Coding agents consume up to 1,000x more tokens than chatbots, reshaping AI cost math

·1 views

Research from Stanford, MIT, and others published in 2026 found that a single agentic coding task consumes roughly 1 to 3.5 million tokens, compared to far fewer in a standard code chat, because agents run multi-turn loops involving file reads, tool calls, edits, and self-correction. About 76% of those tokens are reads, meaning prompt caching — which can cost as little as a fraction of standard input pricing — offers significant savings when stable context is structured correctly. Across models, the same coding task can vary in cost by up to 40 times, suggesting that routing routine work to cheaper models and escalating only difficult tasks is a more impactful lever than choosing a premium model by default. Experts recommend benchmarking AI tools on cost per task rather than cost per million tokens, using real workloads from your own codebase. Full per-request visibility into model choice, latency, and spend is also highlighted as essential, since hidden costs in agentic workflows can quietly compound across long sessions.

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)

How IoT Architecture Enables Scalable Telecom Tower Monitoring at Scale

Telecom towers house complex infrastructure—including power systems, generators, cooling units, and security equipment—that is difficult to supervise efficiently through traditional periodic inspections. A modern IoT monitoring architecture connects heterogeneous site equipment through edge gateways, normalizes telemetry, and transmits operational data to a centralized platform. A key design principle is correlating related data points rather than treating sensor readings in isolation, such as linking grid outages, battery switchovers, and generator startups into a single operational event. Fuel levels, environmental conditions, and security alerts can similarly be cross-referenced to distinguish normal activity from anomalies. The primary engineering challenge is ensuring the architecture remains reliable and consistent as the number of monitored towers, vendors, and data streams continues to grow.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Go 1.26 iterators spark debate over language complexity vs. readability

Go, long praised for its simplicity, introduced iterator support via range-over-func syntax in Go 1.23 (August 2024), expanding further in Go 1.26 (February 2026) with iterator-based reflection methods. The changes require developers to understand push/pull iterators, yield callbacks, and boolean early-termination signals — concepts critics say replace what simple for-loops already handled. Benchmarks by Val Deleplace (September 2024) and researcher Rost Glukhov (January 2025) found custom iterators are generally slower than classical approaches, with the trade-off described as largely aesthetic. A new library called Iterium, released in 2026, brought lazy pipeline functions like Map, Filter, and TakeWhile to Go, drawing comparisons to Rust's iterator model. The community remains divided, with some welcoming the expressiveness and others arguing the additions undermine Go's founding principle that clarity matters more than cleverness.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Four CI checks stayed green while broken — a team's account of silent pipeline failures

A software team discovered that four separate automated checks in a single repository had stopped being able to detect failures, yet continued showing green status in their CI pipeline. Among the issues was a GitHub Actions lint job where the tool's timeout matched the job's timeout, causing silent cancellations instead of diagnosable errors across three consecutive commits. A .gitattributes line-ending rule that appeared fixed locally left over 1,200 files with incorrect CRLF endings in pre-existing working trees, with CI and local environments reporting wildly different counts. A guard written to prevent the timeout misconfiguration from recurring falsely flagged a comment in the workflow file, because it scanned raw text without distinguishing code from prose. The team concluded that a check which cannot fail is more dangerous than a broken one, since silent false-positives eventually train developers to ignore alerts altogether.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Refactors Link View Tracking System to Fix N+1 and Duplicate Count Bugs

A developer has refactored the view-tracking system in their open-source link tool, which previously cached link views in Redis and flushed them every 10 seconds using a manual loop. The old approach suffered from an N+1 problem, where each link ID triggered a separate database update, risking resource exhaustion under heavy traffic. It also caused duplicate view counts because cached data was never cleared after flushing, and wasted resources by running updates even when no data existed. The new system uses Redis's hGetAll to fetch all views at once, validates that data exists before proceeding, and batches all updates into a single bulkWrite database operation. Cached views are only deleted after the database acknowledges the write, ensuring no data is lost if the server crashes mid-process.

0 comments Read more at DEV Community