Anthropic Shipped Claude Opus 5 at the Same Price as the Old One. It Doubled the One Benchmark That Actually Predicts Whether an Agent Ships Working Code.
On July 24, Anthropic released Claude Opus 5 at the exact same price as the model it replaced: $5 per million input tokens, $25 per million output tokens. Frontier-Bench v0.1 (the benchmark that tracks closest to "did the agent ship working code") jumped from Opus 4.8's 18.7% to 43.3%. More than double. Ahead of the larger, pricier Fable 5's 33.7% on the same test. Most model upgrades in 2026 have asked you to pay more, accept slower responses, or trade capability for cost. This one didn't ask for any of that, which is rare enough that it's worth asking why, and worth figuring out whether to actually switch.
What shipped, specifically
Opus 5 is live across the Claude API, Claude Code, and Claude.ai as of July 24. Pricing is unchanged from Opus 4.8: $5/$25 per million tokens. There's also a Fast mode running at roughly 2.5x the default speed for 2x the base price, which matters if your workload is latency-sensitive rather than cost-sensitive: an interactive coding session behaves differently than a batch job, and now you can pick.
The benchmark numbers are the headline. On Frontier-Bench v0.1, Opus 5 scored 43.3%, versus 18.7% for Opus 4.8 and 33.7% for Fable 5, Anthropic's larger and more expensive model. On GDPval-AA v2, a knowledge-work evaluation, Opus 5 beat Fable 5 outright: 1861 versus 1747. On ARC-AGI-3, a test of novel problem-solving rather than pattern-matching, Anthropic reports Opus 5 scoring roughly three times the next-best model. Opus 5 still trails Mythos 5 specifically on cybersecurity-focused tasks, so this isn't a clean sweep across every category, just most of the ones a solo operator actually cares about.
Two beta features shipped alongside the model release, and they're easy to miss under the benchmark chart. Mid-conversation tool changes let you swap which tools are available to Claude partway through a session without invalidating the prompt cache. Previously, changing the toolset meant paying to rebuild context from scratch. Automatic fallbacks let a request that gets blocked by a safety classifier reroute to a different available model instead of just failing. Neither is flashy, but both save real money if you run production agents that occasionally hit either wall.
Why "same price, higher score" is the exception
Model releases over the past year have mostly followed a pattern: capability goes up, and either the price goes up with it, or the new frontier model is positioned as a separate, pricier tier while the "cheap" tier stays flat or gets marginally better. Fable 5 launched at $10/$50 per million tokens (double Opus 5's rate) and went API-only in early July after its subscription access window closed. If you wanted frontier-adjacent performance without the frontier price, your options were limited.
Opus 5 breaks that pattern by sitting at the same price point Opus 4.8 held and simply being better at the thing most solo operators actually run models for: agentic coding, computer use, structured knowledge work. That's not typical, and it's worth being clear-eyed about why it happened: Anthropic is under real competitive pressure from Chinese open-weight models undercutting on price and from OpenAI and Google shipping on a similar six-to-eight-week cadence. A "free" capability jump at the same price point is also a signal that the mid-tier is where the competitive fight is happening right now, not a permanent state of affairs.
The actual decision: do you switch this week
If you're currently defaulting Claude Code or an API integration to Opus 4.8, there's very little reason not to move to Opus 5 immediately: same price, meaningfully better benchmark performance, no migration cost beyond changing a model string. That's close to a free upgrade.
If you're currently on Fable 5 because you needed the extra capability, the calculation is different. Fable 5 still likely wins on raw ceiling for the hardest problems, and it's your call whether the gap between Fable 5's frontier performance and Opus 5's 33.7%-beating 43.3% justifies paying double per token. For most solo operators running agentic coding workflows (not frontier research, not novel-problem benchmarks, just "write this feature and don't break the test suite") I'd bet Opus 5 closes enough of that gap that the 2x price premium on Fable 5 stops making sense for day-to-day work. Reserve Fable 5, if you keep it at all, for the specific tasks where you've actually observed Opus 5 falling short.
The mid-conversation tool-switching beta is worth turning on if you run any agent that dynamically changes its available toolset: a coding agent that sometimes needs browser access, say, or a research agent that toggles between search and code execution. Before this, that kind of switch meant eating a cache invalidation and re-paying for context. Now it doesn't. If your current agent architecture works around that limitation with some awkward design (holding onto a superset of tools it doesn't always need, for instance), this is the moment to simplify it.
The honest counter-take
Benchmark deltas are directional, not guarantees. Frontier-Bench and GDPval-AA are Anthropic-favorable evaluations by construction, and a 43.3% score on a benchmark doesn't tell you how Opus 5 handles your specific messy, half-documented, three-year-old codebase with its own conventions. I'd run your own before-and-after comparison on a real task from your actual backlog before committing a production workflow to the switch, rather than trusting the published numbers alone.
It's also worth remembering that "same price, better benchmark" is the exception precisely because it doesn't hold every time. The next release could easily go the other way: better model, higher price, and a fresh round of the token-math exercise. Don't build a permanent assumption that Anthropic's pricing stays flat as capability rises; build the habit of running that 15-minute comparison every time a new model drops, because this release won't be the last one where the answer matters.
Author
Lukas
@lukcombinator