GPT-6.1 Sol: OpenAI Ships Near-Astra Performance at the Price of the Model It Replaces

Open​AI released GPT-6.1 Sol on September 29, 2026, at its DevDay 2026 keynote — a refresh of the high-end GPT-6 Sol tier that shipped one week earlier, priced at exactly the same $2 per million input and $10 per million output tokens, and positioned by the vendor as approaching GPT-6 Astra, the $10/$50 flagship, at roughly a fifth of its per-token cost. The unusual part is not the price. It is that the same-price, one-week-later refresh is claiming a step change in capability, and that as of this writing no independent evaluator has scored it. Every capability number in this piece that comes from Open​AI is labelled as such, because there is nothing else to check it against yet.

What actually shipped, and when

The release is unusually well documented for a DevDay announcement, which is itself worth noting: within hours there was a model page, a pricing table, a system-card addendum, and a committed entry in the vendor's public API schema — a commit titled "Add gpt-6.1-sol model to model enums" at 17:16 UTC on September 29, where the parent revision contained the string zero times. The vendor's own changelog, however, still ends at the September 22 release of GPT-6 Sol and GPT-6 Luna; the 6.1 Sol entry has not been written up there yet, so the model page and the announcement post are the record.

The concrete shape of it:

• Model ID: gpt-6.1-sol, one snapshot, no dated variant — you are always on the current build.
• Availability: today to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and through the API. Explicitly not in the consumer Chat product, which Open​AI states rather than leaves to be inferred.
• Context: 1,050,000 tokens, 128,000 maximum output, knowledge cutoff April 30, 2026.
• Reasoning control: low, medium (default), high, xhigh and max — five levels, not six. GPT-6.1 Sol and GPT-6 Astra no longer accept the none effort that GPT-6 Sol and GPT-6 Luna do.
• Modalities: text and image in, text out. Audio and video input are not supported.
• Tools: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search — all through the Responses API. Chat Completions works only without tools.
• Not supported: fine-tuning. This is a model you prompt, not one you train.
• Announced, not shipped: an Ultrafast tier. Open​AI says it will offer GPT-6.1 Sol Ultrafast "in the coming days," up to 8x faster token generation in Codex, with no price and no date attached.

Two rows deserve a second read. The five-level effort ladder means a prompt that used none as a cheap escape hatch will fail validation on 6.1 Sol, and the vendor publishes migration guidance for exactly this move from gpt-6-sol. And the Ultrafast line is worth reading carefully: the same tier is already broadly available for GPT-6 Astra and in preview for GPT-5.6 Sol, so 6.1 Sol is being described as an upcoming addition to a tier that exists for other models — not as a new tier.

The pricing, and where it actually bites

The headline rate is unchanged from GPT-6 Sol: $2.00 per million input tokens and $10.00 per million output tokens, with cache writes at $2.50. The vendor's own framing in the announcement is a comparison to its flagship, where $2/$10 is one-fifth of GPT-6 Astra's $10/$50 — which is a real gap and also the least interesting number on the sheet, because GPT-6 Sol already cost $2/$10.

What did change is the cached-input line, and it is the one that moves budgets. GPT-6.1 Sol caches at $0.10 per million tokens against GPT-6 Sol's $0.20 — half the price, and 5% of the uncached input rate. For anything that reuses a long prefix across requests, the effective input cost of that prefix just halved on top of the model getting better, and that is the number to model if you are pricing an agent that carries context between turns.

The multipliers, from the vendor's pricing page:

• Standard: $2.00 input / $0.10 cached / $2.50 cache write / $10.00 output.
• Fast mode: exactly 2x standard — $4.00 / $0.20 / $5.00 / $20.00. Not available with EU data residency.
• Batch and Flex: 50% of standard — $1.00 / $0.05 / $1.25 / $5.00.
• Regional processing: a 10% premium where available; US and EU residency are both supported on this model.
• Long prompts: anything over 272,000 input tokens is repriced for the whole request at 2x input and cache rates and 1.5x output — $4.00 / $0.20 with $15.00 out. The output side rising 1.5x rather than doubling is the softer edge on this family's long-context step.

The long-context row is the one to design around. The context window is a million tokens, but the price break sits at 272,000 — so the practical ceiling on a single request is a budget decision, not a window decision, and a document pipeline that drifts past it pays double on the entire request rather than on the excess.

What OpenAI claims, and what nobody has measured yet

Every performance number published for GPT-6.1 Sol so far is vendor-reported and unreproduced. Open​AI's announcement describes gains on five evaluations:

• DeepSWE v1.1 (complex software engineering in real codebases) — matches GPT-6 Astra at roughly one-fifth the cost, and beats GPT-6 Sol's best score by 6.4 percentage points at a lower reasoning effort.
• GDP.pdf (professional questions over complex PDFs) — scores higher than Claude Opus 5.5 with fallbacks at less than half the cost per task, and approaches Astra at roughly one-fifth the cost per task.
• AutomationBench 1.0.6 (47 tools, multi-step business workflows) — 2.2 percentage points above Claude Opus 5.5 at medium effort at about a third of the cost, and up 4.8 points from GPT-6 Sol at the same setting.
• OSWorld 2.0 offline set (computer use) — seven points above GPT-6 Sol at maximum effort for less than half the cost, and within 2.1 points of Astra at roughly one-seventh its cost per task.
• Terminal-Bench Science 0.1 (data analysis, simulation, theorem proving) — more than doubles GPT-6 Sol's score at maximum effort; $5.47 per task on average against $23.21 for Opus 5.5 and $23.80 for Astra. Open​AI notes Astra still leads the set at 68.1%.

