Accenture Caps Claude Subsidies at $29, Shifting Extra Costs to Projects

01Accenture Cuts Claude Subsidy to $29 a Month for Many Employees, With Projects Paying for Extra Use

Accenture has changed how it pays for Claude Enterprise, Anthropic’s workplace AI service, shifting much of the cost for heavy use from a central budget to individual projects or teams. An internal announcement seen by 404 Media says that, effective October 1, the consulting company would centrally fund only $29 per person each month for many employees.

All Claude costs remain consumption-based. Once an affected employee exceeds the $29 allowance, the remaining expense must be assigned to a specific project or team through what Accenture calls a Work Breakdown Structure, or WBS—its internal identifier for tracking work and costs.

The gap between the subsidy and actual usage can be substantial. One employee said in an internal message that their Claude costs had exceeded $500 in recent months and asked how to accelerate an upgrade to a Premium account, saying the upgrade option appeared to be blocked. The employee warned that the additional charges would have a significant effect on project costs.

Another employee said they used Claude Code extensively to build and test an “AI accelerator” being developed for multiple Accenture clients. That work typically generated between $500 and $1,000 in monthly usage, according to the employee. Under the new funding arrangement, such costs may need to be absorbed by the client-related project or another qualifying team unless the employee receives different access.

Premium plans include $250 of token usage per week, but many employees are not on that plan, 404 Media reported. Some workers have therefore sought Premium upgrades to prevent their projects from being billed for additional consumption. The source does not establish how many employees are covered by the $29 subsidy, how Premium access is awarded, or whether Accenture’s overall Claude budget has fallen.

Access itself is also subject to review. The internal announcement said obtaining a license and supplying a WBS would not guarantee continued use. Allocations may be reconsidered according to business priorities, job requirements, usage patterns and workforce needs. One employee said their Claude access had already been revoked.

Accenture said it continues to make AI available to all employees and aims to provide the appropriate tool for specific work. It also said it remains committed to investing in AI capabilities, skills and support that create value for clients and the company. The firm did not provide its previous average spending per employee or its total budget after the change.

02Google says Gemini 4 Argon can span the robotics development workflow and teach Spot to step on a pedal and toss cans

Google is positioning Gemini 4 Argon as a general-purpose physical reasoning model that can work across the robotics lifecycle, from building simulations and planning actions to executing and checking physical tasks. The company says Argon scored 87% using reasoning alone on ERQA, a benchmark for understanding and solving physical and spatial problems in three-dimensional environments, outperforming other frontier models.

Its most complete demonstration follows a Boston Dynamics Spot quadruped learning to discard soda cans in a foot-pedal trash bin. Argon first created an articulated digital twin of the bin in MuJoCo, a physics simulator, so the robot could practice pressing the pedal and placing cans inside before attempting the task with the physical objects.

During real-world execution, Argon directed Spot to collect two cans and use the bin. The instructions required the robot to check its front cameras and confirm that its front-right foot target was centered on the pedal before stepping. A two-agent verification system then divided the checks: a “foot-verifier” assessed the pedal press, while a “hand-verifier” checked the three-dimensional alignment between Spot’s manipulator and the bin before the can was released.

Google presented that combination as a closed-loop workflow: visual observations can change what the system does next, rather than leaving the robot to follow one fixed sequence, while separate verification agents check distinct stages where an error could derail the task.

Other demonstrations extended the same approach. Argon used wrist-camera images, a coordinate grid and predefined action primitives over several turns to stand up a fallen mug and put a crumpled napkin inside. In Google’s Gemini Omni world model, it inspected intermediate keyframes and rewrote timestamped action prompts to guide a robot in arranging five letter tiles into “ARGON.” Google also showed it iteratively creating and tuning digital twins from real robot teleoperation recordings to reproduce their visuals and contact events.

The evidence establishes the range of workflows Google demonstrated, but not their reliability in broader deployment. The company did not provide repetition counts, task success rates, failure cases, execution costs or independent evaluations, and it did not establish how Argon’s ERQA score translates into dependable performance on physical robots.

03132 Training Runs Find None of Ten Agent Self-Distillation Methods Consistently Beat GRPO

A large-scale benchmark has challenged the case for increasingly complex self-distillation methods—training approaches within agentic reinforcement learning—by finding that none of ten representative methods consistently outperformed GRPO, the study’s comparison baseline. The researchers say earlier evaluations often reused published baseline scores produced under different experimental conditions, while rarely comparing contemporary methods directly or examining training stability and computing costs together.

The new study placed all ten self-distillation methods and GRPO within a unified experimental framework. Researchers tested them on four types of agent tasks: ALFWorld, ScienceWorld, Search-QA and WebShop. The comparison covered three model sizes, although the abstract does not identify those sizes or provide results for each model-task combination.

In total, the team completed 132 training runs and recorded 12,124.4 H800 GPU-hours. It also repeatedly evaluated the pretrained model backbones and all 11 trained methods using eight rollout seeds, helping measure whether reported results held across repeated trials rather than depending on a single evaluation run. This common setup makes the findings more directly comparable than scores collected from separate papers using differing conditions.

“Did not consistently beat” does not mean every self-distillation method failed in every experiment. The paper reports that gains over GRPO were limited in scope: no method maintained an advantage across the different settings. The available abstract does not list the ten methods, their individual scores or which particular combinations produced improvements.

The benchmark also found that a strong final task score could conceal instability during training. For teams choosing a training method, that means the final number alone may not show how reliably the model reached it. The researchers separately measured computational cost and found that spending more on training did not consistently produce better performance.

The practical judgment is therefore three-dimensional: compare final performance, the stability of the training process and the compute required to obtain the result. The detailed cost distribution across methods, tasks, model sizes and runs remains unspecified in the abstract.