Calculate LLM Emissions

<span class="api-method api-method-get">GET</span> /v1/digital/llm/emissions

Estimate CO2e emissions from LLM inference based on provider, model, token counts, and regional grid intensity.

Request

Required Parameters

None — all parameters have defaults. However, for meaningful results you should provide at least provider, model, and token counts.

Model Parameters

Reasoning Parameters

If omitted, the API automatically detects reasoning models (o3, DeepSeek R1, etc.) and applies their default thinking ratio.

Location Parameters

If no region is provided, grid intensity defaults to the provider's primary operating country (e.g., US for OpenAI, France for Mistral, China for DeepSeek).

Output Parameters

Response

Full Response Example

{
  "data": {
    "type": "llm_emission",
    "id": "6863669d-f8ac-47e0-8f8b-4d8ea1767315",
    "attributes": {
      "emissions": {
        "co2e": 0.0557,
        "co2e_unit": "g",
        "co2e_calculation_method": "ipcc_ar6_gwp100",
        "breakdown": {
          "operational_co2e": 0.0057,
          "embodied_co2e": 0.05,
          "unit": "g"
        },
        "energy_usage": {
          "total_wh": 0.1357,
          "per_request_wh": 0.1357,
          "breakdown": {
            "base_energy_wh": 0.1,
            "input_tokens_wh": 0.005,
            "output_tokens_wh": 0.0125,
            "reasoning_overhead_wh": 0,
            "pue_overhead_wh": 0.0136
          }
        },
        "ghg_protocol_scopes": {
          "scope_2": {
            "location_based": 0.0057,
            "method_note": "Based on regional grid intensity for the provider/region. Covers electricity consumed during inference."
          },
          "scope_3_category_1": 0.05
        },
        "source_trail": [
          {
            "data_category": "energy_model",
            "name": "LLM Energy Consumption — Medium tier",
            "source": "Epoch AI / Google / SemiAnalysis",
            "source_dataset": "Composite: How Much Energy Does ChatGPT Use (2025), Gemini Environmental Report (2025), AI Datacenter Energy (2024)",
            "year": "2025",
            "region": "GLOBAL"
          },
          {
            "data_category": "grid_intensity",
            "name": "FR — 42 gCO₂e/kWh",
            "source": "Ember / Electricity Maps",
            "source_dataset": "Cloud provider regional grid intensity",
            "year": "2025",
            "region": "FR"
          },
          {
            "data_category": "embodied_carbon",
            "name": "Hardware amortization — Medium tier GPU cluster",
            "source": "Cloud Carbon Footprint",
            "source_dataset": "CCF methodology — GPU embodied emissions",
            "year": "2024",
            "region": "GLOBAL"
          }
        ]
      },
      "inference": {
        "provider": "mistral",
        "model": "mistral-next-xl",
        "tier": "medium",
        "is_known_model": false,
        "is_reasoning_model": false,
        "thinking_ratio": 1,
        "tokens": {
          "input": 1000,
          "output": 500,
          "cached": 0,
          "reasoning_estimated": 0,
          "effective_output": 500
        },
        "requests": 1
      },
      "grid": {
        "carbon_intensity": 42,
        "carbon_intensity_unit": "gCO2e/kWh",
        "region": "FR"
      },
      "notices": [
        {
          "message": "Unknown model 'mistral-next-xl'. Estimated as 'medium' tier.",
          "code": "unknown_model",
          "severity": "warning"
        }
      ]
    }
  },
  "meta": {
    "methodology": "Energy-per-token estimation with hardware amortization",
    "emission_factors_year": 2025,
    "standards_compliance": {
      "GHG_Protocol": "Scope 2 (purchased electricity) + Scope 3 Category 1 (embodied hardware)",
      "ISO_14040": "Lifecycle assessment aligned"
    },
    "version": "1.0.5",
    "calculated_at": "2026-02-12T10:15:40Z"
  }
}

