GPT-6 Sol Price Cut: New OpenAI API Rates Explained

OpenAI has dropped GPT-6 Sol to 2 dollars per million input tokens and 10 dollars per million output tokens — half what the previous 5.6-series Sol cost — and the gpt-6 sol price cut arrived alongside an even steeper reduction for GPT-6 Luna and a 90 per cent discount on cached input reads, per OpenAI's official API pricing documentation and its 22 September 2026 model announcement as reported by VentureBeat.

📺 Watch: OpenAI Just Dropped Two New GPT-6 Models

🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →

This is the biggest repricing OpenAI has shipped this year, and it landed in the same news cycle as Anthropic's cheaper Claude Opus 5.5, which is no coincidence: the frontier vendors are now openly competing on cost per task, not just capability. Below is exactly what changed, the full new rate card from OpenAI's own pricing page, why the company says it can afford the cut, and what it means if you are choosing a model to power agents and automations this quarter.

GPT-6 Sol Price Cut: The New Numbers

Here is the new schedule for the updated GPT-6 models, per million tokens in US dollars, as listed on OpenAI's API pricing documentation on 25 September 2026, with the previous 5.6-series rates that VentureBeat's report of the announcement cites for comparison.

Sol is a straight 50 per cent cut on both lines. Luna is 50 per cent on input and just over 58 per cent on output. Cached input reads on Sol now bill at 0.20 dollars per million — the 90 per cent discount against fresh input — with cache writes at 2.50 dollars. Long-context traffic on Sol runs 4 dollars in and 15 dollars out. And per the VentureBeat report, an OpenAI spokesperson confirmed these are permanent prices, not promotional or introductory pricing, so you can build unit economics on them rather than treating them as a launch sale.

Why OpenAI Says It Can Halve the Price

The company attributes the gpt-6 sol price cut to improvements in inference and caching rather than a margin squeeze. The concrete figure in the announcement coverage: measured across the billions of model requests flowing through GitHub Copilot, OpenAI's caching upgrades cut the volume of tokens needing fresh processing by more than half. When most of your fleet's traffic is cached re-reads billed at a tenth of the fresh rate, the blended cost of serving falls fast, and OpenAI is passing a slice of that downstream. For you, the mechanism matters as much as the number: it signals that cache-heavy workloads — which is what agent workloads are — are exactly the traffic these vendors are optimising and repricing for.

If you want ready-made agent workflows that cash in on cheaper models the week the cuts land, check out the AI Profit Boardroom → see what members are building. Want a personal plan for your site first? Book a free SEO strategy session and map it 1-on-1.

Where the GPT-6 Sol Price Cut Leaves the Rest of the Range

The pricing page makes the tiering explicit. GPT-6 Astra stays the premium option at 10 dollars input and 50 dollars output per million tokens, with 1 dollar cached reads — five times Sol's rates on every line. Sol is now the mid-tier default at 2 and 10. Luna is the volume tier at 0.10 and 0.50, with cached reads at a remarkable 0.01 dollars per million. In other words, OpenAI now offers a 100-to-1 spread between its priciest and cheapest GPT-6 input rates, and the full comparison of the two cheaper tiers — which one your workloads actually need — is broken down in the companion piece on GPT-6 Sol vs GPT-6 Luna.

📺 Watch: Claude Opus 5.5 + NEW GPT 6 Models!

How It Compares With Claude Opus 5.5

Anthropic released Claude Opus 5.5 on 22 September 2026 at 4 dollars input and 20 dollars output with 0.20 dollar cache reads — the same week, and exactly double Sol's new headline rates, with cache reads at parity. The two announcements read as a coordinated price war: Anthropic cut Opus-tier pricing by 20 per cent and OpenAI answered at the mid-tier with 50. Which model wins for a given job is not something a rate card can tell you — that is what task-level testing is for, and the Goldie Bench write-up covers how these frontier brains compare on hands-on agent tests. On the funnel's own pages, the Hermes Claude Opus 5.5 guide covers the subscription route on the Claude side, useful context for judging whether Sol's API pricing or a flat Claude plan serves you better.

