Claude Opus 5 release announcement
Official release page with benchmark framing, effort settings, availability, and Anthropic’s performance and cost claims.
← Back to Compute College
Compute College
Claude Opus 5 benchmark explained
Read Claude Opus 5 benchmark claims through AI compute economics: task quality, effort settings, token pricing, fast mode, and inference demand.
Plain-English definition
Claude Opus 5 is Anthropic’s July 24, 2026 Opus release. Anthropic says it improves on Opus 4.8 across coding, knowledge-work, science, and agent evaluations while keeping standard API pricing at $5 per million input tokens and $25 per million output tokens. For compute buyers, the useful question is whether higher task completion changes cost per accepted result, token use, latency, or the amount of premium inference capacity a workload needs.
Memory trick: The price tag is per token; the economic result is per accepted task.
Why it matters
A model can improve the economics of a workload without becoming cheaper per token. If Opus 5 completes more complex tasks with fewer retries, tool calls, or review cycles, its cost per accepted outcome can fall at the same listed price. But if better capability prompts teams to run longer agents, larger contexts, or more autonomous workflows, total inference demand can rise even when the unit rate is unchanged.
- Same listed token rates make the release a quality-per-dollar question rather than a simple list-price reduction.
- Effort settings can trade more tokens and latency for greater task reliability, so a headline benchmark needs its configuration.
- Long-running coding and research agents can turn better performance into more serving hours and capacity demand.
Simple example
At Anthropic’s listed standard rate, a request using 100,000 input tokens and 20,000 output tokens costs $0.50 for input plus $0.50 for output, or $1.00 before caching and platform differences. Fast mode is listed at twice the standard price, so the same token mix would cost $2.00 while aiming for lower latency. The cheaper choice depends on whether faster completion avoids enough waiting, retries, or downstream work.
- The arithmetic uses current published standard and fast-mode rates; it is not a quote or workload forecast.
- Output tokens can equal input spend despite being far fewer, so log both sides of the token mix.
- Compare accepted task outcomes, latency, retries, and tool use before treating a faster mode as economically better.
Example figures are illustrative calculations, not current quoted market prices.
Current example
What Anthropic published
Anthropic announced Claude Opus 5 on July 24, 2026. Its release page describes higher performance than Opus 4.8 across selected evaluations and says Opus 5 keeps the same standard token price. Anthropic also publishes fast mode at twice the standard price. Those are first-party release and pricing claims, not independent Compute College benchmarks.
Claude API pricing
Official pricing reference for standard usage, caching, and fast-mode rates.
Historical: Claude Opus 4.8 benchmark explained
Compare the prior release without treating it as the current model example.
Source discipline: evaluate Anthropic benchmark, customer, and safety claims as first-party evidence. Measure your own workload before making a routing, procurement, or capacity decision. Last checked: Sep 16, 2026.
Common mistake
A better benchmark score does not prove a lower production bill. A model may need more effort, longer contexts, different tools, or a different agent harness, and each can change latency, token use, and cost per completed task.
Practical takeaway
What you can do with this
Run a recorded comparison on representative production tasks. For each candidate model and effort setting, capture input and output tokens, cache behavior, tool calls, retries, latency, human review, and whether the task was accepted. Use the official price page for the cost calculation, then compare accepted outcomes per dollar and within the required latency budget.
- Buyers: require a workload-level result before moving meaningful traffic to a premium model.
- Developers: log effort and mode because those settings can materially change both performance and cost.
- Analysts: distinguish early benchmark evidence from observable adoption and capacity-demand evidence.
Decision check: does Opus 5 improve the completed outcome your team needs enough to offset its token use, latency, and operating requirements under the chosen effort and speed mode?
Compute College
Follow model releases as market signals
Follow model releases as AI compute market signals in the ComputeTape Market Brief.
Compute College track
Model Benchmarks & AI Compute Economics
Step 15 of 25: Claude opus 5 benchmark explained