Cheaper intelligence is expanding the surface area of AI operations
Today's signal is that greater model capability does not make the operating layer disappear. It expands where AI can be used: prompts get smaller, software becomes more disposable, agents gain control of production surfaces, and lower unit costs unlock more workloads. The result is more leverage—and more systems that need attribution, evaluation, and governance.
Anthropic reportedly cut Claude Code’s system prompt by 80%
TestingCatalog reports that Anthropic substantially reduced Claude Code's system prompt as newer models became more capable. The team's explanation is that excessive instructions, constraints, and examples can narrow a stronger model's behavior instead of improving it.
Revalidate the harness whenever the underlying model changes. Run controlled tests that remove examples and legacy instructions, then compare task success, latency, and token use. Prompt complexity should be earned by evaluation—not inherited indefinitely.
Source: Anthropic Cuts Claude Code System Prompt by 80% — TestingCatalog
Paid-ad agents are being given websites they can control
GTM practitioner Cody Schneider says his team gives paid-ad agents their own websites so they can generate and control landing pages. The claim points to a shift from agents that recommend campaign changes to agents that can produce and operate the conversion surface themselves.
Treat an agent-managed website as a production system. Constrain domains and approved claims, separate publishing from campaign budgets, preserve every revision, and require rollback and approval paths before an experiment can change what customers see.
Source: Cody Schneider on X
AI is turning software into disposable working material
Hiten Shah frames AI-generated software as scratch paper: a team can create a dashboard to answer one question, explore an interface, and then move on rather than treating every application as a durable asset that must be maintained and reused.
Short-lived software still touches real data and systems. Give ephemeral apps an expiration date, least-privilege access, and an owner; capture the decision or insight they produced; and remove credentials, storage, and deployments when the question has been answered.
Source: Hiten Shah on X
Lower AI prices are likely to increase total consumption
Box CEO Aaron Levie argues that cheaper tokens make a wider range of tasks economical, so falling unit prices can increase total AI spending rather than reduce it. The relevant budget effect is the number of newly viable workflows, not the cost curve of a fixed workload.
Do not translate a lower model price directly into a lower budget. Forecast adoption elasticity, attribute usage to workflows, and set spend limits and outcome thresholds before cheaper inference turns previously invisible demand into an uncontrolled operating expense.
Source: Aaron Levie on X
Every says Claude is now handling about 70% of its routine copy edits
Every CEO Dan Shipper says Claude has automatically completed roughly 70% of the copy edits the publication would ordinarily perform by hand over the past week. He describes it as the first time the workflow has worked after years of attempts, suggesting that model capability crossed a practical production bar.
Define the quality threshold before automating editorial work. Measure accepted edits, reversals, defect rates, and reviewer time against the manual baseline; preserve human approval for material changes; and expand scope only when the full workflow improves.
Source: Dan Shipper on X
Forward-deployed AI may be a service layer on the way to a platform
Developer Dax Raad compares today's forward-deployed AI engineering firms to the service shops that built e-commerce systems for large companies during the early internet era. Those businesses created value, but the larger platform opportunity emerged in reusable infrastructure such as Shopify.
Use deployments to identify the repeated control plane beneath bespoke work. Standardize identity, connectors, evaluation, observability, and billing while leaving workflow-specific judgment at the edge. Reuse—not project volume—is the signal that services are becoming a platform.
Source: Dax Raad on X
A public AMD repository may reveal Anthropic as a customer
SemiAnalysis spotted a reference to Anthropic as a customer in public GitHub code associated with AMD's senior director of AI. Neither company is cited as confirming the relationship, but the reference is a notable signal that a frontier lab may be diversifying beyond a purely NVIDIA-based supply path.
Accelerator diversification is valuable only when the software path travels with it. Track model compatibility, kernel maturity, serving performance, and observability across vendors so alternate capacity is operational—not merely available on a procurement plan.
Source: SemiAnalysis on X
Bottom line
Model progress is compressing the instructions needed for each task while expanding the number of tasks worth automating. That combination drives more software, more autonomous production surfaces, more infrastructure demand, and more total spend. The durable advantage remains the operating layer that can measure outcomes, limit authority, and retire systems as quickly as AI creates them.
Sources
- Anthropic Cuts Claude Code System Prompt by 80% — TestingCatalog
- Cody Schneider on X
- Hiten Shah on X
- Aaron Levie on X
- Dan Shipper on X
- Dax Raad on X
- SemiAnalysis on X
Chompute Daily is a weekday briefing on the operating consequences of enterprise AI news. Reporting belongs to the linked publishers; analysis and operator takeaways are Chompute's.