📺 Watch: GPT-6 Sol + Luna Just Changed AI Agents

What Sol Is Actually For

Per the announcement coverage, OpenAI positions Sol at the complex work developers and knowledge workers perform repeatedly — building features, reviewing code, debugging and analysing data — while Luna handles high-volume, tightly defined jobs such as summarisation, extraction and answering straightforward questions. That division maps cleanly onto agent architecture: Sol-class models take the reasoning-heavy steps, Luna-class models take the bulk steps, and your orchestrator routes between them. If you are assembling that kind of stack, the Hermes agent OS overview shows the operating layer, and the best Hermes agent LLM guide ranks the candidate brains across price tiers, from frontier APIs down to local Ollama models that cost nothing per token at all.

Availability: Who Gets the Updated Models

The updated models are live in the API under the IDs gpt-6-sol and gpt-6-luna. Beyond the API, the reported rollout covers ChatGPT Work and Codex for Plus, Pro, Business and Enterprise subscribers, with free and Go users getting Luna through the desktop apps. If your usage lives inside a subscription product rather than the API, the price cut reaches you as capacity and model-quality improvements rather than a smaller bill — the direct saving accrues to API builders, which is one more nudge towards owning your own automation stack instead of renting outcomes through someone else's app.

What the GPT-6 Sol Price Cut Means for Your Automation Budget

Run the arithmetic on a concrete shape: an automation pipeline pushing 100 million input tokens and 20 million output tokens a month on Sol cost about 800 dollars on the old rates and now costs about 400. Add heavy caching — say 300 million cached reads at 0.20 dollars — and the marginal cost of a standing agent loop collapses to double digits. Cheaper tokens do not make anyone money by themselves, though. The gap between a lower bill and higher profit is workflow design: what you automate, what you sell, and how much of the saving you keep as margin. That build-out is what the Agent OS resource exists for, and pieces like the Qwen 3.8 Omni Flash capabilities review and the Kimi 2.6 benchmark cover the fast-moving budget end of the model market where these price cuts bite hardest.

How to Take Advantage of the New Rates This Week

If you already run on OpenAI's API, the saving arrives automatically — the updated rates apply to the gpt-6-sol and gpt-6-luna model IDs you are calling now — but capturing the full benefit takes three deliberate moves. First, re-run your cost model: any pricing you quoted clients against the old 4 and 20 dollar Sol rates now carries double the margin, and you decide whether to bank that or price sharper to win work. Second, audit which of your Sol calls are genuinely Sol-shaped; the price cut is the perfect excuse to move summarisation and extraction steps down to Luna, stacking a 20x tier saving on top of the 50 per cent cut. Third, lean into caching deliberately — stable system prompts and tool definitions are what turn the 90 per cent cached-read discount from a footnote into the biggest line on the saving. None of this requires new tooling, just an afternoon with your usage dashboard.

The Catch: Cheaper Is Not Free, and Sol Is Not Always the Answer

Two honest caveats before you re-platform everything. First, the fetched documentation lists a long-context surcharge — Sol at 4 dollars in and 15 out beyond the short-context window — so genuinely long-document workloads pay roughly double the headline rate, and prompt discipline still matters. Second, a 50 per cent cut on a model that was the wrong tool remains the wrong tool: extraction jobs overpay on Sol when Luna does them at a twentieth of the price, and hard reasoning that needs Astra or an Opus-class model will not become good on Sol because Sol is cheap. Route by task, then let the new rates compound the saving.

The bottom line: the gpt-6 sol price cut halves the going rate for mid-tier frontier intelligence to 2 dollars in and 10 out, permanently per OpenAI's statement, with Luna at a tenth of that and a 90 per cent cache discount underneath both. Whether you build on OpenAI, Anthropic or the open-weight challengers, the cost floor for serious AI automation just dropped — the builders who reprice their own services slowest keep the most margin.

If you want to be in the room where members turn model price cuts into new service margin — with the full Agent OS zip, prompt libraries and weekly live coaching — check out the AI Profit Boardroom → get inside. Or start with a personal roadmap: book a free SEO strategy session today.

Real wins from inside the AI Profit Boardroom

See all 3,000+ members →

← Back to all posts