danielrosehill/OR-Batch-Inference-Models

OpenRouter Batch Inference Models

Every model with a :batch id on OpenRouter on 2026-10-31 (71 models), with batch and sync prices, the providers that serve batch, and a price-tier label.

• Source: public https://openrouter.ai/api/v1/models and /api/v1/models/{id}/endpoints, fetched 2026-10-31T14:52:01Z. No key needed.
• Prices: US$ per million tokens, base tier, from pricing.prompt / pricing.completion × 1e6.
• Each snapshot is a split named snapshot_YYYY_MM_DD. New snapshots are added, never overwritten.
• Visualiser: danielrosehill/OR-Batch-Inference-Costs.
• Source notebook and scripts: danielrosehill/OpenRouter-Batch-Notes (rubric: notebook/07-price-tiers.md, tiering: scripts/price_tiers.py).

Prices change often. Re-check on the day you rely on them.

Data dictionary

Machine-readable copy: data_dictionary.json. Origin says where a value comes from:

• api: retrieved from OpenRouter's public API at fetch time, reformatted only (units, joins).
• enriched: added by us, either computed from API fields by a stated rule or a judgement label. OpenRouter does not publish these.

Column	Origin	Type	Meaning
snapshot_date	enriched	date	UTC date the API was fetched. Same for every row of a file.
model_listed_utc	api	date	Date OpenRouter first listed the model (created, Unix time to UTC date). Not the date batch support arrived.
model	api	string	Sync model slug (id with :batch removed).
batch_id	api	string	Batch slug (id) to submit batches with.
name	api	string	Display name (name).
context_length	api	integer	Context window in tokens (context_length).
input_modalities	api	string	architecture.input_modalities, ;-separated.
batch_input_usd_per_m	api	number	Batch input price, US$ per million tokens (pricing.prompt x 1e6, base tier).
batch_output_usd_per_m	api	number	Batch output price, US$ per million tokens (pricing.completion x 1e6, base tier).
sync_input_usd_per_m	api	number or null	Input price of the sync id for the same model. Empty if the sync id is not listed.
sync_output_usd_per_m	api	number or null	Output price of the sync id for the same model. Empty if not listed.
batch_providers	api	string	Providers serving the :batch id, from /api/v1/models/{id}/endpoints, ;-separated.
canonical_slug	api	string	canonical_slug: OpenRouter's permanent slug, usually the model name plus a release-date stamp.
hugging_face_id	api	string or null	hugging_face_id: weights repo for open-weight models. Empty for closed models.
expiration_date	api	date or null	expiration_date: announced retirement date. Empty if none announced.
tier	enriched	string	Cheap / Mid / Expensive by batch output price: < $3, $3-20, >= $20 per million. Rule: notebook/07-price-tiers.md, scripts/price_tiers.py.
tier_rank	enriched	integer	1 = Cheap, 2 = Mid, 3 = Expensive.
sub_band	enriched	string	1a <= 1.00, 1b 1-3, 2a 3-8, 2b 8-20, 3a 20-50, 3b >= 50 (batch output $/M).
input_band	enriched	string	Cheap < 0.60, Mid 0.60-3.00, Expensive >= 3.00 (batch input $/M).
input_mismatch	enriched	boolean	True when input_band differs from tier: the model changes tier on input-heavy jobs.
out_in_ratio	enriched	number	batch_output_usd_per_m / batch_input_usd_per_m, 1 decimal.
reference_job_usd	enriched	number	Cost in US$ of 4.5M input + 0.5M output tokens at batch prices (bulk-labelling job shape, 9:1).
stable_endpoint	enriched	boolean	True if the id resolves to one fixed model version. False for dynamic aliases (~ prefix or latest in the id, which OpenRouter says 'always redirects to the latest model') and for closed models whose canonical_slug has no date stamp (an upstream alias the provider can repoint). Rule: scripts/build_hf_dataset.py.
stability_basis	enriched	string	Which branch of the stable_endpoint rule applied.
offers_cache_pricing	enriched	boolean	True when the batch id's API record publishes a prompt-cache read price (pricing.input_cache_read). False means no cache price is listed, not proof the provider never caches.
cache_pricing	api	string (JSON) or null	Cache prices from pricing, US$ per million tokens, as compact JSON: read_usd_per_m (input_cache_read), write_usd_per_m (input_cache_write, null when not published: usually automatic caching with no write premium), and long_context {min_prompt_tokens, read_usd_per_m, write_usd_per_m} when a long-context override lists cache prices. Empty when offers_cache_pricing is False.

Tier boundaries sit in the widest gaps of the 2026-10-31 output-price ladder (2.50 → 3.75, 7.50 → 10, 15 → 25, 37.50 → 60).

Endpoint stability

stable_endpoint says whether the id always resolves to the same model version. OpenRouter's dynamic aliases (~google/gemini-flash-latest, ~anthropic/claude-sonnet-latest and 18 others on 2026-10-31) silently move to each new release. None of them has a :batch id, so every alias is excluded from batch by construction. Three batch models are still labelled unstable: openai/gpt-4o, openai/gpt-4o-mini and openai/gpt-4-turbo. Their canonical_slug carries no date, so they map to OpenAI's undated names, which OpenAI can repoint to a newer snapshot (it did so for gpt-4o in 2024). That they are aliases is inferred from the undated slug. OpenRouter does not state it. The other 68 have a dated canonical_slug (66) or are open weights with a fixed Hugging Face repo (2, gpt-oss).

Prompt caching

On 2026-10-31, 60 of 71 batch ids publish a cache read price (offers_cache_pricing). Read price as a share of batch input price: 0.025× (Fable 5.1), 0.05× (Opus/Sonnet 5.5, GPT-6.1, Gemini 4 Argon, GPT-6.1 Sol), 0.10× (most models), 0.16–0.25× (Grok 4.3, GLM, o3/o4-mini, GPT-4.1) and 0.50× (GPT-4o, o3-mini). A write price is published for 31: 1.25× input for Anthropic (15) and OpenAI GPT-5.6 Sol, GPT-6 and GPT-6.1 (12), 0.11–0.19× for Gemini 3.6–4 (4). The other 29 list no write price (automatic caching, no write premium). No cache price: gpt-oss-20b/120b, Qwen 3.7 Plus, Mistral Medium 3.5, Gemini 3 Flash / 3.1 Pro preview, GPT-4 Turbo and the GPT-5 / 5.2 / 5.4 / 5.5 Pro models. OpenRouter's batch guide says only that "prompt caching rates vary by model"; whether requests inside one batch actually hit each other's cache is not confirmed.

Compiled by Claude Opus 5.5 for Daniel Rosehill, 2026-10-31.