Claude Opus 5 Ships at Opus 4.8's Price, With Looser Security Guardrails

What happened

Anthropic launched Claude Opus 5 on July 24. It costs the same as its predecessor, Opus 4.8: $5 per million input tokens and $25 per million output tokens. A token is roughly a word fragment, so this is the price per rough word-count of text you send in and get back. Anthropic also offers a faster “Fast mode” for Opus 5, at double the base price, that runs about 2.5 times quicker.

Opus 5 is now the default model on Claude Max (Anthropic’s higher-usage subscription tier) and the strongest model available on Claude Pro. It’s live across Claude.ai, Claude Code, Claude Cowork, and the Claude API (the option developers use to call Claude directly from their own app’s code, instead of through the chat interface).

The same day, GitHub Copilot added Opus 5 to its model picker, rolling out gradually to Copilot Pro+, Max, Business, and Enterprise users in VS Code, Visual Studio, the Copilot CLI, the Copilot cloud agent, the GitHub Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Copilot bills Opus 5 usage at Anthropic’s list price under its own usage-based billing, meaning you pay based on how many requests and tokens you actually use rather than a flat add-on fee. Business and Enterprise plan admins have to turn on the “Claude Opus 5” policy in Copilot’s admin settings before anyone on their team can pick the model.

Anthropic reports Opus 5 as its new best coding and agentic model, ahead of Opus 4.8 on the benchmarks it publishes, and close to Claude Fable 5 (Anthropic’s most capable model) on several of them at a fraction of Fable 5’s cost. Devin’s CEO is quoted saying Opus 5 “approaches Fable-level performance at half the cost,” and Zapier’s CEO says it hit a 100% pass rate on Zapier’s own internal agentic test, AutomationBench, something no earlier model managed. These are vendor and partner claims, not independent benchmarks, so treat them as a starting point rather than a verdict on your own workload.

The guardrail change worth knowing about

Opus 5 ships with new automated safety checks, called classifiers, that scan requests for high-harm cyber content, like requests that resemble building malware or exploiting a vulnerability, and can block them. Anthropic says these classifiers are tuned to be less restrictive on Opus 5 than the equivalent ones on Fable 5, intervening roughly 85% less often.

That’s a real behavior change if you write security tools, scripts that test for vulnerabilities, or anything that looks like it’s probing for an exploit, even for entirely legitimate work. GitHub Copilot’s own notes confirm the classifiers can still block cyber-related or security-adjacent requests, and suggest rephrasing with more benign context or switching models if that happens.

There’s a second, quieter change: automatic fallback. In Claude.ai, Claude Code, and Claude Cowork, a request that gets flagged by these classifiers now falls back to Opus 4.8 by default, and answers instead of simply refusing. On the API, you can turn on the same behavior so a blocked request reroutes to another model instead of failing outright. Either way, the model that actually answered your prompt may not be the one you picked. For anything where the specific model’s judgment matters, like reviewing your own security code, that’s worth knowing before you trust the output.

Who should care

If you already pay for Claude Max or Claude Pro, you’re using Opus 5 by default or can switch to it now, at no added cost over what Opus 4.8 charged. If you use Claude through the API for an app you’re building, the price is unchanged, so this is a capability upgrade at the same spend. If you use GitHub Copilot on a paid individual or team plan, check your model picker, since the rollout is gradual and Business or Enterprise admins need to flip a setting first.

This is the same pattern BuilderWithin covered three days ago when Google shipped Gemini 3.6 Flash into GitHub Copilot the same day it launched: a model vendor ships a new model, and it shows up in a familiar coding tool on day one. If you already work inside Copilot, you don’t need to change tools to try Opus 5.

What builders should do next

Don’t take Anthropic’s benchmark numbers as your answer. Pick one coding task you already run regularly with a known correct outcome, like a bug fix you can check against your existing tests. Run it once with your current default model and once with Opus 5, in Claude Code or Copilot, using the identical prompt and starting point both times.

Compare four things: did the result actually pass your tests, how many follow-up corrections did you need before it was usable, how long each run took, and how many tokens or how much of your usage allowance it consumed. That gives you a real answer for your own work instead of borrowing Anthropic’s.

If your task touches security code, security review, or anything that could read as exploit-adjacent, check which model actually produced the response after the fact. Claude Code and the Claude API both let you confirm the model in play. If a request that reached Opus 5 silently fell back to Opus 4.8, you’re reviewing a different model’s judgment than the one you intended, and that’s worth knowing before you ship.

End of article