Why mid-tier models like Claude Sonnet often win real agentic coding work — drag the podium to orbit, pick a task, watch the recommendation.
| Fast tier | — |
| Mid tier | — |
| Frontier tier | — |
Agentic coding is not one clever answer — it is a loop. The model plans, edits files, runs tests, reads the errors, and iterates, often 20–100 times per task.
That rewards three traits over raw brilliance: tool-use reliability (calling the right tool with valid arguments nearly every time), instruction-following that survives long loops (still respecting constraints on step 40), and self-correction — treating a failing test as information, not a dead end.
Per-step quality only has to clear a “good enough” bar, because the loop catches mistakes. What compounds across every iteration is cost × latency.
A rough model: total task cost ≈ iterations × tokens/iteration × price/token, and wall-clock ≈ iterations × latency. A mid-tier model at ~1/5 the price and 2× the speed of a frontier model can run far more loop iterations per dollar — which is exactly how models like Claude Sonnet 5 (June 2026), pitched as Anthropic’s most agentic mid-tier coder, earn their keep in Claude Code.
Don’t pick a tier on vibes or leaderboards. Take 10–20 real tasks from your own backlog, run each tier on them, and score completion rate, dollars per completed task, and time to green tests.
Rule of thumb: fast tier for short, well-specified loops; mid tier as the default workhorse for building and refactoring; frontier tier for one-shot deep reasoning where a wrong answer poisons everything downstream. Then re-run the eval when a new model ships — tiers shift under your feet.
The page assigns tasks-per-dollar constants three hundred eighty to fast, one hundred twenty to mid, and fourteen to frontier. At budget ten it multiplies to thirty eight hundred, twelve hundred, and one hundred forty illustrative tasks. The bars compare fast and mid at point zero seven pixels per task, spanning two hundred sixty six and eighty four. Doubling budget doubles every count without accounting for task complexity, failure, retries, tokens, actual model pricing, or throughput limits. These constants are source fixtures, not current benchmark results or a recommendation justified by evidence. The offline missing Three dependency stops bindings before the budget recalculates. Each selected tier column targets its authored score times three point two. Fast capability point five two therefore maps to height one point six six four and frontier point nine nine to three point one six eight. At one hundred pixels per geometry height the bars span one hundred sixty six point four and three hundred sixteen point eight. Scores for capability, speed, and cost efficiency are manually supplied values from zero to one. An unselected column targets point one five instead, so visual height also depends on selection. Smooth interpolation and glowing podiums do not establish comparative model capability, accuracy, reliability, or actual speed. Four task buttons are assigned to tier indices zero, one, one, and two by a fixed lookup. The explanatory text does not evaluate a user repository or test whether the chosen tier succeeds. The optional orbit has five named decorations: plan, edit, run tests, read errors, and iterate. At sixty pixels per count the bars span two hundred forty task buttons and three hundred orbit labels. Dragging or zooming would move a Three camera, not launch tools or model workers. The offline original fails at the first Three renderer access, leaving a blank positive canvas and editable native controls without application behavior. Preserve that limitation and source rather than claim actual agent execution, provider facts, exhaustive functions, or public playback.