← The Polylog AI Intelligence Brief
Morning Edition · Monday, September 28, 2026Published at 2:48 AM EDT · New York
Anthropic Prices Claude Opus 5.5 at 40 Percent Below Opus 5 Running Cost
Input tokens drop from $5 to $4 per million and output from $25 to $20, with cached reads cut 60 percent to $0.20, while Anthropic claims performance level with its larger Fable 5.1 model on most work.
Anthropic's Claude Opus 5.5 is the clearest current example of a frontier lab selling efficiency rather than a new capability ceiling. The company says the model performs at the level of Claude Fable 5.1, released on September 1, on most work, while costing 40 percent less to run than Opus 5 on typical workloads.
The published price moves are concrete. Input falls from $5 to $4 per million tokens and output from $25 to $20, with cached reads at $0.20 per million, a 60 percent reduction against Opus 5, and output generation more than 30 percent faster, according to reporting on the launch. Subscription users get roughly 25 percent more usage headroom, per Memeburn. The model is available on the Claude platform and through Amazon Web Services, Google Cloud and Microsoft Azure.
Two claims deserve separation. The prices are verifiable and immediately actionable for anyone running production traffic. The performance parity claim with Fable 5.1 is Anthropic's own characterization of "most work," not a third-party leaderboard result, and the company's statement that Opus 5.5 is the strongest model it has tested on its automated behavioral audit is also an internal measurement. Cached-read pricing is the line that matters most for agent workloads, where long system prompts and tool schemas are re-sent on every step, and a 60 percent cut there changes the unit economics of long-running agents more than the headline output price does.
What this means
The mechanism is capability per dollar rather than capability per parameter. Anthropic is pushing its most expensive tier down toward the price band where agent loops, which consume tokens in proportion to step count rather than to user messages, stop being the dominant line item in a software budget. Who gains is any vendor whose product wraps many model calls per user action, including coding agents and customer support automation, and who loses is the mid-tier model segment, since a cheaper Opus compresses the reason to route to a smaller model. The precedent is the 2024 to 2025 sequence in which prompt caching and batch pricing arrived first as discounts and then became assumed defaults in application architectures. The likelier outcome over the next few months is that competitors respond on cached-input and long-context pricing specifically rather than on headline output rates, because that is where agent margin now sits. Independent evaluation showing Opus 5.5 trailing Fable 5.1 on agentic or long-horizon coding tasks, rather than matching it, would undercut the parity claim and slow the migration.
Part of a tracked trend
Frontier Model Efficiency Gains
Capability per unit of training and inference compute keeps improving, letting newer models match prior frontier performance far more cheaply and gradually loosening the link between raw scale and capability.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Amodei Gets a Private White House Dinner Weeks Before Anthropic's Planned October Listing
- Stanford Runs a 37,000-Agent Virtual Biotech and Reports a Target Hypothesis That Matched a Real Therapy
- Meta Sells Muse Spark 1.3 at $1.25 per Million Input Tokens Through Its Own Metered API
- Rails Creator Says His Team Stopped Writing Code by Hand, and Anthropic Stops Cutting Agents Off Mid-Task
- Anthropic's 2030 Economic Model Puts United States Output Between $34.1 Trillion and $44.4 Trillion
- Berkeley Lab Cuts Beamline Annotation From a Month to 15 Minutes Using Meta's Open Vision Models
- New Benchmark Tests Whether Penetration-Testing Agents Stay Inside the Client's Authorized Scope
Comments
0No comments yet.