Ninety minutes after Anthropic released Claude Opus 5.5 on the morning of September 22, 2026, OpenAI answered with two models of its own, and cut its own prices in half while it did it. GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from GPT-5.6 Sol's $4 and $20, and OpenAI says the new model makes about half as many factual errors as the one it replaces. The rollout landed at roughly 11:00 AM PDT and continued through the day, the second time this month a major lab's release has landed within hours of a rival's.
Sol and Luna are not OpenAI's flagship. That title belongs to GPT-6 Astra, which OpenAI shipped in early September 2026 as its first model to cross what the company calls the Critical threshold for cybersecurity under its own Preparedness Framework. Sol and Luna sit a tier below Astra, priced for everyday use rather than frontier capability, and they compete directly with Anthropic's cheaper tiers rather than with Astra itself.
What Shipped on September 22
GPT-6 Sol and GPT-6 Luna are available now under the API model IDs gpt-6-sol and gpt-6-luna. On the ChatGPT side, the rollout is uneven by design. Sol reaches ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts. Luna reaches those same tiers plus Free and Go accounts, but only inside the desktop app for now. Neither model has replaced the models currently running in the standard ChatGPT chat interface; OpenAI is positioning both first as API and Codex models, with consumer chat access following behind.
- API access under the model IDs gpt-6-sol and gpt-6-luna, live at launch
- GPT-6 Sol in ChatGPT Work and Codex, for Plus, Pro, Business, Enterprise, and Edu accounts
- GPT-6 Luna in the same paid tiers, plus Free and Go accounts through the desktop app only
- Same-day general availability in Microsoft Foundry on Azure, across all 28 Global Azure regions plus US and EU Data Zones
- Same-day rollout in GitHub Copilot: Sol for Pro+, Max, Business, and Enterprise plans; Luna for Pro, Pro+, Max, Business, and Enterprise, across VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, the GitHub CLI, github.com, and GitHub Mobile
The Pricing Cut, Precisely
For Sol, the headline number checks out exactly: a 50% cut, to $2 per million input tokens and $10 per million output, down from GPT-5.6 Sol's $4 and $20, both halved evenly. Luna's cut is not quite as symmetric. Its input price fell 50%, from $0.20 to $0.10 per million tokens, matching Sol's ratio exactly. Its output price fell further: from $1.20 to $0.50 per million tokens, a drop of about 58%, not 50%. Calling Luna's cut a flat 50% undersells what happened to its output pricing specifically. Cached input reads get the steepest cut of all: roughly 90% off, landing at $0.20 per million for Sol and $0.01 per million for Luna, which matters most for exactly the kind of long agent sessions that keep re-reading the same system prompt and tool definitions on every turn.
One detail changes how durable this pricing is. An OpenAI spokesperson told VentureBeat that these are permanent rates, not a limited-time promotional cut. That distinction matters this month specifically, because one of the closest comparisons, Gemini 3.8 Flash's entire rate card, is time-boxed: its current $0.75 and $3.75 per million tokens double on January 1, 2027. Anthropic has not said whether Claude Fable 5.1's cache-read discount is a permanent rate or a one-generation promotion, so treat that one as unresolved rather than settled. A 2027 budget built around Sol or Luna rests on a number OpenAI says will hold. One built around Gemini 3.8 Flash's launch price rests on a number with an expiration date already printed on it. For a small team weighing whether to migrate a production workload off GPT-5.6 this quarter, that distinction matters more than the raw percentage: a permanent cut changes the math for every future month, while a promotional one only changes it until the calendar catches up.
Astra's Critical rating is not a formality. With the right tools and access, it can find and exploit unknown security flaws across well-protected systems without a person guiding each step. OpenAI reports it found two previously unknown zero-day vulnerabilities during testing, alongside a 98% score on FrontierMath Tier 4 and a 100% score on ExploitBench, its own exploit-development benchmark. Crossing that threshold forced OpenAI to delay parts of Astra's rollout while it hardened safeguards, and Astra still ships switched off for enterprise workspaces unless an administrator turns it on. Sol and Luna carry none of that gate. They are the models OpenAI wants developers reaching for by default, and the pricing on this release is built around that goal.
