Claude Opus 5: What Anthropic's Benchmarks Show
Claude Opus 5 was released on July 24, 2026. NanoGPT lists it at the same standard input and output rates as Opus 4.8. Anthropic's launch results emphasize software engineering, computer use, and professional knowledge work.
The case for switching depends on the work. Anthropic reports stronger results on several demanding evaluations, but a higher reasoning effort can still make an individual request slower and more expensive. Teams should compare Opus 5 and Opus 4.8 on representative tasks rather than assume every prompt will improve.
The change from Opus 4.8
Anthropic describes Opus 5 as more proactive and more likely to check its work before stopping. Its examples include tracing a package-manager bug to an underlying edge case rather than repairing only the visible symptom, and building an image-processing tool when a coding task did not provide direct access to its image input.
These are selected launch anecdotes, not controlled evaluations. They show the sort of multi-step behavior Anthropic wants to highlight, but they do not establish how frequently Opus 5 will act this way.
On NanoGPT, Opus 5 supports:
- Text, image, and PDF input
- Tool calling, so it can use search, code, and other connected functions
- Structured output for applications that need defined fields
- Adaptive reasoning with
low,medium,high,xhigh, andmaxeffort - A one-million-token context window, with outputs capped at 128,000 tokens
- Prompt caching for repeated instructions or reference material
NanoGPT uses high reasoning effort by default. Lower settings can reduce latency and token use on direct tasks. max gives the model a larger reasoning budget and is intended for work where lower settings miss a specific requirement.
What Anthropic's benchmarks report
Anthropic presents most Opus 5 results as charts that compare benchmark scores with estimated task cost across several effort settings. Every row below comes from Anthropic's launch post or a launch partner. We had not found substantial independent testing as of July 28, 2026.
| Evaluation | Reported launch result | What it tests |
|---|---|---|
| Frontier-Bench v0.1 | Anthropic's chart shows more than twice Opus 4.8's score at a lower reported cost per task | Completing difficult software work with tools |
| CursorBench 3.2 | Within 0.5% of Claude Fable 5's peak score at max effort, at about half the reported task cost | Working on an existing codebase inside an editor |
| ARC-AGI 3 | About three times the next-highest score among the plotted models and settings | Adapting to unfamiliar visual and logical problems |
| GDPval-AA v2 | Highest score among the models in Anthropic's comparison | Completing professional knowledge-work tasks |
| Zapier AutomationBench | About 1.5 times the next-highest pass rate at a similar reported task cost in Zapier's test | Completing business workflows from start to finish |
| OSWorld 2.0 | A higher score than Fable 5's best plotted result at just over one third of its reported cost | Operating desktop applications through visual interaction |
Frontier-Bench and CursorBench are designed to resemble repository-level development rather than short code completion. The model inspects context, edits files, uses tools, and may need to revise an initial approach.
OSWorld tests whether a model can operate graphical applications instead of describing the required clicks. Anthropic's chart shows Opus 5 ahead at each plotted cost point, although the comparison covers selected models and configurations rather than the whole market.
Anthropic also reports improvements over Opus 4.8 on every life-sciences evaluation shown in the launch post. It highlights a 10.2 percentage-point gain on an internal spectroscopy task and a 7.7 percentage-point gain on predicting how changes to a protein sequence affect its function. These internal results are vendor-reported indications that still need independent replication.
Why effort settings matter
An Opus 5 request does not have one fixed cost. The same prompt at low uses a smaller reasoning budget than at max, although the model can use less than the available allowance.
low and medium are worth testing for editing, analysis with clear instructions, and routine code changes. high provides more room for evidence review, tool use, and constraint checking. xhigh and max may improve difficult repository or research tasks, but they can also increase latency and output-token use without changing the final result.
Anthropic's launch charts show competitive results at lower effort levels. Zapier reports that even the lowest setting passed more AutomationBench tasks than any other model included in its comparison. This is a partner-reported launch result, not independent replication.
Keep a small set of representative tasks and compare pass rate, latency, and token use. Raise effort when a lower setting fails a specific check, not merely because the task is important.
Reading the benchmark claims carefully
The launch announcement combines Anthropic-run evaluations with results from selected partners. Some benchmarks use internal test systems, and the model's effort setting changes both score and cost. Frontier-Bench is also version 0.1, so its tasks and test setup may still change. In the reported setup, requests blocked by Opus 5's safety filter could fall back to Opus 4.8. The headline Frontier-Bench result was therefore not produced by Opus 5 alone in every attempt.
The charts' "cost per task" figures are benchmark-specific estimates, not NanoGPT request prices. Depending on the evaluation, retries, successful completions, tool calls, and reasoning-token use can change that number. Opus 5 and Opus 4.8 share the same standard token rates on NanoGPT, but two requests can still have different total costs.
We did not run a separate NanoGPT benchmark for this article. Anthropic's results justify testing Opus 5, but they do not establish that it will improve every existing Opus prompt.
Price and context on NanoGPT
NanoGPT lists Opus 5 at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8. Prompt-cache reads cost $0.50 per million cached tokens; creating a cache costs more than ordinary input, so caching is useful when the same prefix will be reused.
The one-million-token context window can hold large repositories, document collections, and extended agent histories. Supplying irrelevant files raises cost and gives the model more material to inspect. A focused set of source files, requirements, logs, and tests can work better than uploading an entire workspace.
The anthropic/claude-opus-latest alias now points to Opus 5. Use anthropic/claude-opus-5 when the application should remain on this model family; the latest alias may move to a future release.
Work worth testing with Opus 5
The launch results make Opus 5 a candidate for tasks where planning, tool use, and revision determine whether the result passes review.
- Repository-wide changes that require planning, implementation, tests, and revision
- Debugging where the visible failure may not be the root cause
- Source-based analysis where every claim can be checked by a human reviewer
- Computer-use tasks that cross several applications or screens
- Long reports, presentations, and other artifacts that need iterative refinement
- Agent workflows where failed attempts and repeated tool calls are expensive
For short summaries, first drafts, classification, and bulk transformations, compare a faster or cheaper model on the same quality checks. Opus 5 is easier to justify when its measured pass rate offsets the extra latency or token use.
