Claude Opus 5.5 Lowers Prices and Adds Safety Routing
Matthew Leo · September 23, 2026 · AI
Anthropic released Claude Opus 5.5 on September 22 with lower prices, faster output and a more visible set of restrictions around sensitive work.
The model costs US$4 per million input tokens and US$20 per million output tokens, down from US$5 and US$25 for Opus 5. Cache reads drop from US$0.50 to US$0.20 per million tokens. Anthropic says typical workloads cost about 40 per cent less because the model also uses fewer tokens to finish a task.
Those are provider prices in US dollars. Canadian customers will pay the converted amount plus any taxes or platform markup, and subscription plans use separate limits rather than direct token billing.
What changed
Anthropic says Opus 5.5 produces output more than 30 per cent faster than Opus 5 at its normal setting. A faster mode is available through Claude Code and the Claude Platform at double the standard token price.
The company positions the model mainly for long coding jobs, computer use and knowledge work. Its published benchmarks show gains over Opus 5, but most of those numbers were produced or commissioned by Anthropic and early partners. They are useful for understanding the target, not a substitute for independent testing on your own workload.
GitHub has also started a gradual rollout in Copilot for Pro+, Max, Business and Enterprise users. It is available through the model picker in tools including Visual Studio Code, Copilot CLI, the coding agent and GitHub Mobile. Business administrators can disable it through model policy settings.
How the safety routing works
The release is unusual because the model you select may not complete every sensitive task itself. Anthropic says its production safeguards can route restricted cybersecurity work to Claude Opus 4.8, while some biology and frontier-model work may be handled by Opus 5.
That means a benchmark labelled Opus 5.5 can include work completed by a fallback model when a safeguard triggers. Anthropic discloses this in its evaluation notes, and says one business-workflow benchmark counted those interventions as failures because it was run without fallbacks.
The distinction matters for professional users. If you work in security or life sciences, the model’s answer quality and behaviour can change when a request crosses a policy boundary. Anthropic offers verification programs for approved organizations, but ordinary users should not assume every permitted task will run on Opus 5.5 from start to finish.
What the price cut does not prove
A cheaper token price does not guarantee a cheaper result. Total cost depends on prompt length, output length, tool calls, retries and whether an agent finishes without human correction. Anthropic’s own claim is based on typical workloads, so teams should compare complete tasks rather than token rates alone.
The same caution applies to safety claims. Anthropic says Opus 5.5 performs better on its behavioural audit and resists prompt injection more effectively than Opus 5. The company used outside evaluators before release, but the detailed claims still come from the developer and its chosen testing process.
That is why the routing disclosure is useful. It tells customers more about how the deployed service behaves, even if it makes model comparisons less tidy. It also connects to a question Mapletechie raised in its look at Anthropic’s proposed safety gates: controls are easier to judge when users can see when they change the system’s behaviour.
For most buyers, the next step is simple. Run the same real task on Opus 5 and Opus 5.5, record total tokens, time, corrections and any safeguard handoff, then compare the completed-work cost.
Sources
- Anthropic: Claude Opus 5.5, September 22, 2026
- GitHub: Claude Opus 5.5 in Copilot, September 22, 2026
- Reuters: Anthropic unveils Claude Opus 5.5, September 22, 2026
Tags: Claude, Anthropic, AI Models, GitHub Copilot, AI Safety