OpenAI Models in Borg: GPT-5.5, GPT-5.6 and GPT-6

Borg includes OpenAI GPT-5.5, GPT-5.6, GPT-6, and gpt-oss-120b models. Their tiers target different tasks, and gateway speed figures are observations rather than guaranteed response times.

Models and model pages

The catalog lists GPT-5.5; GPT-5.6 Luna, Terra, and Sol; GPT-6 Astra, Luna, Sol, and 6.1 Sol; and gpt-oss-120b. See GPT-5.5, GPT-5.6 Luna, GPT-5.6 Terra, GPT-5.6 Sol, GPT-6 Astra, GPT-6 Luna, GPT-6 Sol, GPT-6.1 Sol, and gpt-oss-120b.

Model names point to different tiers and roles.

What the names mean

GPT-5.5 is a flagship model for difficult professional work, not a small throughput tier. GPT-5.6 Luna is the cost-first option for repeated focused requests, Terra is the middle tier, and Sol is the flagship. The unsuffixed GPT-5.6 alias points to Sol.

GPT-6 Astra is the most capable tier for difficult end-to-end work. Sol emphasizes coding and agent work, while Luna is the efficient option for focused, high-volume tasks. GPT-6.1 Sol is a newer variant for complex work, distinct from GPT-6 Sol.

The task, not the name alone, should guide your choice.

The catalog says these models accept text and image inputs and generate text, with adjustable reasoning effort. A large context capacity alone does not establish which model will do the best work.

The open-weight option

gpt-oss-120b is the larger open-weight member of the gpt-oss family. Its weights are downloadable. It accepts text and produces text, with configurable reasoning effort and agent-style tool capabilities.

The catalog says its Apache 2.0 license permits customization and local deployment. The smaller 20b sibling favors low latency, according to its model description. OpenAI does not publish a standard hosted API token rate for gpt-oss-120b.

gpt-oss-120b has downloadable weights.

Measured gateway speed

The figures quoted below use the Borg telemetry snapshot recorded at 2026-10-08T03:42:10.773437+00:00, covering a 30-day observation window from SearchAtlas AI Gateway production request logs. They are dated observations; the linked model pages show the current published measurements and each metric’s sample count.

Borg telemetry reports production gateway requests over a 30-day window. Time to first response event measures how long the gateway takes to receive the first streamed event. Completion tokens per second measures the full turn, including queueing and reasoning, not the rate at which each token streams to the screen.

GPT-5.5 recorded a median 8,716 ms to first event and 24.9 completion tokens per second. GPT-5.6 Luna measured 3,998 ms and 17.2 tokens per second; Terra measured 5,138 ms and 18.7; Sol measured 4,614 ms and 14.3.

These are observed medians, not guaranteed response times.

GPT-6 Astra measured 4,780 ms and 16.0 tokens per second. Luna measured 3,324 ms and 15.3; original Sol measured 5,288 ms and 11.3; GPT-6.1 Sol measured 3,194 ms and 17.7. gpt-oss-120b measured 6,121 ms and 46.8 tokens per second. Measurement counts differ, so comparisons need care. GPT-5.5 had 1,158 first-event measurements, while GPT-5.6 Luna had 22,631. These are first-event sample counts, not the completion-speed denominators; check the linked model sections for metric-specific counts.

Coding bench results

This snapshot includes coding-bench results for GPT-5.6 Luna and GPT-6 Astra. Luna passed 73.33% of measured cases; Astra passed 66.67%. Hidden pass rates were 73.33% for both.

Each result used only 15 samples across 3 tasks.

The hard-graded bench used tests and ruff, with no language-model judge, and ran September 20, 2026. It is a small sample, not a general measure of coding quality. No results are listed here for the other OpenAI models.

Picking a tier

For focused, repeated tasks, compare Luna with Terra. For top-tier work, try GPT-5.6 Sol or GPT-6 Astra. The catalog describes GPT-6 Sol for coding and agent work, and GPT-6.1 Sol as its newer successor. Consider gpt-oss-120b when downloadable weights, customization, or local deployment matter.

Borg Fusion can route work to model lanes, or a user can select a lane directly for a fixed model choice.

FAQ

Which OpenAI model is the GPT-5.6 flagship?

GPT-5.6 Sol is the flagship member of its family and the target of the unsuffixed GPT-5.6 alias.

What is gpt-oss-120b?

It is the larger open-weight member of OpenAI's gpt-oss family, with text input and output and configurable reasoning effort.

Which OpenAI models have published coding-bench results?

Borg telemetry lists results for GPT-5.6 Luna and GPT-6 Astra, based on a small set of hard-graded tasks.