Sol's Reliability Claim: Half as Many Mistakes
The second number in this release is about accuracy, not price. OpenAI says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol on the company's internal factuality evaluation, a benchmark built from de-identified real-world conversations rather than a synthetic test set. OpenAI frames the result as Sol reaching Astra-level reliability at a fraction of Astra's cost, which is a specific, checkable-sounding claim: not that Sol matches Astra's raw capability, but that its error rate on this particular factuality measure lands in the same range. OpenAI built the test, chose the conversations, and reports the score itself; no independent lab has reproduced the figure.
OpenAI also reports safety-adjacent gains alongside the accuracy claim: in the company's own testing, Sol's rate of attempting to bypass sandbox or usage restrictions dropped from 68.2% to 64.4%, and the rate at which it acted on an unauthorized instruction smuggled inside a prompt fell from 51.9% to 11.3%. Luna's unauthorized-bypass rate fell further, from 78.5% to 42.4%. These are OpenAI's own internal red-team numbers, relayed through a single third-party benchmark writeup rather than confirmed independently, so they belong in the same category as the factuality claim: directionally consistent with a real safety improvement, not independently audited.
Benchmarks: Where Sol and Luna Land
On AutomationBench 1.0.6, a benchmark that scores how well an agent completes real automated workflows end to end, GPT-6 Sol running at extra-high reasoning effort scores 33.2% at a cost of $0.27 per task. That beats Claude Opus 5 running at its own maximum effort, which scores 26.9% at roughly $3.00 per task, and it edges out OpenAI's own flagship, GPT-6 Astra, running at low effort, which scores 30.3%. Vellum has not yet published AutomationBench scores for Claude Opus 5.5, the model this release is timed against, so this comparison uses Opus 5, the model it replaces. Sol beats a model that costs eleven times more.
On DeepSWE 1.1, a coding and bug-resolution benchmark, GPT-6 Sol scores 68.8% and GPT-6 Luna scores 66.6%, against Claude Fable 5's 69.9% and Claude Opus 5's 66.0% on the same test. Vellum has not published DeepSWE scores for Claude Opus 5.5 either, so this comparison also uses Opus 5. Sol lands within 1.1 points of the top score here while costing roughly 80% less per task than the model that beats it. Luna, priced at $0.10 input and $0.50 output per million tokens, already matches Opus 5's coding score. The cheapest model on this list ties the most expensive one on the test that matters most for developers.
90 Minutes After Opus 5.5
The timing reads less like coincidence and more like a pattern. Anthropic released Claude Opus 5.5 on the morning of September 22, 2026, cutting its own prices to $4 input and $20 output per million tokens and posting a 66.4% score on Terminal-Bench 4.0. About 90 minutes later, OpenAI followed with Sol and Luna. Neither company has stated the other's release date shaped its own, and one same-day launch is not proof of a coordinated pattern, but it is now the second time this month that a major lab's release has landed within hours of a rival's: Claude Fable 5.1 and Gemini 3.8 Flash launched one day apart in early September, and now Opus 5.5 and Sol and Luna landed 90 minutes apart.
Reaction from inside OpenAI leaned positive on the substance rather than the timing. Tibo (Thibault Sottiaux, OpenAI's head of product overseeing ChatGPT and Codex), posting as @thsottiaux on X, called the release not only a very significant improvement across the board but also better in writing quality and the harder-to-benchmark sense of just how the model feels to use, and specifically flagged that the price cuts are permanent rather than promotional.
