Skip to content

Model registry

The registry lists every model route ExtractIt knows: where it processes data, how long the provider keeps it, what it costs and what it can do. Policy resolution reads only this data (resolve() in extractit.llm.resolution), so a policy such as eu_only or zdr_required can only choose among the rows below, and fails closed when none qualifies.

Registry facts depend on (provider, model, route, contract) and change often. Every row carries the date it was verified and its source; review them whenever models are updated. The values here are what the public API gives without a contract: Zero Data Retention (zdr) and regional processing are agreements between a deployment and the provider, so cloud rows are registered conservatively (non-ZDR) until a deployment records otherwise.

OpenAI-compatible servers (vLLM, llama.cpp...) are not listed: their capabilities are discovered at run time by extractit.llm.probe and added as local routes (locality: local, retention_class: local, zdr: true, free).

Registry verified on 2026-11-10.

Reading the table

  • Retention: zdr zero data retention in force; 30d-provider the provider may keep inputs and outputs (abuse monitoring, up to 30 days or as stated in the notes); subscription handled under a consumer/plan account's data settings; local never leaves the machine.
  • Locality: cloud data leaves the deployment; local it never does (needed by air_gapped).
  • Structured output: native provider-side schema enforcement; tool forced tool call; emulated JSON by prompt plus validation and repair; none.
  • Prices are USD per 1M tokens (input / output); - means billed by plan, not per token. The cost filter of a policy uses an estimate for 1 000 pages (1 000 input and 150 output tokens per page by default).

Models

Route Tier Regions ZDR Retention Locality USD/1M in / out Structured Vision PDF Logprobs Batch Context
anthropic:claude-fable-5-2 - global, us no 30d-provider cloud 8 / 40 native yes yes no yes 1,000,000
anthropic:claude-haiku-5-5 fast global, us no 30d-provider cloud 0.1 / 0.5 tool yes yes no yes 1,000,000
anthropic:claude-opus-5-5 strong global, us no 30d-provider cloud 4 / 20 native yes yes no yes 1,000,000
anthropic:claude-sonnet-5-5 standard global, us no 30d-provider cloud 2 / 10 native yes yes no yes 1,000,000
codex:codex-default - global no subscription cloud - native yes no no no 200,000
gemini:gemini-3.1-pro-preview strong global no 30d-provider cloud 2 / 12 native yes yes no yes 1,048,576
gemini:gemini-3.5-flash-lite fast global no 30d-provider cloud 0.3 / 2.5 native yes yes no yes 1,048,576
gemini:gemini-3.8-flash standard global no 30d-provider cloud 0.75 / 3.75 native yes yes no yes 1,048,576
openai:gpt-6-luna fast us no 30d-provider cloud 0.1 / 0.5 native yes yes no yes 1,050,000
openai:gpt-6.1-sol strong us no 30d-provider cloud 2 / 10 native yes yes no yes 1,050,000
openai:gpt-6.1-terra standard us no 30d-provider cloud 1 / 5 native yes yes no yes 1,050,000

Sources and notes

anthropic:claude-fable-5-2

  • Verified on 2026-11-10; source: https://platform.claude.com/docs/en/about-claude/pricing
  • Cached input: 0.2 USD per 1M tokens
  • The temperature parameter is not sent to this model
  • The Messages API of this SDK takes no temperature parameter: it is not sent. Covered Model: the docs require 30-day retention, so ZDR is not available. Forced tool_choice is rejected (400): structured output goes through output_config.format. No EU option. No logprobs.

anthropic:claude-haiku-5-5

  • Verified on 2026-11-10; source: https://platform.claude.com/docs/en/about-claude/pricing
  • Cached input: 0.01 USD per 1M tokens
  • The temperature parameter is not sent to this model
  • The Messages API of this SDK takes no temperature parameter: it is not sent. Prices for prompts up to 100k tokens (above: 0.50 in / 2.50 out). Docs: content not retained by default and ZDR is an organisation agreement; registered as non-ZDR (fail closed). inference_geo 'us' costs 1.1x; there is no EU option. No logprobs.

anthropic:claude-opus-5-5

  • Verified on 2026-11-10; source: https://platform.claude.com/docs/en/about-claude/pricing
  • Cached input: 0.2 USD per 1M tokens
  • The temperature parameter is not sent to this model
  • The Messages API of this SDK takes no temperature parameter: it is not sent. Forced tool_choice is rejected (400): structured output goes through output_config.format. Docs: content not retained by default and ZDR is an organisation agreement; registered as non-ZDR (fail closed). inference_geo 'us' costs 1.1x; there is no EU option. No logprobs.