Response Fields

emissions

energy_usage

ghg_protocol_scopes

inference

grid

notices

Examples

Single GPT-5.2 request

curl "https://api.emissions.dev/v1/digital/llm/emissions?\
provider=openai&\
model=gpt-5.2&\
tokens_input=5000&\
tokens_output=2000" \
  -H "Authorization: Bearer em_live_xxxx"

→ ~1.2g CO₂e (frontier tier, US grid)

Claude Opus with prompt caching

curl "https://api.emissions.dev/v1/digital/llm/emissions?\
provider=anthropic&\
model=claude-opus-4.5&\
tokens_input=100000&\
tokens_output=4000&\
tokens_cached=90000" \
  -H "Authorization: Bearer em_live_xxxx"

→ 90,000 cached tokens save ~80% of input energy. Notice confirms cache applied.

Reasoning model with explicit effort

curl "https://api.emissions.dev/v1/digital/llm/emissions?\
provider=openai&\
model=o3&\
tokens_input=2000&\
tokens_output=1000&\
reasoning_effort=high" \
  -H "Authorization: Bearer em_live_xxxx"

→ 6× thinking overhead applied. 1,000 output tokens → 6,000 effective tokens.

Batch calculation — 100k API calls

curl "https://api.emissions.dev/v1/digital/llm/emissions?\
provider=openai&\
model=gpt-4.1-mini&\
tokens_input=800&\
tokens_output=400&\
requests=100000" \
  -H "Authorization: Bearer em_live_xxxx"

→ Total emissions for your monthly API usage in a single call.

Self-hosted model on UK grid

curl "https://api.emissions.dev/v1/digital/llm/emissions?\
provider=self_hosted&\
model=llama-4-maverick&\
tokens_input=2000&\
tokens_output=1000&\
region=GB" \
  -H "Authorization: Bearer em_live_xxxx"

→ 1.5× efficiency penalty for self-hosted (higher PUE, less optimised hardware). UK grid at 217 gCO₂e/kWh.

DeepSeek R1 on China grid

curl "https://api.emissions.dev/v1/digital/llm/emissions?\
provider=deepseek&\
model=deepseek-r1&\
tokens_input=3000&\
tokens_output=1500" \
  -H "Authorization: Bearer em_live_xxxx"

→ Reasoning model auto-detected (6× thinking ratio). China grid at 544 gCO₂e/kWh increases operational emissions significantly.

With equivalents

curl "https://api.emissions.dev/v1/digital/llm/emissions?\
provider=google&\
model=gemini-3-pro&\
tokens_input=10000&\
tokens_output=5000&\
equivalents=true" \
  -H "Authorization: Bearer em_live_xxxx"

→ Includes comparisons: Google searches, smartphone charges, km driven, Netflix hours.

Provider Efficiency

Different providers have different infrastructure efficiency. We apply a multiplier based on published data and disclosed PUE values:

Calculation Methodology

Energy Model

1. Base energy — Fixed overhead per request (model loading, routing): 0.03–0.5 Wh depending on tier
2. Input token energy — Per 1K tokens: 0.001–0.05 Wh depending on tier
3. Output token energy — Per 1K tokens: 0.005–0.25 Wh depending on tier (output is 5× more energy-intensive than input)
4. Cache efficiency — Cached tokens use 10% of normal input energy (KV-cache reuse)
5. Reasoning overhead — Output tokens multiplied by thinking ratio for reasoning models
6. PUE — 1.1× default Power Usage Effectiveness (cooling, networking overhead)
7. Provider efficiency — Multiplier based on provider infrastructure

CO₂e Conversion

Energy (kWh) × Grid Intensity (gCO₂e/kWh) = Operational CO₂e (g)

Plus embodied carbon: 0.01–0.3g per request depending on tier.

Sources

• Epoch AI — "How Much Energy Does ChatGPT Use?" (2025)
• Google — Gemini Environmental Report (2025)
• Samsi et al. — "From Words to Watts" (2023)
• SemiAnalysis — "AI Datacenter Energy" (2024)
• Cloud Carbon Footprint — GPU embodied emissions methodology