Developer audit finds session history, not AI errors, drives 80% of LLM costs

·1 views

A developer analyzed 660 AI session transcripts spanning 45 days to identify what was driving up costs in their local AI assistant setup. The audit found 729 failed tool calls, but those recovery costs amounted to only 1.5% of total spending — far less than expected. The dominant cost driver turned out to be context re-reading: as sessions grow longer, the model re-processes its entire conversation history on every reply, silently inflating token usage. Three long-running sessions alone accounted for 20% of one week's total token burn, with the worst single session consuming 240 million cache-read tokens across 665 messages. The developer responded with three practical fixes, the most impactful being a strict habit of archiving long sessions and starting fresh with a brief handoff summary.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Claude Suffers ~20 Outages in 24 Days, Anthropic's Status Page Confirms Pattern

Anthropic's AI assistant Claude experienced approximately twenty separate service incidents in the first twenty-four days of August 2026, according to the company's own public status page at status.claude.com. Issues ranged from degraded performance and elevated error rates to full service disruptions, with one major incident on August 24 simultaneously affecting claude.ai, the Claude API, Claude Code, and Claude Cowork. The high frequency is particularly problematic for paying subscribers, whose usage is metered against rolling five-hour and weekly caps — meaning failed requests can still consume a user's allowance. Anthropic has been transparent by maintaining a detailed public status page, but critics argue that near-daily disruptions constitute a systemic reliability problem rather than isolated incidents. Paying customers are effectively absorbing the cost of instability in the form of lost time and depleted usage quotas.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

TinyML Project Brings Offline Skin Lesion Scanning to Raspberry Pi 3

A developer named circuitrocks has built a TinyML-powered skin lesion scanner called LesionIQ that runs on a Raspberry Pi 3. The project was published on DEV Community on August 27 and falls under the domains of edge AI and machine learning. LesionIQ is designed to analyze skin lesions entirely offline, without requiring a cloud or internet connection. The use of TinyML allows the machine learning model to run efficiently on the low-power, low-cost Raspberry Pi 3 hardware.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

JWT Scope and Audience Flaws Leave Go Microservices Open to Privilege Escalation

A technical analysis highlights how most Go microservice deployments correctly verify JWT signatures but fail on authorization semantics, creating exploitable security gaps. Two key failure modes are identified: scope inflation, where a broad token claim is accepted by unintended services, and audience misrouting, where tokens issued for one service are accepted by another due to skipped or misconfigured audience validation. The widely used golang-jwt/jwt library makes audience validation opt-in rather than enforced by default, meaning developers under deadline pressure often omit it entirely. This oversight can allow tokens meant for one service, such as a payments API, to be replayed against unrelated services like reporting, with no trace in logs. The article provides corrected Go code that enforces both audience and scope checks, arguing that without a coherent boundary contract, RBAC enforcement scattered across individual services creates privilege escalation paths that are nearly impossible to audit.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Angular 19 Signals Enable Zoneless, 60fps Reactivity for Enterprise Apps

Angular 19 introduces fine-grained Signals as a modern alternative to the long-standing Zone.js change detection model. Unlike Zone.js, which triggers dirty-checking across the entire component tree on every async event, Signals track exact DOM dependencies and update only the nodes that have changed. This approach eliminates Zone.js runtime overhead entirely via the provideExperimentalZonelessChangeDetection() configuration, reducing memory leaks and frame drops common in large enterprise dashboards. Derived values using computed() are lazily evaluated and memoized, replacing complex RxJS chains like combineLatest and switchMap. The result is a more performant and maintainable reactive architecture capable of sustaining 60fps rendering in data-heavy applications such as live telemetry monitors and grid streams.

0 comments Read more at DEV Community