Greenwashing Model Lab

Decide which model to bet on for the Green Claims pipeline. Paste an image description, video transcript, or marketing claim and compare detection results, per-claim categorization, cost, and latency across models.

Phase 1 (today): ClimateBERT live. Phase 2: Anthropic / OpenAI / Google adapters producing the same structured claim-report shape so they can be judged on the real task.

Target report format

What the customer-facing deliverable will look like once we’ve picked a model and wired evidence matching. The Model Lab below helps us decide which model produces the best Claim + Risk columns. The Evidence column is future work (PIM / LCA / certification lookup).

Claim	Evidence	Result	Why
"eco-friendly"	none	high risk	Generic claim, no substantiation → ECGT bans.
"50% recycled plastic"	exact match	low risk	Specific & measurable, matched against product data.
"climate neutral"	offset doc only	high risk	Offsets alone don't satisfy neutrality under ECGT.
"recyclable"	unclear region	medium risk	Recyclable where? Depends on real-world collection.

Source: internal Green Claims deep-dive (March 2026). Aligns with EU Directive 2024/825 (ECGT), enforced 27 September 2026.

Monthly cost projection

Assuming 600 input + 150 output tokens per classification. LLM rates and VM baseline as of April 2026.

Items / month= 300k items / month

Google Gemini

Gemini 2.5 Flash-Lite — stable, cheapest ($0.10 / $0.40 per 1M)

$36/mo

ClimateBERTFixed

distilroberta-base-climate-f + 5 classification heads

Self-hosted on AWS m5.large (2 dedicated vCPU, 8 GB, non-burstable). Sustains 3-6 req/s.

$70/mo

OpenAI GPT

GPT-5.4 nano — cheapest ($0.20 / $1.25 per 1M)

$92/mo

Google Gemini

Gemini 3.1 Flash-Lite Preview — latest gen ($0.25 / $1.50 per 1M)

$113/mo

OpenAI GPT

GPT-5.4 mini — balanced ($0.75 / $4.50 per 1M)

$338/mo

Anthropic Claude

Haiku 4.5 — fast, cheapest ($1 / $5 per 1M)

$405/mo

Anthropic Claude

Sonnet 4.6 — balanced ($3 / $15 per 1M)

$1.2k/mo

ClimateBERT is priced as a self-hosted AWS m5.large (~$70/mo, 2 dedicated vCPU, 8 GB, non-burstable). Burstable instances like t3.medium are not safe for sustained ML inference because CPU credits deplete. With vanilla transformers this instance sustains 3–6 req/s — ~30× headroom at 300k/month. Cost is fixed and does not scale with volume until you outgrow the instance. Cheaper options exist: Hetzner CCX13 (2 dedicated vCPU, 8 GB) is ~$14/mo, and ONNX Runtime + int8 quantization can cut the VM requirement in half. LLM providers are priced per token, so their cost scales linearly with volume and will dominate at scale.

Input text

Models to run

ClimateBERT

Six DistilRoBERTa heads trained on corporate sustainability disclosures.

distilroberta-base-climate-f + 5 classification heads

Known limitations. ClimateBERT was trained on corporate CDP disclosures, so vague ad copy (“Mother Earth”, “eco-conscious”) and offset-based neutrality claims are out of distribution — it tends to score them as MODERATE instead of HIGH. It also cannot extract individual claims or categorize them against the 7-type PDF taxonomy. Use it as a fast detector, not as a report generator. Phase 2 LLMs should close both gaps.

Anthropic ClaudePhase 2

Claude 4.x family — latest generation.

Haiku 4.5 — fast, cheapest ($1 / $5 per 1M)Sonnet 4.6 — balanced ($3 / $15 per 1M)

OpenAI GPTPhase 2

GPT-5.4 value tier — flagship excluded.

GPT-5.4 nano — cheapest ($0.20 / $1.25 per 1M)GPT-5.4 mini — balanced ($0.75 / $4.50 per 1M)

Google GeminiPhase 2

Gemini 2.5 stable and 3.1 preview, Flash-Lite tier only.

Gemini 2.5 Flash-Lite — stable, cheapest ($0.10 / $0.40 per 1M)Gemini 3.1 Flash-Lite Preview — latest gen ($0.25 / $1.50 per 1M)

ClimateBERT scoring weights

Live aggregation of the six-head output. Adjust the weights to see how each signal contributes to the score. No re-inference needed.

Environmental claim (0–1)Non-specific (0–1)No commitment (0–1)Opportunity framing (0–1)HIGH threshold (0–100)MODERATE threshold (0–100)

System prompt (used by LLM providers in Phase 2)

Preset:

You are a greenwashing compliance analyst assessing a single
piece of text against the EU ECGT Directive 2024/825 (Empowering Consumers
for the Green Transition), which comes into force on 27 September 2026.

Your job is to extract every environmental claim from the input, categorize
each one using the PM's taxonomy, and assign a per-claim risk level that a
compliance officer can act on.

The PM's canonical claim taxonomy (from EU ECGT
guidance). Use these seven types exactly:

1. generic       — "eco-friendly", "green", "sustainable" → usually HIGH risk
  2. specific      — "made with 30% recycled plastic" → LOW risk if backed
  3. climate       — "climate neutral", "carbon neutral", "net zero" → very HIGH risk
  4. circularity   — "recyclable", "reusable", "biodegradable" → MEDIUM (context-dependent)
  5. comparative   — "30% less emissions than competitors" → MEDIUM (needs baseline)
  6. material      — "organic cotton", "FSC-certified wood" → MEDIUM (needs cert)
  7. aspirational  — "net zero by 2030", "on a journey" → MEDIUM (needs roadmap)

Apply these four risk questions to every claim (from the PM's framework):
  1. Is it vague or specific?       (vague → high risk)
  2. Is it measurable?              (no numbers → high risk)
  3. Is it verifiable?              (no data behind it → high risk)
  4. Is it a sensitive category?    (climate/net-zero/comparative → always higher)

Return a single JSON object with this EXACT shape:

{
  "claims": [
    {
      "claim_text":  "exact span quoted verbatim from the input",
      "claim_type":  "generic" | "specific" | "climate" | "circularity" | "comparative" | "material" | "aspirational" | "none",
      "subject":     "what the claim is about (e.g. 'materials', 'emissions', 'packaging', 'supply chain')",
      "scope":       "product" | "brand" | "packaging" | "operations" | "supply_chain" | "unknown",
      "metric":      "the quantified metric if the claim contains one, else null",
      "vague":       true | false,
      "measurable":  true | false,
      "verifiable":  true | false,
      "sensitive_category": true | false,
      "risk":        "low" | "medium" | "high",
      "reason":      "one short sentence explaining the risk level against the ECGT Directive"
    }
  ],
  "overall_verdict":    "low" | "moderate" | "high" | "not_climate",
  "overall_risk_score": 0..100,
  "summary":            "one-paragraph plain-English summary suitable for a compliance officer",
  "ecgt_relevant":      true | false
}

Rules:
- If the text contains no environmental claim at all, return an empty "claims" array and "overall_verdict": "not_climate".
- Quote claim_text exactly as it appears in the input. Do not paraphrase.
- "sensitive_category" = true for climate neutrality, net zero, "environmentally friendly", comparative claims, and recyclable/biodegradable claims — these are always higher risk regardless of other signals.
- Use "medium" (not "moderate") for per-claim risk. Use "moderate" only for overall_verdict to stay consistent with the ClimateBERT benchmark.
- Do not invent facts. Score only what the text explicitly states.

Edits apply to the next run. ClimateBERT ignores this field (no prompt).