CodingSubRanking adjusts real subscription quotas by measured output-token consumption — so AI coding plans compare on the work they finish, not the tokens they advertise.
Raw $/MTok comparisons assume every model spends the same tokens on the same job. In reality, concise configurations finish a complex task in 3,000 – 19,000 output tokens while verbose ones burn 115,000 – 204,000 — up to a 65× swing that token-price charts never see.
Three numbers turn a raw token quota into purchasing power you can actually compare.
How many real tasks your monthly token pool can actually solve.
What you actually pay for each completed task.
The true efficiency metric: completed tasks delivered per dollar.
T = monthly usable tokens · C = output tokens per task (DeepSWE v1.1 medians) · P = monthly fee.
All evaluated reasoning effort tiers (Low to Max) are included by default in the rankings.
All 70 evaluated configurations from Datacurve’s contamination-free, long-horizon engineering benchmark — 113 tasks, 91 repos, one harness for every model. These median token figures drive every ranking on this page.
| Rank | Model & effort tier | Pass@1 | API cost/task | Output tokens/task | Steps |
|---|
Ranked by effective cost per task — cheapest completed task first. Shift shows how each plan moves once verbosity is priced in.
| Rank | Shift vs Raw | Plan | Served model | Effort | Fee/mo | Tok/task | Tasks/mo | Cost/task | Tasks/$1 | Raw $/MTok |
|---|
gemini-3.8-flash offers tens of billions of tokens and tops raw $/MTok charts — but at 120,488 median output tokens per task, its effective task cost falls behind far leaner models.
ChatGPT’s gpt-5.6-luna (7,918 tokens) and gpt-5.6-terra (11,453) pair huge pools with lean consumption — terra plans climb up to +41 spots once verbosity is priced in, and glm-5.3-flash breaks the top 10 on quota alone.
DeepSWE ranks 22 model families with plans — nearly double CursorBench’s 11. DeepSeek V4, Kimi K3, GLM-5.2/5.3, Qwen3.8 Max, Opus 4.8/5, GPT-5.5–6 Astra, Gemini 3.x Flash and Grok 4.x all enter; claude-opus-5.5, gpt-6-sol, grok-4.7 and swe-2 now await evaluation.
Top plans per monthly fee range at the Medium baseline. Figures show tasks delivered per dollar.
These models from the real-api-pricing dataset have no DeepSWE v1.1 evaluation — their output tokens per task are unknown, so they’re flagged [Pending DeepSWE] and excluded from the ranking.
API-only plans for benchmarked models (e.g. “Claude Opus 5 API”) carry no monthly token quota, so they can’t be ranked as subscriptions — their absence is intentional, not a coverage gap.
When DeepSWE or a community saturation run publishes completion-token figures for Opus 5.5, GPT-6 Sol, Grok 4.7, MiMo v2.6, Step-5, SWE-2 or any other pending model: update data/model-token-consumption.json and re-run compute_task_ranking.py.
Devin Pro × SWE-2 is unmetered (≈$0/MTok, unbounded quota) under a promotion running until 2026-10-31 — it carries a promo flag in the derived data rather than a rank, and must be re-verified when the promo ends.
The full pipeline is four scripts. Every figure on this page can be regenerated from raw sources.
# regenerate the dataset, rankings and report $ python3 scripts/extract_deepswe.py $ python3 scripts/build_model_consumption_map.py $ python3 scripts/compute_task_ranking.py $ python3 scripts/build_readme.py
Every figure here derives from public benchmarks and community saturation tests — and every table is four commands away.