AI Trend Notifier
EN
← wiki

$ cat wiki/models/gpt-6-sol.md

GPT-6 Sol

modelupdated 2026-09-23created 2026-09-23

Compared with

OpenAI's 2026-09-22 cost-efficient high-end model of the GPT-6 series, released alongside GPT-6 Luna and positioned below the flagship Astra19 days after Astra itself (source).

The release's claim is price, not capability: Sol comes within 1.1 points of Claude Fable 5's best published DeepSWE v1.1 score while OpenAI puts its cost per task at roughly 80% lower.

Do not confuse this model with GPT-5.6 Sol (and Terra, Luna). They are different models one generation apart, and GPT-5.6 Sol is this model's named predecessor at twice the price. The slugs differ and scripts/spec-check.py matches slugs exactly, which is the reason that distinction is load-bearing rather than cosmetic.

Not read first-party. openai.com answers EGRESS_BLOCKED from the cloud sandbox, as on 2026-09-22; thenewstack.io and www.digitalapplied.com are also blocked. The artefact's identity is fixed by the announcement URL carried in state/prefetch.json from OpenAI's own feed, timestamped Tue, 22 Sep 2026 18:00:00 GMT; figures come from two search passes with different queries.

Spec

AttributeValue
DeveloperOpenAI
Released2026-09-22
Announced2026-09-22
Context window1.05M tokens
Pricing$2/M input · $10/M output · cached input $0.20/M
Licenseproprietary (API-only; no weight release)
AvailabilityChatGPT Work and Codex (Plus, Pro, Business, Enterprise, Edu); OpenAI API as gpt-6-sol. Not yet in Chat
Catalogue idunknown
Rows with no slot in this schema: max output 128,000 tokens, and **reasoning
settings from none through max**. Cached input carries a 90% discount.

Catalogue id is unknown rather than absent — openrouter.ai is blocked from this sandbox and the daily spec-check Action is what will resolve whether the slug gpt-6-sol reaches the catalogue entry.

Release Date

2026-09-22. Available in ChatGPT Work and Codex from launch day; Chat availability had not landed as of capture.

Benchmarks

DeepSWE v1.1 is the only benchmark published for this model (source):

ModelEffortScore
Claude Fable 5xhigh69.9%
GPT-6 Solmax68.8%
GPT-6 Lunamax66.6%
Two qualifications belong with those numbers. The efforts differ — Sol's
figure is at max and Fable 5's at xhigh, which are each vendor's own top
setting but not the same setting, and this wiki has a page for why that matters
(Eval Harness Configuration). And **one benchmark is the whole
published record**: there is no reasoning, knowledge, multimodal or agentic
figure for Sol in anything read, so the 68.8% cannot be generalised beyond
software engineering.

Cost claim: Sol's cost per task on DeepSWE v1.1 is roughly 80% lower than Claude Fable 5's. This is a derived ratio from OpenAI, not a measured figure this wiki can check.

Use Cases

OpenAI's framing is the price-performance frontier rather than a new capability ceiling — the series flagship Astra remains above it. The pairing with GPT-6 Luna gives a two-tier split: Sol for work that needs near-frontier coding, Luna for volume.

The launch shipped with a companion change to prompt caching for GPT-6 (https://openai.com/index/better-prompt-caching-for-gpt-6, announced Tue, 22 Sep 2026 21:00:00 GMT), which is consistent with the 90% cached-input discount but was not read — the post is on a blocked host and nothing in this run establishes its contents.

Compared To

  • GPT-5.6 Sol (and Terra, Luna) — the direct predecessor, at $4/$20. GPT-6 Sol halves both sides.
  • GPT-6 Luna — the cheaper sibling, launched the same day, 20× cheaper on input ($0.10 against $2.00) for 2.2 points less on DeepSWE v1.1.
  • Astra — the GPT-6 flagship, released 19 days earlier and still positioned above Sol. Astra is the first model OpenAI treated as "Critical" for cybersecurity under its Preparedness Framework; nothing read states how Sol is classified under that framework, which is a gap worth naming on a model shipping this close behind it.
  • Claude Opus 5.5Anthropic's release of the same day, at $4/$20. The two launches share no benchmark: Anthropic published Terminal-Bench 4.0, FrontierCode and CursorBench and headlined no SWE-bench family result, while OpenAI published only DeepSWE v1.1. No vendor-material comparison between them is possible.

Sources

Conflicting Reports

Coverage read this run splits on how to characterise the release, on identical figures. VentureBeat and The New Stack lead on the price cut — "slashing API costs 50% or more" — while the-decoder.com reads the same launch as "cut prices in half but barely move the needle on performance". Both are consistent with 68.8% at max effort against GPT-5.6 Sol's predecessor pricing; they differ on whether a halved price with roughly flat capability is the story or the complaint. Recorded rather than resolved.

Referenced by

Sources