Cheap per token ≠ cheap per task.
Text models bill you for every token spent reaching an answer, not for the answer. On the same problems that spend varies by 9× at default settings, and 17× with the effort dial at max — and it drives your invoice, your energy, your carbon and your water.
In plain words
When you ask a chatbot something, it often “thinks out loud” first, writing text you never see. That text is counted in tokens — roughly pieces of words.
Some models write far more tokens than others to reach the same answer, and every token costs money, electricity and water. The chart compares them on the same set of test problems.
Same 200 problems, wildly different meters
per problem · lightest firsthow much each model writes to finish one problem — measured × published
The experiment
Every model gets the same 200 problems and only the final answer is graded. Each model chooses how much working-out it writes on the way — and you pay for every page.
7 pages → 86% correct
63 pages → 78% correct
9× the paper for 8 points fewer — and the paper is tokens on your invoice and watt-hours in a datacenter.
Wondering about your own prompts? Try one in the Playground →
Run it on your own workflow
Pick two models and a monthly volume. The per-problem gap above becomes dollars, energy, carbon and water at your scale.
- MoneyDeepSeek V4 Flash (high) costs $55,127 less
- EnergyGPT-5.5 (medium) uses 8,650 kWh less
- CarbonGPT-5.5 (medium) emits 4,152 kg less
- WaterGPT-5.5 (medium) uses 15,571 L less
- AccuracyGPT-5.5 (medium) scores 8.0 pts higher
GPT-5.5 (medium)
86% correct · 4,692 tok/task
$84,456
1,126 kWh
≈ 37.5 days of a home's power · range 563 kWh – 1,689 kWh
541 kg
≈ 1,351 mi in a gas car · range 270 kg – 811 kg
2,027 L
≈ 4,054 bottles (0.5 L) · range 113 L – 6,756 L
DeepSeek V4 Flash (high)
78% correct · 40,735 tok/task
$29,329
9,776 kWh
≈ 326 days of a home's power · range 4,888 kWh – 14,665 kWh
4,693 kg
≈ 11,732 mi in a gas car · range 2,346 kg – 7,039 kg
17,598 L
≈ 35,195 bottles (0.5 L) · range 978 L – 58,658 L
The cheap-per-token trap. Over a year at 600,000 tasks, DeepSeek V4 Flash (high) costs $55,127 less on the invoice — but burns 4,152 kg more CO₂e, 8,650 kWh more energy and 15,571 L more water, a 8.68× heavier footprint for the same work. The lower sticker price is subsidised by the atmosphere, not by efficiency. DeepSeek V4 Flash (high) also scores 8.0 pts lower — more footprint, less accuracy.
How this projection is madetap to open
Tokens per task measured by OckBench — Measuring the Efficiency of LLM Reasoning (2025); prices from BenchLM.ai — LLM API pricing comparison (Aug 2026) (output tokens only). Energy, carbon and water are derived through the bands in the assumptions below (Ember — Global Electricity Review (world-average grid carbon intensity), Li et al. 2023 — Making AI Less Thirsty (water per kWh band)); mid shown, band printed. Bars are drawn to a fixed scale — full width is the heaviest model in the table at your volume. A projection, not a quote — your serving stack, grid and prompt mix will shift the absolute numbers, but the ratio between models holds.
Share it
1200×630 PNG · X, Open Graph, Slack
The card carries the pair, the volume and the band, so the number holds up in the replies. The link re-runs the comparison against the live table.
On long agent tasks
Multi-step work that runs for minutes, measured by Artificial Analysis. Much longer tasks than the chart above, so compare these rows with each other, not with that chart.
Output tokens per task · lightest first
About this benchmark6 notestap to open
- Tokens per task are the output (reasoning + answer) tokens Artificial Analysis recorded across the evaluations in its Intelligence Index: long agent tasks such as terminal work, office documents and research. The index is AA's own composite score out of 100, not a percentage correct.
- Each model runs at the effort setting printed beside it, as AA chose; most are at max. Effort moves token use a lot (DeepSeek V4 Flash doubles from high to max on the hard-problems chart), so two rows at different settings are not a like-for-like pair.
