OPENAI · MODEL RELEASE TRACKER
GPT-6 Luna
GPT-6 Luna is an OpenAI GPT-6 variant designed as a low-cost software agent for healthcare administrative workflows. It operates by interacting with existing user interfaces (mouse clicks and keystrokes), was integrated into the Agents API in late September, is available via the API and in ChatGPT for Healthcare/Clinicians, and scored 91.8% on the HealthAdminBench benchmark. OpenAI reported a cost of about 8.5 cents per task and US deployment including use at CVS Health.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
RELEASE TIMELINE
Published, source-backed release events only.
Anthropic launches Claude Haiku 5.5 with big price cuts and benchmark gains
Anthropic released Claude Haiku 5.5, a faster, lower‑cost small model aimed at high‑volume tasks; Anthropic says Haiku 5.5 costs ~75% less on average than Haiku 4.5 and up to 90% less for prompts ≤100,000 tokens. The company also halved Sonnet 5.5 cache‑read costs, added monthly API credits for subscribers, updated SDKs with beta computer/browser use, and says Haiku 5.5 outperforms its predecessor (and compares favorably to OpenAI’s GPT‑6 Luna) on several benchmarks and agentic coding tests.
OpenAI rolls out GPT-6 with Intelligent UI in ChatGPT
GPT-6 is rolling out globally inside ChatGPT and introduces an "Intelligent UI" that delivers faster responses along with visuals and interactive experiences you can explore and use directly.
OpenAI reopens $200 Pro plan but halves API credits per dollar
OpenAI is reopening its $200-per-month Pro subscription to new sign-ups while cutting the amount of API credits granted per dollar by half; the previous 5-hour usage cap for subscribers has been removed. Employee Thibault Sottiaux cited recent releases—GPT-6 Sol and GPT-6 Luna, reportedly released this week at half the API price of predecessors—and framed the changes as a shift toward pay‑per‑use pricing.
VERIFIABLE FACTS
Every value stays attached to a source and date.
Announced/launched on September 22, 2026.
Described as suited to clerical, high-volume tasks with clear goals — e.g., summarizing documents, extracting information, and answering quick questions.
Positioned to make focused, repeatable tasks practical at high volume, where small differences in latency and cost multiply across requests.
Features are available through the Agents API and in ChatGPT for Healthcare and ChatGPT for Clinicians.
US companies, including CVS Health, are reported to use the agents to process referrals and specialist consultations across separate systems.
Provided in the API under the name gpt-6-luna.
About $0.085 (8.5 cents) per task; reported as 98% less than GPT-6 Astra.
Reportedly priced at about half the token price of the GPT-5.6 Sol and Luna models (≈50% reduction reported).
Positioned for high-volume, well-defined, repeatable tasks — examples given include summarizing documents, extracting information, and answering short questions. Coverage describes Luna as making focused, repeatable tasks practical at high volume where small differences in latency and cost matter.
Operates clinic programs and insurance portals by interacting with existing user interfaces using mouse clicks and keystrokes.
One report states Luna 'improves on its predecessor by 5.4 percentage points while costing 58 percent less' on the evaluations referenced by that coverage; the excerpt does not name the specific benchmark.
Reportedly the Decisions API is built on GPT-6 Luna for fast, single decisions and classification tasks (i.e., Luna is used as the model for that API).
Some coverage characterizes Sol and Luna as largely matching their predecessors' performance while delivering significant price reductions (phrases used include 'barely move the needle on performance'), indicating that the primary change in these releases was cost efficiency rather than large capability gains.
The computer-use agents (including Luna) were integrated into the Agents API in late September, enabling them to operate existing UIs the way a human would.
GPT‑6 can produce responses while still 'thinking' (streaming), which OpenAI says reduces wait times by 44 percent in some cases.
In internal tests using the streaming method, GPT‑6 scored higher than its predecessor GPT‑5.6 on difficult web searches.
Accepts text and image input.
Designed to make focused, repeatable tasks practical at high volume.
Supports client-side tool use and prompt caching.
Accepts configurable reasoning-effort levels: none, low, medium, high, xhigh, and max.
Google states that customer data submitted to OpenAI models (including GPT-6 Luna) via Google Cloud will not be used for training.
OpenAI positions GPT-6 Luna as a fast, cost-efficient option for everyday use.
On OpenAI's MultiLangBench evaluation (maximum reasoning effort), GPT-6 Luna scored 40.3 percent.
In a third-party comparison published by Inception, GPT-6 Luna is reported to generate about 130 tokens per second on Nvidia GPUs (Inception's internal throughput comparison).
Inception's internal benchmarks report that their Mercury 3 model outperformed GPT-6 Luna in eight of 12 tests and led GPT-6 Luna by 15 points on IFBench in the company's comparison.
In experiments involving GPT-6 Luna, same-model interactions tended to narrow the opinion gap on average, while cross-model interactions tended to widen it; conversational influence depended on the participating models.
A blanket instruction to persuade did not reliably increase a model's influence in the LLM-LLM debate experiments involving GPT-6 Luna.
Scored 31 on the Coding Agent Index v2.0 (reported value: 31).
In the paper's dietitian-approved weekly planning evaluation, a locally hosted 14B planner incurred fewer rule violations than its GPT-6 Luna teacher, indicating GPT-6 Luna had higher rule violations in that evaluation.
GPT-6 Luna was evaluated on StateBench, a benchmark for long-context state tracking.
On StateBench, GPT-6 Luna "still breaks down on harder forms of state tracking as inputs grow longer."
In tests, reliably handled medical administrative tasks such as transferring data between electronic health records and external web portals.
Scored 91.8% on the HealthAdminBench benchmark; ran at maximum reasoning effort during the benchmark.
Built for low API pricing (described by OpenAI as a lower-cost agent option compared with Astra).
WHAT CHANGED