anthropic:claude-sonnet-5-5

  • Verified on 2026-11-10; source: https://platform.claude.com/docs/en/about-claude/pricing
  • Cached input: 0.1 USD per 1M tokens
  • The temperature parameter is not sent to this model
  • The Messages API of this SDK takes no temperature parameter: it is not sent. Forced tool_choice is rejected (400): structured output goes through output_config.format. Docs: content not retained by default and ZDR is an organisation agreement; registered as non-ZDR (fail closed). inference_geo 'us' costs 1.1x; there is no EU option. No logprobs.

codex:codex-default

  • Verified on 2026-11-10; source: https://developers.openai.com/codex
  • ChatGPT subscription through the Codex CLI: billed by the plan, handled under the account's data settings, never zero-retention. For local/personal use; policies with ZDR or air-gapped exclude it.

gemini:gemini-3.1-pro-preview

  • Verified on 2026-11-10; source: https://ai.google.dev/gemini-api/docs/pricing
  • Cached input: 0.2 USD per 1M tokens
  • Preview model; prices for prompts up to 200k tokens (row read from a summarised page: re-check). Gemini API (AI Studio key): no regional guarantee, prompts may be retained for abuse monitoring.

gemini:gemini-3.5-flash-lite

  • Verified on 2026-11-10; source: https://ai.google.dev/gemini-api/docs/pricing
  • Cached input: 0.03 USD per 1M tokens
  • Gemini API (AI Studio key): no regional guarantee, prompts may be retained for abuse monitoring, and on the free tier used to improve Google products. Output price includes thinking tokens.

gemini:gemini-3.8-flash

  • Verified on 2026-11-10; source: https://ai.google.dev/gemini-api/docs/pricing
  • Cached input: 0.075 USD per 1M tokens
  • Introductory price until 2026-12-31; from 2027-01-01 it doubles (1.50 in / 7.50 out). Gemini API (AI Studio key): no regional guarantee, prompts may be retained for abuse monitoring.

openai:gpt-6-luna

  • Verified on 2026-11-10; source: https://developers.openai.com/api/docs/pricing
  • Cached input: 0.01 USD per 1M tokens
  • The temperature parameter is not sent to this model
  • Abuse-monitoring logs up to 30 days; ZDR and EU residency (eu.api.openai.com, +10%) need prior approval from OpenAI. Cached input 0.01. Prompts over 272K tokens cost 2x input and 1.5x output. Logprobs and temperature support not verified.

openai:gpt-6.1-sol

  • Verified on 2026-11-10; source: https://developers.openai.com/api/docs/pricing
  • Cached input: 0.1 USD per 1M tokens
  • The temperature parameter is not sent to this model
  • Abuse-monitoring logs up to 30 days; ZDR and EU residency (eu.api.openai.com, +10%) need prior approval from OpenAI. Cached input 0.10. Prompts over 272K tokens cost 2x input and 1.5x output. Does not accept reasoning effort 'none'. Logprobs and temperature support not verified.

openai:gpt-6.1-terra

  • Verified on 2026-11-10; source: https://developers.openai.com/api/docs/pricing
  • Cached input: 0.05 USD per 1M tokens
  • The temperature parameter is not sent to this model
  • Abuse-monitoring logs up to 30 days; ZDR and EU residency (eu.api.openai.com, +10%) need prior approval from OpenAI. Cached input 0.05. Prompts over 272K tokens cost 2x input and 1.5x output. Logprobs and temperature support not verified.

How a policy uses this data

Policy field Rejects a route when Reason code
allowed_providers its provider is not listed provider_not_allowed
air_gapped it is not local not_local
eu_only not every region is in the EU (a local route passes) region_not_eu
zdr_required zdr is false zdr_unavailable
(the call's needs) it lacks structured output, vision, native PDF, logprobs or context capability_missing
max_usd_per_1k the estimated cost per 1 000 pages is above the limit over_budget

Checks run in that order and the first failing reason is the one recorded. The remaining routes are ranked by (index in preference, input + output price, id): no randomness, so the same policy and registry always give the same route and the same policy_hash.