Anthropic launched Claude Opus 5 on July 24 with pricing that creates an unusual gap inside its own model family. Claude Opus 5 pricing starts below Fable 5 despite offering a 1M context window, adjustable effort levels, Fast mode and stronger results on several demanding evaluations.

Claude Opus 5 pricing remains unchanged

Anthropic has priced Claude Opus 5 at $5 per million input tokens and $25 per million output tokens in standard mode. Those rates match Claude Opus 4.8, the model it replaces as the company’s flagship Opus offering.

The company also offers Fast mode for developers who need quicker responses. Anthropic says this setting runs about 2.5 times faster than standard mode and costs $10 per million input tokens and $50 per million output tokens. Fast mode remains a research preview and is currently limited to Anthropic’s first-party API.

The pricing creates a notable contrast with Claude Fable 5, which costs $10 per million input tokens and $50 per million output tokens in its normal mode. As a result, Opus 5’s standard rate is half the price of Fable 5, even though both models sit within Anthropic’s current-generation lineup. Anthropic positions Fable 5 for highly demanding, long-running work with stricter safeguards, while Opus 5 serves as its strongest generally available model.

Opus 5 features expand developer control

Claude Opus 5 supports a one-million-token context window and can produce outputs of up to 128,000 tokens. This allows developers to process large codebases, extensive research files and long-running agent workflows without dividing the material into as many separate requests.

Anthropic has also made effort level the main control for steering the model. Developers can choose low, medium, high, xhigh or max effort depending on the complexity of the task. Lower settings can reduce token use and latency for routine work, while higher levels allow the model to spend more computing resources on difficult reasoning, coding or analytical problems.

This approach gives teams more flexibility within one model. Instead of moving between several model tiers, they can adjust how much effort Opus 5 applies to each request.

Frontier-Bench result highlights reasoning gains

Anthropic reported that Opus 5 reached 43.3% on Frontier-Bench v0.1, a benchmark designed to test advanced agentic coding and reasoning. However, the result should be described carefully: Anthropic said it came from an internal run using the mini-SWE-agent harness, five attempts per task and Opus 4.8 as a fallback for some safety-related refusals.

The score therefore offers useful evidence of progress, but buyers should not treat one benchmark as a complete measure of real-world performance. Reliability, latency, output quality, tool use and total task cost can matter more than headline scores when companies deploy models at scale.

AI pricing war puts pressure on Anthropic

Opus 5 enters a market where competitors increasingly offer several intelligence and cost tiers. OpenAI’s GPT-5.6 family ranges from the cost-focused Luna model at $1 per million input tokens to the flagship Sol model at $5 per million input tokens. Their output prices range from $6 to $30 per million tokens.

Moonshot AI’s Kimi K3 also targets long-context and complex agent workloads, with public API listings showing lower cached-input pricing alongside an open-weight release. These alternatives give developers more leverage to compare quality against the full cost of completing a task, rather than comparing token prices alone.

Anthropic is clearly betting that stronger judgment, coding reliability and long-horizon reasoning will justify Opus 5’s premium over cheaper models. Its challenge is proving that customers can complete difficult work with fewer retries, fewer tool calls and less supervision. The launch strengthens Claude’s technical offering, but it also makes Anthropic’s own pricing hierarchy more complicated.