Anthropic releases Opus 5.5 with lower prices and higher limits
The new model cuts token prices and increases subscription allowances. Anthropic claims typical workloads cost 40% less than with Opus 5.
The briefing
Anthropic cuts Opus prices and raises subscription limits. NVIDIA releases free robotics development tools, while Qualcomm announces phone chips designed to run larger AI models locally. The practical gains are clearer than the performance promises: cheaper tokens, reusable software and more processing on the device.
Anthropic releases Opus 5.5 with lower prices and higher limits
Anthropic has released Claude Opus 5.5 with cheaper input, output and cache reads. Subscribers get higher five-hour limits, while the company reports improvements in coding, speed and resistance to prompt injection.
Claude users get two distinct changes: lower metered prices and more room within subscriptions. Anthropic is raising five-hour limits on Pro, Max, Team and seat-based Enterprise plans, and giving subscribers a usage-limit reset they can save for later.
Standard pricing is $4 per million input tokens and $20 per million output tokens, both 20% below Opus 5. Cache reads fall 60% to $0.20 per million tokens, a larger reduction for applications that repeatedly reuse stored context.
The bill depends on the workload
Anthropic's claimed 40% saving combines lower prices with fewer tokens used per task in its tests at default settings. That is a workload estimate, not a universal discount. The company also reports output generation more than 30% faster than Opus 5.
Fast mode in Claude Code and the Claude Platform costs $8 per million input tokens and $40 per million output tokens, with Anthropic claiming up to 2.5 times the speed.
Anthropic says external evaluators including METR and Frontier Design tested the model before release. It reports stronger prompt-injection defences, rather than immunity. Vetted organisations can apply for biology research access through its Life Sciences Verification Program; expanded cybersecurity verification access is planned.
Teams running repeated coding or document tasks can compare the new token rates against their own workloads. Heavy reuse of cached context brings the largest stated price cut, while subscribers gain extra capacity without changing models solely to avoid limits.
Market signal
NVIDIA releases free Isaac ROS 5.0 tools for robot developers
NVIDIA's free, open-source Isaac ROS 5.0 adds support for ROS Lyrical and Ubuntu 24.04. New workflows help developers configure robotics software, adapt stereo perception and assemble picking-and-placing applications.
Robot developers can now use Isaac ROS 5.0, NVIDIA's collection of GPU-accelerated packages built on the ROS framework from Open Robotics. Released at ROSCon in Toronto, it supports ROS Lyrical and Ubuntu 24.04.
In NVIDIA's announcement, Katie Washabaugh describes reusable workflows for developers and AI agents, alongside documentation structured for agents to use. These cover setup and manipulation tasks rather than just help with individual pieces of code.
A FoundationStereo fine-tuning workflow lets an agent help adapt stereo vision to a developer's cameras and operating environment. A separate picking-and-placing workflow connects object detection, depth estimation and position information, and can be used beyond Isaac ROS.
From development machine to robot
The release supports NVIDIA hardware from Jetson Orin Nano to Jetson Thor. NVIDIA also worked with the Open Source Robotics Alliance on a standard data-handling interface for ROS Lyrical, intended to help software work across different computing hardware.
FoundationPose, which estimates and tracks an object's position and orientation, gains an agent-ready inference library. NVIDIA claims it runs up to 5.5 times faster; that is a component-level claim, not a measured speed increase for an entire robot.
The software is available free and open source. Its GPU-accelerated packages provide reusable building blocks, while camera adaptation and application testing remain part of the development work.
Developers building vision-guided robot arms can reuse setup, perception and manipulation workflows instead of assembling every step themselves. Support across Jetson devices also offers a shared software foundation for moving applications between NVIDIA robot computers.
What to watch
Qualcomm announces two Snapdragon chips built for local AI agents
Qualcomm's Snapdragon 8 Elite Gen 6 processors add sensing hubs for small AI models and support for local voice agents. The Extreme version can run a larger mixture-of-experts model, the company says.
Qualcomm has announced the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6 at its annual Snapdragon Summit. The processors put more of the machinery for personal AI assistants inside the phone.
Ivan Mehta reports for TechCrunch that the new sensing hubs can run models with up to 200 million parameters. Qualcomm says these can support a local transcription assistant that distinguishes speakers, alongside memory built from usage to inform suggestions for automating tasks.
The company also says the chips can run a complete agent taking voice input and producing spoken output.
Larger models, and a handset to follow
The Extreme version can run a 30-billion-parameter mixture-of-experts model locally, Qualcomm says. Such a model activates only part of its parameters for a given task, so the headline size does not mean all 30 billion are working on every request.
Both processors support AI-based vocal enhancement and noise reduction. The Extreme also supports 8K video at 60 frames per second and 4K at 240 frames per second, plus the Advanced Professional Video codec.
Motorola announced the Signature 27 smartphone using the Extreme chip, with general availability planned for 2026. That provides a named device for the processor, but the chip's supported capabilities should not be read as a promise that every handset will offer every feature.
Phone makers gain hardware for transcription and voice interactions that can run locally rather than requiring remote processing for those tasks. For buyers, the useful distinction will be the features implemented in a handset, not simply its supported model size.
What to watch next
- Teams running repeated coding or document tasks can compare the new token rates against their own workloads. Heavy reuse of cached context brings the largest stated price cut, while subscribers gain extra capacity without changing models solely to avoid limits.
- Developers building vision-guided robot arms can reuse setup, perception and manipulation workflows instead of assembling every step themselves. Support across Jetson devices also offers a shared software foundation for moving applications between NVIDIA robot computers.
- Phone makers gain hardware for transcription and voice interactions that can run locally rather than requiring remote processing for those tasks. For buyers, the useful distinction will be the features implemented in a handset, not simply its supported model size.
The takeaway
Teams running repeated coding or document tasks can compare the new token rates against their own workloads. Heavy reuse of cached context brings the largest stated price cut, while subscribers gain extra capacity without changing models solely to avoid limits.
The editor’s view
Developers building vision-guided robot arms can reuse setup, perception and manipulation workflows instead of assembling every step themselves. Support across Jetson devices also offers a shared software foundation for moving applications between NVIDIA robot computers.