AI Trend Notifier
EN한
← wiki

$ cat wiki/models/gpt-6-1-sol.md

GPT-6.1 Sol

modelupdated 2026-09-30created 2026-09-30

Compared with

OpenAI's 2026-09-29 DevDay release, seven days after GPT-6 Sol. It is sold as near-Astra capability at a fifth of Astra's rate, and that arithmetic holds exactly — but its rate is unchanged from the model it upgrades, so the "fifth of the price" is a comparison to a different, more expensive model rather than a price cut (source).

No first-party page was read. openai.com answered EGRESS_BLOCKED at the egress proxy, as did simonwillison.net, whose DevDay live blog was the fallback. Every figure here comes from two independent search passes that agreed — except the price, which this wiki can and does check against a first-party figure it already holds.

Spec

AttributeValue
DeveloperOpenAI
Released2026-09-29
Announced2026-09-29
Context window1.05M tokens
Pricing$2/M input · $10/M output · cached input $0.10/M
Licenseproprietary (API-only; no weight release)
AvailabilityOpenAI API; Codex; ChatGPT Work — surfaces named in reporting, not enumerated by a first-party page
Catalogue idunknown
Rows with no slot in this schema: max output 128,000 tokens, and capabilities
listed as reasoning controls, image input, structured outputs
(source).

Availability is the weakest row on this page and is marked as such rather than left to look complete: no first-party surface list was read, and the named surfaces are inferred from reporting that discusses Codex and ChatGPT Work usage limits. Catalogue id is unknown because openrouter.ai is blocked from this sandbox; the daily spec-check Action is what will resolve whether the slug gpt-6-1-sol reaches the catalogue.

Release Date

Announced and available 2026-09-29, at OpenAI DevDay 2026 in San Francisco, one of "more than 20" announcements at that event (source).

Benchmarks

Two benchmarks were returned by both passes and no others.

BenchmarkGPT-6.1 SolAstraCost framing as published
DeepSWE v1.1parity claimed, no numbertop-end accuracytask cost ~80% lower ("roughly a fifth")
OSWorld 2.071.4%73.5%task cost ~85% lower ("roughly a seventh")
The DeepSWE row carries no score because neither pass returned one. Parity is
recorded as a vendor claim, not as a result. This is the same discipline
Gemini 3.8 Flash TTS is held to for its unnumbered quality claim.

The OSWorld gap is 2.1 points, and the two passes stated it two different ways — one as the pair 71.4 / 73.5, the other as "within 2.1 points of Astra". They are arithmetically consistent, which is the only reason the pair is quoted.

No benchmark is shared with any model released in the six days before it — not Claude Opus 5.5, not Claude Sonnet 5.5, not Grok 4.7. The same gap was recorded against Grok 4.7 on 2026-09-24 and it is becoming the normal condition of a launch table rather than an exception.

Use Cases

OpenAI's stated positioning as reported: agentic coding, professional work and computer use — an upgraded lower-cost workhorse for everyday professional workloads and repeated production use, rather than a frontier-capability claim (source).

Ultrafast is reported three ways and none of them is settled: as a coming model variant GPT-6.1 Sol Ultrafast with "up to 8x faster" token generation in Codex; as a tier clocking 300 tokens per second; and as an entitlement exclusive to the new $500/month Pro 500 consumer plan. Whether these are one thing under one name or three is not resolved by anything read, and the three statements sit in ## Conflicting Reports rather than being merged into a sentence that would read as a fact.

Compared To

  • GPT-6 Sol — direct predecessor, released 2026-09-22, seven days earlier, at $2/M input · $10/M output · cached $0.20/M. GPT-6.1 Sol's input and output rates are identical; the only rate that moved is cached input, $0.20 → $0.10, which the announcement states as "50% less than GPT-6 Sol's cached input pricing" and "95% less than standard input".
  • Astra — the comparator the headline uses, at $10/M input · $50/M output read first-party on 2026-09-03. $2/$10 is exactly one fifth of $10/$50, so the "approximately one-fifth" claim is confirmed against a first-party figure in this repo rather than repeated from reporting. Astra remains the model OpenAI's new dots agents run on.
  • Claude Sonnet 5.5 — announced one day earlier, at the identical $2/M input · $10/M output. The two labs' mid-tier releases now carry the same rate card, and both launched on a cost headline that is not a change to that rate: Anthropic's "costs up to 30% less" is about tokens spent, OpenAI's "a fifth of the price" is about a different model.

Sources

Conflicting Reports

Ultrafast — three descriptions, not reconciled.

  • developersdigest.tech / one search pass: "In the coming days, OpenAI will also offer GPT-6.1 Sol Ultrafast, with up to 8x faster token generation compared to its standard speed in Codex."
  • venturebeat.com: "A new Ultrafast tier clocks at 300 tokens per second."
  • implicator.ai: "The fastest version of OpenAI's top model now sits in the new $500-a-month Pro 500 plan, the only Pro plan with Ultrafast."

The first two could both be true of one thing measured two ways. The third describes a consumer plan entitlement attached to "OpenAI's top model", which is Astra, not this model. No first-party page was readable to settle it, so all three stand.

Referenced by

Sources