Skip to main content
v2.14.2 is a maintenance release for Polyglot and Instructor. It moves the bundled OpenAI defaults off retired models and makes OpenAI Chat Completions tool calls work for both reasoning and non-reasoning models.

​
OpenAI Defaults

​
gpt-6-luna replaces gpt-5.4-mini and gpt-4.1-nano

OpenAI shut down gpt-4.1-nano on October 23, 2026. The bundled openai and openai-responses presets now default to gpt-6-luna, and all Polyglot docs, the cheatsheet, and examples that referenced gpt-4.1-nano or gpt-5.4-mini now use gpt-6-luna. Example cost calculations use the gpt-6-luna standard rates ($0.10 input, $0.01 cached input, $0.50 output per 1M tokens).

​
Model catalog records for current OpenAI models

The bundled model catalog adds exact records for gpt-6-luna, gpt-6-sol, gpt-6-astra, gpt-6.1-sol, gpt-6.1-terra, gpt-5.6-luna, gpt-5.6-sol, and gpt-5.6-terra on both the openai and openai-responses drivers, including context limits, modalities, and reasoning facts. The gpt-5.4-mini records were removed.

​
OpenAI Chat Completions Tool Calls

Current OpenAI reasoning models (gpt-6-*, gpt-5.6-*) reject function tools on /chat/completions unless reasoning_effort is none, while non-reasoning models such as gpt-4o-mini reject the reasoning_effort field entirely. Because Instructor uses tool calls by default, this made structured output fail on the new models. Model capabilities gain a toolsWithReasoning support fact, so the catalog states what each offering requires or forbids: unsupported on the openai records for the gpt-6-*, gpt-6.1-*, and gpt-5.6* models, supported on their openai-responses records. The openai/gpt-4o-mini record now declares no reasoning controls (reasoning: {selections: [], default: disabled}), so typed reasoning selections are rejected before the request is sent. The openai driver now sends reasoning_effort: none for tool calls only when the exact model record marks toolsWithReasoning: unsupported and reasoning can be disabled. With a model profile attached, preflight rejects tool requests that would keep reasoning enabled on such an offering. Non-reasoning and uncataloged models are sent unchanged, so withModel('gpt-4o-mini'), gpt-4.1-mini, and o3 keep working on the same preset. Explicit reasoning_effort options and typed withReasoning(...) selections are never overridden.

​
Upgrade Notes

  • Applications that relied on the bundled openai or openai-responses preset now call gpt-6-luna. Pin a model with withModel(...) or your own preset to keep a different default.
  • Unlike gpt-5.4-mini, gpt-6-luna reasons by default (medium) outside tool calls, and with reasoning enabled it rejects temperature values other than 1. Reasoning tokens count toward maxTokens, so very small limits (for example 64) can end with length. Raise the limit, or use withReasoning(ReasoningSelection::disabled()) to restore the previous non-reasoning behavior.
  • Models that cannot disable reasoning (gpt-6-astra, gpt-6.1-sol) and requests that combine explicit reasoning with tools need the openai-responses driver.
  • Tool calls on uncataloged reasoning models still need reasoning_effort: none from your own options or model record.