Price Per Task vs Price Per Token: AI Cost Guide 2026
The Short Answer
Price per token is what vendors print; price per task is what you actually pay. In August 2026 OpenAI and Anthropic both started pushing task-level cost metrics (Bloomberg, August 13) — partly genuine (reasoning models made token math misleading), partly competitive (Chinese open-line models undercut them 3-10× per token). The right response isn’t to adopt either side’s framing — it’s to measure cost per completed task on your own workload.
The Formula
cost per attempt = (input_tokens × input_price + output_tokens × output_price) / 1,000,000
cost per task = cost per attempt / success_rate (+ human review cost of failures)Three things the sticker price hides:
- Thinking tokens bill as output. Reasoning models can emit 2-10× the visible output. Some models (e.g. GLM-5.3, released Aug 14, 2026) make thinking mandatory.
- Success rate dominates. A 5× cheaper model that fails half the time isn’t 5× cheaper.
- Caching changes everything for agents. Agent loops re-read context constantly; cache-hit rates of $0.26-$0.50 per MTok (vs full input price) can cut real bills 40-70%.
Worked Example (30K in / 5K out coding task)
| Model | Per-attempt | At 90% success | At 60% success |
|---|---|---|---|
| Claude Opus 5 ($5/$25) | $0.28 | $0.31 | — |
| GPT-5.6 Sol ($5/$30) | $0.30 | $0.33 | — |
| Gemini 3.6 Flash ($1.50/$7.50) | $0.08 | $0.09 | $0.14 |
| GLM-5.3 ($1.40/$4.40) | $0.06 | $0.07 | $0.11 |
| DeepSeek V4 Pro (~$0.435/$0.87 off-peak) | $0.017 | $0.019 | $0.029 |
Reading it: if the cheap model’s success rate on your tasks is within ~10 points of the frontier model’s, the cheap model wins decisively. If it’s 30+ points behind on long-horizon work, the frontier model is often cheaper per completed task once you price retries and human review.
How to Actually Decide (5 Steps)
- Build a 20-50 task eval set from your real workload — not benchmarks.
- Run each candidate model 3× per task; record tokens (including thinking) and pass/fail.
- Compute cost per completed task with the formula above.
- Add failure externalities: engineer review minutes × loaded cost.
- Re-run monthly. Prices moved constantly in 2026 (Luna cut 80% on July 30; Sonnet 5’s $2/$10 intro ends August 31).
The Trap to Avoid
Vendor task-cost marketing always picks tasks their model wins. Treat “per-task pricing” announcements the way you treat benchmarks: directionally interesting, decided only by your own eval. Related guides: best value AI API 2026 and how to choose an AI model in 2026.
Last verified: August 15, 2026. Prices from vendor pricing pages.