollick, who praised coding and practical work. These are user testimonials rather than controlled independent proof that the model is best for every application.