GPT-6 Sol: The Complete Reference for OpenAI's Mid-Tier GPT-6

GPT-6 Sol is the middle tier of the family — the model between the flagship GPT-6 Astra above it and the cheap, fast GPT-6 Luna below it, shipped on September 22, 2026 at $2.00 per million input tokens and $10.00 per million output tokens. That is half of what GPT-5.6 Sol cost for a model the vendor describes on its own documentation as "built to power complex coding and agentic workflows": 1,050,000-token context window, 128,000-token output ceiling, text and image input, and six levels of reasoning effort running from none to max.

This page is the reference for that model — the one you land on after typing its name. It is deliberately not a second announcement. The launch date is used here to date the model, not to report an event: the model is a live, generally available API product today, and what a reader arriving at this page needs is the rate card, the envelope, the vendor's claims clearly labelled, and what an independent evaluator measured. The datetime anchor is September 22, 2026, six days before this page was written, so the launch also sits inside a seven-day freshness window; the section you are reading is the reason the page is a reference rather than a news item. The reason it exists at all is search demand measured first-party on this site: people type the model's own name, and the loose traffic around that name is entirely versus-shaped and already served by the matchup pages we publish separately.

Where GPT-6 Sol sits in the GPT-6 stack

OpenAI's own documentation names three GPT-6 models and gives each one job. GPT-6 Astra is the flagship — the company's own guide calls it its most intelligent model, aimed at computer use, browsing, software engineering, science and document creation, and priced at $10.00 / $50.00 before any tiers. GPT-6 Luna is the efficiency tier, aimed at well-defined high-volume work, and costs $0.10 / $0.50. GPT-6 Sol is the middle: OpenAI's model page describes it in one line as built for complex coding and agentic workflows, and its family guide tells developers to pick gpt-6-astra for the highest level of capability, gpt-6-sol for "strong reasoning on demanding tasks", and gpt-6-luna for efficient repeatable work at scale.

Two facts make the tier real rather than a pricing label. Sol carries the same 1,050,000-token window and 128,000-token output ceiling as Astra — the middle tier is not a smaller window, it is fewer capability dollars per token. And Sol supports a reasoning setting Astra does not: reasoning.effort accepts none on Sol, so the model can be run without a reasoning pass at all, while Astra's ladder starts at low. For everything else the effort ladder is identical — none, low, medium (the default), high, xhigh, max.

Artificial Analysis dates GPT-6 Astra to September 3, 2026 and GPT-6 Sol and Luna to September 22, 2026 — three weeks that separate the flagship generation's first release from its cheaper siblings. Note that our own catalogue row for GPT-6 Astra carries a catalogue date of September 4; that is a catalogue artefact, and the independent tracker's September 3 is the release date we use here.

The rate card, in full

Every figure in this section is from OpenAI's own pricing documentation and the GPT-6 Sol model page, both read on September 28, 2026, and both per million tokens in USD.

• Standard, up to 272,000 input tokens — $2.00 input, $0.20 cached input, $2.50 cache writes, $10.00 output.

• Cached input ratio — OpenAI bills cached input at 10% of the uncached rate, a 90% discount, and states that cache writes are billed at 1.25× the uncached input rate. The GPT-6 launch also removed a friction that used to cost money quietly: changing reasoning effort or tool availability no longer invalidates the cache.

• Long context, above 272,000 input tokens — $4.00 input, $0.40 cached input, $5.00 cache writes, $15.00 output. The threshold is 272K, and it is stated on OpenAI's GPT-6 Sol model page itself rather than borrowed from the Astra page, which carries the same number. The same page states the repricing rule in one sentence: prompts with more than 272K input tokens are priced at 2× input and cache rates and 1.5× output for the full request. This is not an overage on the excess. A 300,000-token request bills all 300,000 tokens at the long-context rate, so the cost of crossing the line is a step change rather than a slope — and the step arrives at a boundary most long-document pipelines cross by accident.

• Batch and Flex — both are priced at 50% of the applicable standard rates, so Batch and Flex on Sol are $1.00 / $5.00 short context and $2.00 / $7.50 long context, with cached input at $0.10 and $0.20 respectively.