- "Max" for Claude Opus 5.5 and Sonnet 5.5 is AA's "max with fallback": when safeguards intervened, a task was finished by an older Claude model.
- Energy counts output tokens only, at the same 0.0002–0.0006 Wh per token as the rest of this page. Agent tasks also re-read long contexts, mostly from cache, which this leaves out, so read the watt-hours as a floor for comparing models rather than a full bill.
- No cost column: agent work is billed mostly on input and cache reads, which a per-output-token price cannot capture.
- Qwen3.8 Max: AA's model page gives 62,782 tokens per task, while its all-models listing printed 107,730 for the same model on the same day. The model-page figure is used.
Source: Artificial Analysis Intelligence Index — output tokens per task, read 2026-10-02, each figure read from that model's own page on 2026-10-02.
What this means for your rollout
Assumptions & sources6 numbered assumptionstap to open
- Energy per output token assumed at 0.0002–0.0006 Wh (mid 0.0004). Assumes current production serving: H100/H200-class accelerators with FP8 quantization and batched inference, GPU + host power, datacenter overhead (PUE) ≈ 1.3; excludes idle capacity, networking and training. Anchors: ~0.39 J (0.00011 Wh) per token measured for a 70B model on H100/FP8 (GPU only), and the Joule 2026 median of 0.31 Wh for a frontier query. The full literature band is 0.0001–0.002 Wh/token; older hardware or unbatched serving sits at the high end. Independently measured serving energy brackets this band: the ML.ENERGY Leaderboard measures DeepSeek V3.1 on 8×B200 at 1.34 J (0.00037 Wh) per token at batch 256 up to 10.1 J (0.0028 Wh) at batch 8 — batching alone spans 7.5×, which is why this stays a band.
- One printed page ≈ 500 words ≈ 650 tokens.
- One full iPhone 17 charge ≈ 15 Wh at the wall (≈12.5 Wh battery capacity plus charging losses).
- Cost per task = measured output tokens × list price per 1M output tokens (Oct 2026), verified against vendor pricing pages. Output tokens only — input tokens, caching discounts, batch tiers and volume deals excluded. DeepSeek moved the V4 family to peak/off-peak pricing on 16 Aug 2026 and cut the Flash price with the V4.1 release on 10 Sep 2026; the standard (peak) rate is used here ($1.20/M Flash, $3.96/M Pro), since every other row is likewise quoted undiscounted. The DeepSeek V4 rows were measured on builds DeepSeek has since replaced: V4 Flash was retired on 10 Sep 2026 and its API name now routes to V4.1 Flash. They stay until OckBench measures V4.1. The DeepSeek API defaults to high effort, which is why the high row is the one featured. Qwen3.5 9B has no official list price; a typical hosted rate of $0.40/M is assumed, so it is excluded from cost superlatives.
- Carbon = energy × 480 gCO₂e/kWh, the world-average grid intensity (Ember). Equivalents: 30 kWh per home-day; 0.4 kg CO₂e per gas-car mile (EPA).
- Water = energy × 0.2–4 L/kWh (mid 1.8), the same world-average datacenter water band the calculator uses (Li et al. 2023). It is the widest band on this page: cooling water depends on climate, cooling design and whether the water behind the electricity is counted.
Token counts and accuracies read from the OckBench live leaderboard — data snapshot 2026-08-22 — a shared set of 200 math, coding and science problems, method described in OckBench — Measuring the Efficiency of LLM Reasoning (2025). List prices from BenchLM.ai — LLM API pricing comparison (Aug 2026), cross-checked against each vendor's published pricing page. Energy is derived from those counts, not measured — vendors don't disclose serving efficiency, so the token ratio is the solid part and the watt-hours stay an estimate; see ML.ENERGY Leaderboard — measured serving energy per token for the measured batching spread quoted in assumption 1. Water uses the calculator's own band, from Li et al. 2023 — Making AI Less Thirsty (water per kWh band).
Static exhibit, last reviewed 2026-10. Figures are third-party published measurements (see also Artificial Analysis — model pages (output tokens to run the Intelligence Index) for per-model token usage across a wider suite); text models are not yet part of the calculator's dataset. Verbosity also varies with each model's reasoning-effort setting — the same model at "max" can spend several times the tokens of "medium".