Model catalog

Living map of model string → emission class used by carbonmd-factors-2026-08.

This page is steered weekly (Sunday evening, Europe/Zurich): new public releases and models seen in agent usage are classified, documented here, and wired into src/core/factors.ts.

Last steered: 2026-10-04<br> Factors version: carbonmd-factors-2026-08

How to read this

• Classes feed the gCO₂e bands in Methodology & factors.
• Matching is heuristic and honest — unknown strings fall to medium + guessed: true.
• Small markers win first (mini, flash, luna, fast, lightning…), so tiered families classify correctly.

Frontier

High-capability flagships. Central 4.5 gCO₂e / 1k output tokens.

Large

Workhorse coding / agent models. Central 2.5 gCO₂e / 1k output tokens.

Small

Cheap / fast tiers. Central 0.15 gCO₂e / 1k output tokens.

Decision note (2026-08 steer): grok-composer-2.5-fast is classified small, not large. The composer … fast naming marks xAI's cheap/fast tier, and the explicit composer + fast rule wins over the generic grok → large family rule. This is a deliberate behavior change vs the pre-2026-08 catalog; it lowers the central estimate for that model from 2.5 to 0.15 gCO₂e / 1k output tokens.

Decision note (2026-08-16 steer): grok-4.6 stays large, not frontier. OpenRouter describes it as xAI's smartest coding/STEM model; the catalog still treats the whole grok family as the large workhorse band, consistent with grok-4.3 / grok-4.5. qwen3.8-max / qwen3.8-2.4t-a95b stay large for the same family-rule reason (qwen → large), matching qwen3.7-max. seed-2-1-turbo is large (coding/agent workhorse); lite/mini Seed IDs remain small.

Decision note (2026-08-23 steer): z-ai/glm-5.3 stays large (glm family, reasoning / agent workhorse). meta/muse-spark-1.2-contributor stays large — contributor is not a small marker; the muse family rule wins even though the contributor SKU is cheaper than Spark. Tencent hy-mt2* (Hunyuan-MT translation specialists, including 30b-a3b with 3B active) is classified small. deepseek-v4-flash-vision-exp is small via flash.

Decision note (2026-08-30 steer): stealth/ox-alpha is classified small, not medium/guessed. Z.ai unmasked it on 2026-08-26 as GLM-5.3-Flash (320B-A18B). The flash marker already maps glm-5.3-flash / qwen3.8-flash-next to small; the explicit ox-alpha rule keeps the stealth ID aligned with that SKU. glm-5.3 (non-flash) stays large. Factor bands unchanged.

Decision note (2026-09-06 steer): gpt-6-astra / gpt-6-astra-pro are classified frontier, not medium/guessed. OpenAI's GPT-6 flagship began a phased rollout on 2026-09-03 (API IDs also listed as openai/gpt-6-astra). The gpt-6 rule is checked after small markers, so a hypothetical gpt-6-mini / luna still lands small. claude-fable-5.1 and claude-mythos-5.1 stay frontier via existing fable / mythos. gemini-3.8-flash (and the Fairwind-gated Cyber variant) is small via flash. muse-spark-1.3 / muse-spark-1.3-contributor stay large via muse. qwen3.8-max-0902 stays large via qwen. ibm-granite/granite-4.2-8b is small via 8b. inception/mercury-2.5-preview is classified small (diffusion LLM, vendor-compared to Luna / Flash-Lite / Haiku). tencent/hy4-preview is large (Hunyuan 4, 770B-A49B coding flagship) — the hy-mt translation family remains small; hy4 is a different product. Factor bands unchanged. No new uncatalogued IDs in Hermes usage this week. Decision note (2026-09-13 steer): deepseek-v4.1-flash / deepseek/deepseek-v4.1-flash / API alias deepseek-flash are small via flash. DeepSeek launched V4.1-Flash on 2026-09-10 (MIT weights; older V4 Flash / Flash Vision API names now route here). gemini-3.8-flash-cyber stays small via flash. inclusionai/ling-3.0-flash-fin is small via flash. meituan/longcat-2.0 (and :free) is classified large, not medium/guessed — Meituan LongCat-2.0 is a 1.6T-A48B coding workhorse, same band as Kimi / DeepSeek-Pro / Qwen-Max. Factor bands unchanged. Decision note (2026-09-20 steer): Moonshot's kimi-k2.8-preview is large through the existing kimi family rule; its Kimi Code rollout also retains the kimi-for-coding endpoint. Shanghai AI Lab's Atria-Dawn-Preview is large, not medium/guessed: it is a 744B agentic-MoE release built on the GLM-5.2 lineage, so an explicit atria rule aligns it with GLM/Kimi/DeepSeek workhorses. Hermes usage also contains upstage/solar-pro4:free; Upstage describes Solar Pro 4 as its 512K-context agentic LLM, and the explicit solar-pro rule maps it large. The complete gpt-5.2-codex, gpt-5.3-codex, and gemini-3.1-flash-lite-preview IDs were documented as already-covered family variants. Factor bands unchanged.