• Fast mode — 2× the applicable rate, which is $4.00 / $20.00 short context and $8.00 / $30.00 long context. OpenAI renamed priority processing to Fast mode on July 30, 2026 and accepts either service_tier: "priority" or "fast".

• Two uplifts worth knowing — regional processing endpoints carry a 10% uplift for models released on or after March 5, 2026, and FedRAMP endpoints carry a 10% uplift on standard rates. For GPT-6 Sol and Luna, EU data residency is available only with Standard processing.

• Context for the price — OpenAI's price sheet still carries GPT-5.6 Sol at $4.00 / $20.00 with promotional pricing stated to run at least through November 21, 2026. Sol's $2.00 / $10.00 is a 50% cut on both sides of its direct predecessor, and OpenAI told reporters the new rates are permanent rather than promotional. That is a spokesperson statement rather than a written commitment on the price sheet, so it is worth treating as vendor-stated rather than as a contractual fact.

The envelope

• Context and output — 1,050,000-token context window, 922,000 maximum input tokens, 128,000 maximum output tokens. The 1,050,000 figure is OpenAI's own; Artificial Analysis lists 872,000 on its spec table for the same model, and the FAQ on that page says 870k. Treat the vendor number as the contract and the tracker's number as a measurement of a particular served configuration.

• Modalities — text and image in, text out, per OpenAI's model page. Our own catalogue row for the model also lists a file input alongside text and image, and OpenAI's page does not. The vendor page governs, so the honest reading of this model today is text and image input.

• Knowledge cutoff — April 20, 2026. GPT-6 Luna's is May 18, 2026, and Astra's is April 30, 2026, so the three models in the family do not share a cutoff.

• Endpoints — Chat Completions and Responses both supported, and Batch supported. Realtime, Realtime translation and transcription, Assistants, fine-tuning, embeddings, image generation and editing, speech, audio transcription and translation, moderation, and the legacy Completions route are all unsupported.

• Tools and structured output — streaming, structured outputs, function calling, file search, image input, web search and prompt caching are all listed. There is one sharp edge: on Chat Completions, function calling works only when reasoning_effort is none. Reasoning with tools requires the Responses API.

• Unsupported parameters when reasoning is on — with reasoning effort set to anything other than none, remove temperature, top_p and top_logprobs; on Chat Completions also remove logprobs.

What OpenAI claims, and what kind of claim it is

OpenAI's launch material for GPT-6 Sol, as reported by The Decoder on the launch day, is best read as a cost-at-parity argument rather than a capability claim. The figures below are vendor-reported, have not been reproduced independently, and are the shapes OpenAI chose to publish; the benchmark set is narrower than the usual lineup, which the same coverage points out — GDPval and Terminal-Bench 4.0 are absent from it.

• FrontierCode 1.1, vendor-reported — GPT-6 Sol scores 49.3% at maximum effort for $2.14 per task. OpenAI places that roughly on par with Claude Fable 5.1 at 50.3% maximum effort, which costs $12.83 per task on OpenAI's own comparison.

• DeepSWE v1.1, vendor-reported — 68.8% for GPT-6 Sol at maximum effort, within 1.1 points of Claude Fable 5's best published 69.9%, and ahead of OpenAI's own GPT-6 Luna, which OpenAI says matches Sol at xhigh for far less.

• AutomationBench and OSWorld 2.0, vendor-reported — OpenAI reports GPT-6 Sol at maximum effort beating Claude Opus 5 at maximum effort on the business-workflow benchmark at 9% of the cost per task, and says GPT-6 delivers computer-use results similar to Claude Opus 5 at roughly 80% lower cost, with GPT-6 Astra still leading that category inside OpenAI's own family.

Read the shape of those sentences rather than the numbers. Each one pairs a score that is close to a rival's with a cost that is a fraction of it. That is a parity-at-lower-cost claim, not a dominance claim, and it is a materially weaker statement than "the best model" — which is what OpenAI reserves for Astra. The vendor claims here say the middle tier has closed the price gap; they do not say it has closed the capability gap.

The independent read

Artificial Analysis evaluated GPT-6 Sol at maximum reasoning effort and published the results on the launch day. Its figures are independent and are the counterweight to the paragraph above.