The two numbers here that are not capability scores are the more interesting ones, because they are failure-rate claims rather than leaderboard claims. On factuality, Open​AI reports the share of responses containing at least one factual error falling from 11.4% to 7.7% at low reasoning effort — a reduction of about 32% — and says the error rate stays within 1.9 points of Astra across tested settings. On tool reliability, the share of cases where the agent fails to tell the user its search tool is broken falls to 2.1%, against 4.9% for GPT-6 Sol, 1.5% for GPT-6 Astra and 28.7% for GPT-6 Luna.

Both come with the vendor's own caveats, and they are unusually specific ones: the factuality set is de-identified conversations where users had flagged an earlier model's error, and the tool-disclosure tasks were selected to elicit failures. Neither is a claim about typical usage, and Open​AI says so. Treat them as the vendor's description of its own reliability work, not as measured facts.

The independent picture is thinner than you would expect for a flagship-adjacent launch. Artificial Analysis — the evaluator whose Intelligence Index the rest of this family is usually quoted against — has no GPT-6.1 Sol entry: its model URL returns 404 and the string does not appear anywhere in its leaderboard data, which does carry GPT-6 Sol, GPT-6 Astra and GPT-6 Luna across their reasoning configurations. The closest available anchor is the model this one replaces: Artificial Analysis scores GPT-6 Sol at 48 on its Intelligence Index in the max-effort configuration, at $1.05 of evaluation cost per task, generating 77 million tokens across the run, at 78 tokens per second. Until 6.1 Sol has a comparable row, the honest summary of the capability claim is that Open​AI says it is a large step and no one has yet published a number that agrees or disagrees.

The name, and the model that never arrived

The GPT-6 line reads as three tiers — Astra at the top, Sol beneath it, Luna at the bottom — plus one ghost. That ghost is GPT-6.1 Astra, the planned flagship refresh that the Wall Street Journal reported on September 28 had been scrapped, with frontier training paused. We covered that decision in its own piece rather than re-litigating it here. What matters for reading this release is the shape it leaves behind: a 0.1 version bump to the mid-tier carrying the improvements that would otherwise have gone to a flagship, with the model page's own positioning line stating the intent plainly — near-Astra performance for complex work at a lower cost. GPT-6.1 Sol is not an iteration on last week's Sol so much as the replacement for a model that never shipped.

Where this sits for us, honestly

OrcaRouter does not route GPT-6.1 Sol yet. The public catalogue has no entry for it, and we are not going to describe a model we cannot serve as though we could. What we do serve is the model it replaces: GPT-6 Sol on OrcaRouter costs the same $2.00 input / $10.00 output list price with both context tier rows published, alongside GPT-6 Astra above it and GPT-6 Luna below.

The reason that matters for a same-price refresh rather than a price cut: because we pass provider list price through with no markup added, the day GPT-6.1 Sol lands in the catalogue it costs what Open​AI charges for it, with no repricing lag on our side — and because one API key reaches every provider we carry, switching a workload onto it is a model-ID change, not a new contract. For a substitution where the incumbent and the replacement cost exactly the same, that is the difference between a one-line edit and a migration project. Until then, the sensible move on an unscored model is to keep it off a production path and use failover to a proven sibling when the primary errors — which is also the cheapest way to find out whether the vendor's claims hold on your workload.

What to watch over the next two weeks

Three things would settle the questions this release leaves open, and none of them are in yet. First, an independent score — a single Artificial Analysis row for GPT-6.1 Sol would turn the largest claim in the announcement into something checkable, and it is the number most likely to move procurement decisions. Second, the Ultrafast price: Open​AI has promised the tier "in the coming days" for Codex with up to 8x generation speed, but an announced serving tier with no number attached is a promise, not a product, and Fast mode's 2x-standard rate is the only anchor for guessing what it will cost. Third, whether the 11.4%-to-7.7% factuality improvement survives contact with a harness that is not Open​AI's — because that, and the search-tool disclosure rate, are the two claims that would justify moving production traffic rather than a leaderboard line.

The honest summary is that GPT-6.1 Sol is a real release with a real price sheet, a documented API surface, and a set of vendor numbers that are larger than a 0.1 bump usually carries. The price argument is easy: it costs the same as the model it obsoletes, and its cached input costs half as much. The capability argument is entirely Open​AI's until someone else measures it — which is the whole reason to read the spec sheet now and the benchmarks in a week.

GPT-6.1 Sol is not on our routes yet. When it is, it will appear in the full model catalogue at Open​AI's own list price with no markup added, next to the 200+ models already behind one API key.