Claude Opus 5: Anthropic's Cheaper Everyday Flagship
Source: Anthropic — 2026-07-24
Summary
Anthropic launched Claude Opus 5, an "everyday" flagship model that lands near the top of its own model lineup on Frontier-Bench and GDPval-AA while staying at Opus 4.8's existing price of $5/$25 per million input/output tokens — half of what Anthropic's higher-end model costs. It pairs a per-request "effort dial" (low/medium/high reasoning budget) with an optional 2.5x-speed Fast mode, and becomes the new default model on Claude Max and the strongest option on Claude Pro. It's Anthropic's fourth Claude 5-series release in under two months, reinforcing a shift toward frequent, incremental capability/cost releases rather than infrequent blockbuster launches.
Key Takeaways
- Pricing holds flat at $5 per million input tokens / $25 per million output tokens — unchanged from Opus 4.8, but delivering roughly double the Frontier-Bench v0.1 score at that same price point.
- The "effort dial" lets developers cap extended-thinking token spend per request (low/medium/high), since reasoning tokens are billed as output — cheap for routine tasks like lint fixes, dialed up for hard problems.
- Fast mode trades higher per-token cost for up to 2.5x lower latency, aimed at interactive workflows like live debugging or rapid iteration; available on both Opus 5 and Opus 4.8.
- On OSWorld 2.0 (the computer-use benchmark where the model operates a mouse/keyboard/software directly), Opus 5 beats every model at any price point, matching the top score from Anthropic's higher-end model at roughly a third of the cost.
- Opus 5 becomes the new default on Claude Max and the top available model on Claude Pro, immediately available across Anthropic's platforms.
Reel Script
Hook
Anthropic just shipped a model that beats their own flagship on benchmarks and costs half as much. That's not a typo — it's a pricing strategy, and it tells you exactly where AI margins are headed next.
Core Concept
Here's the mechanism. Opus 5 ships with extended thinking on by default — meaning before it answers, it generates internal reasoning tokens, and those get billed just like output. That's normally expensive. So Anthropic added an "effort dial": low, medium, or high, set per request. Low caps the reasoning budget hard, so a quick lint fix or a one-line summary barely spends any thinking tokens. High lets it reason as long as it needs on a gnarly problem. Think of it like a car's eco mode versus sport mode — same engine, but you choose how much fuel it's allowed to burn before it gives you an answer. That single dial is what lets one model serve both "cheap and fast" and "slow and maximally capable" use cases without forcing you to switch models entirely.
Hands-On
The numbers are the real story here. Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — exactly what Opus 4.8 cost — but on Frontier-Bench v0.1 it's roughly doubled that predecessor's score. On GDPval-AA, a knowledge-work benchmark, and on OSWorld 2.0, the computer-use benchmark where the model literally drives a mouse and keyboard through real software, Opus 5 beats every model at every price point Anthropic tested, matching the peak score of its own higher-end model while using roughly a third of the budget to get there. Then there's Fast mode, a separate toggle that trades a higher per-token price for up to two-and-a-half times lower latency — you flip it on for live debugging sessions where waiting five seconds per turn kills your flow, and flip it off for background batch jobs where cost matters more than speed. And it's now the default model on Claude Max, so most subscribers get this upgrade without changing a setting.
Takeaway
This is Anthropic's fourth Claude 5-series release in under two months, and the pattern is clear: they're optimizing cost-per-capability, not just raw capability. If you're running agentic workloads at any real volume, this is the release to benchmark against your own tasks this week — the effort dial alone could meaningfully cut your reasoning-token bill without touching output quality.
Discussion
(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)