2026-10-04 steer: releases, catch-up and provenance

The previous catalog refresh was 2026-09-20. This steer covers 2026-09-28 through 2026-10-04, plus missing releases since that refresh.

• GPT-6.1 Sol ↗ (September 29) → frontier through existing sol / gpt-6 rules. Hermes also recorded gpt-6.1-sol. Cheaper API pricing alone does not justify changing an emission band.
• Claude Sonnet 5.5 ↗ (September 28), API claude-sonnet-5-5 → large through sonnet; also seen in Hermes.
• Gemini 4 Argon ↗ (September 30) → frontier through an explicit Gemini + Argon product-name rule. Access is restricted to trusted Fairwind cyber defenders. The spelling gemini-4-argon is a catalog label, not a confirmed public API ID or a claim of general availability.
• Index-Translate ↗ → small: 2B / 9B already match parameter markers; the 35B-A3B-preview text-translation checkpoint gets a narrow explicit rule for its 3B-active MoE, consistent with Hy-MT2-30B-A3B. Local inference is still outside these cloud-token estimates.
• Catch-up: GPT-6 Sol / Luna ↗ and Claude Opus 5.5 ↗ (September 22) → frontier / small / frontier through existing rules. Preserve the Hermes alias claude-opus-5-5[1m] literally; [1m] does not change its class.
• Catch-up: Grok 4.7 ↗ (September 21) → large, as for prior Grok flagships; seen in Hermes. MiMo-V2.6 ↗ (September 22, vendor date) → large for Pro / Pro-RL / Pro-UltraSpeed (42B active), small for Flash / Flash-RL through flash. UltraSpeed is a serving mode, not the whole-token ultra marker.
• Pareto 26.10 Preview ↗ (October 1) uses the unchanged API ID pareto. It remains medium + guessed: routing/backend composition is not disclosed sufficiently to infer an emission class from price or benchmarks.

All Hermes model strings checked in this run have a known classification. Emission bands, input-token weight and factors version are unchanged. New mappings only; no historical event rewrite.

Decision models: tracked, not validated by the token formula

Cloudflare Clef / Clef-flash ↗, Perplexity pplx-decider-v1-27b ↗ and Strands strands-decider-2b ↗ were also reviewed. They return decisions/probabilities rather than autoregressive text. Their generic heuristic results (e.g. flash / 2b → small) are not validated emissions factors for decision inference; do not substitute option counts for output tokens. No new decision-specific bands are introduced. Audio/video-generation releases are outside this text-token catalog's supported accounting.

Medium (guessed)

Anything without a known marker. Central 0.8 gCO₂e / 1k output tokens, range widened in status.

pareto is now documented here as medium + guessed, pending backend provenance; it was not found in Hermes usage.

If your production model lands here, open an issue or wait for the weekly steer.

Weekly steer

Every Sunday 20:00 Europe/Zurich, Hermes:

1. Scans public release notes + Hermes session_model_usage for new model IDs
2. Proposes class mappings (frontier / large / medium / small)
3. Updates factors.ts + this page + methodology examples
4. Rebuilds and deploys docs.carbonmd.dev ↗
5. Commits and pushes changes to both public and private GitHub repositories, then reports what changed

No silent factor-band rewrites: class values (gCO₂e table) only change with an explicit factors version bump and human review.

Related

• carbon-md factors — print the active table from the CLI
• Methodology — derivation and token rules
• GitHub issues ↗ — suggest a mapping