• Intelligence Index — 48 on the Artificial Analysis Intelligence Index v4.3.2, which folds in ten evaluations including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1. The page places it 18th of 211 models in its comparison class — a denominator that moves as the tracker adds models.

• Cost per task — $1.06 per Intelligence Index task, against a $1.54 blended rate per million tokens on Artificial Analysis's 7:2:1 cache-hit/input/output mix. The model generated 77 million output tokens during the index run, against a median of 88 million for reasoning models in its price tier, so it is not winning on brevity.

• Speed and latency — 88.7 output tokens per second measured through OpenAI's API, above the tracker's median of 82.6 for comparable reasoning models, with a time to first token of 138.06 seconds against a median of 3.79 seconds. Those two numbers describe the same trade: the model is fast once it starts and slow to start, which is the expected signature of a long reasoning pass.

• Where it went backwards — The Decoder's reading of the Artificial Analysis results reports regressions on two knowledge-work evaluations against the previous generation: roughly 100 Elo points lost on GDPval-AA v2.1 and about 75 on Luna, with manual inspection tracing the drop mainly to presentation quality and incomplete results. Artificial Analysis's own summary, as reported, is that per-task costs roughly halved while intelligence scores stayed at GPT-5.6 levels, with gains on some evaluations and regressions on others.

One caveat belongs on all of that. GPT-6 Sol and GPT-6 Astra have not been benchmarked head to head at matched reasoning effort on a matched harness by any party we can cite, and the two figures that get quoted side by side — Sol at 48 and Astra at 53 — come from the same index but not necessarily the same configuration. The honest version of that comparison is short: on the index, Astra scores above Sol, and no published run isolates how much of the gap is the model and how much is the effort setting.

Availability, and the one thing that changed after launch

OpenAI shipped both models on September 22, 2026 through its own API as gpt-6-sol and gpt-6-luna, and its launch material placed them in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers, with Free and Go users reaching GPT-6 Luna through the desktop app. Neither model was available in the regular ChatGPT Chat interface at launch — that is OpenAI's own framing as reported on the day, and it is the detail most likely to confuse a reader who went looking for Sol in the model picker and did not find it.

The post-launch change that matters for anyone feeding this model images is in OpenAI's own API changelog, dated September 25, 2026: OpenAI fixed a bug in image encoding that it says degraded image understanding in GPT-6 Sol and GPT-6 Luna, affecting visual tasks in the API and in Codex, including computer use. The changelog's own recommendation is to rerun evaluations and retry affected workflows if image inputs are part of your use case. That is a three-day-old fix on a six-day-old model, and it is the reason a benchmark run from launch week and a run from this week can disagree on anything visual.

Running GPT-6 Sol on OrcaRouter

GPT-6 Sol is routed on OrcaRouter as openai/gpt-6-sol on OrcaRouter, at OpenAI's own list price with provider list price passed through at 0% markup, alongside openai/gpt-6-astra and openai/gpt-6-luna. All three sit behind one OpenAI-compatible endpoint and one API key, which is what makes the tier question testable rather than theoretical: the same request body can be pointed at Sol for a production path and at Luna for the bulk job beside it, with no second contract and no code change, and a vendor price change is live here the same day rather than after a repricing cycle. If you are moving a workload onto a model this new, automatic failover across providers is the other half of the argument — it is how you try a six-day-old model on real traffic without betting a production path on its first month.

Two things the rate card does to a real bill, worked through:

• A 900,000-token document-analysis run bills at long context for the whole request: 900,000 input tokens at $4.00 per million is $3.60, plus output. The same run kept under 272,000 tokens bills at $2.00 per million. Splitting one oversized request into four sub-threshold calls is not a hack here; it is a straightforward 2× saving on input.

• A cached agent loop is where the 90% cached-input discount does the work: a stable prefix re-read hundreds of times costs $0.20 per million on Sol rather than $2.00, and because changing reasoning effort no longer invalidates the cache, the effort ladder can be walked inside one conversation without paying to rebuild the prefix.

If what you actually want is the two-model decision — this or the flagship — we have run the full comparison of GPT-6 Sol against GPT-6 Astra separately, and the pre-launch record of what was then an unreleased model is preserved in its own piece rather than re-litigated here.