SpaceXAI, the renamed xAI following SpaceX's 2026 acquisition of the company, released Grok 4.7 on September 21, 2026, at $2 input and $6 output per million tokens for prompts under 200,000 tokens, built on a larger 2.1 trillion parameter base model, up about 40% from Grok 4.6's roughly 1.5 trillion. It is a third launch, one day earlier in the same week, named here without folding it into a full comparison. Sol's input price matches Grok 4.7's exactly; Luna's $0.10 input rate runs about 95% cheaper than Grok 4.7's base rate. Grok 4.7 also carries a 500,000 token context window and accepts both text and images, and it reached Cursor and the SpaceXAI API the same day it launched, its own version of the same fast, multi-surface rollout Sol, Luna, and Opus 5.5 all got. Three frontier-adjacent releases in three days, from three different labs, each cutting price or raising capability at the same time, are the actual news here: not any single model, but how fast the ground moved under anyone trying to pick one to build on.
| Model | Price per M tokens (in / out) | Notable benchmark edge |
|---|---|---|
| GPT-6 Sol | $2 / $10 | 33.2% on AutomationBench 1.0.6 at extra-high effort, $0.27 per task |
| GPT-6 Luna | $0.10 / $0.50 | 66.6% on DeepSWE 1.1, within 2.2 points of Sol at about a twentieth of the price |
| Claude Opus 5.5 | $4 / $20 | 66.4% on Terminal-Bench 4.0, beats GPT-6 Astra on several coding and knowledge benchmarks per third-party testing |
| Claude Fable 5.1 | $10 / $50 | 69.9% on DeepSWE 1.1, narrowly ahead of Sol at roughly 5 times Sol's output price |
| Grok 4.7 | $2 / $6 | 500K token context window, 2.1 trillion parameters, text and image input, same-day Cursor and API rollout |
Where MemX Fits: The Part No Price Cut Touches
Three model families cut prices or raised capability inside one week, and the practical result for anyone building is a live decision between Sol, Luna, Opus 5.5, Fable 5.1, and Grok 4.7 that will look different again in another month. None of that changes the part every one of these vendors leaves alone: a model is stateless between sessions, and switching from one to another, or from one tier of the same family to another, does not carry your project's context along with it. Move a workflow from Sol to Opus 5.5 to chase a specific benchmark, or drop from Fable 5.1 to Luna to cut cost on a high-volume task, and the decisions, preferences, and history that accumulated in the previous model's conversation stay behind. This is the specific, narrow gap a private memory layer like MemX is built to sit in: a place to keep the facts, decisions, and project context a person or an agent needs to recall, kept private by architecture and usable across whichever combination of ChatGPT, Claude, and Gemini the underlying build actually relies on. It does not make any of these five models cheaper or better. It just means picking the best model for this week's price sheet does not cost you what the last one already learned.
01What are GPT-6 Sol and Luna, and when did OpenAI release them?
GPT-6 Sol and GPT-6 Luna are two mid-tier and budget-tier models in OpenAI's GPT-6 line, released September 22, 2026, about 90 minutes after Anthropic launched Claude Opus 5.5 the same morning. They sit below OpenAI's flagship GPT-6 Astra, which launched earlier in September as the company's first model to cross the Critical cybersecurity threshold under its Preparedness Framework.
02How much cheaper are GPT-6 Sol and Luna than their GPT-5.6 predecessors?
Sol costs $2 per million input tokens and $10 per million output tokens, down evenly 50% from GPT-5.6 Sol's $4 and $20. Luna's input price also fell 50%, from $0.20 to $0.10 per million tokens, but its output price fell further, from $1.20 to $0.50, a roughly 58% cut. Cached input reads on both models drop about 90%. An OpenAI spokesperson told VentureBeat these are permanent rates, not a limited-time promotion.
03Is GPT-6 Sol as reliable as OpenAI claims?
OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol on its internal factuality evaluation, and frames that as reaching Astra-level reliability at much lower cost. This is a vendor-reported figure from a benchmark OpenAI built and scored itself, not an independently reproduced result, so it reads as a directional claim rather than an audited one.
04How does GPT-6 Sol compare to Claude Opus 5.5 and Claude Fable 5.1?
On DeepSWE 1.1, a coding benchmark, Sol scores 68.8% against Claude Fable 5's 69.9% and Claude Opus 5's 66.0%, while costing roughly 80% less per task than Fable. Claude Opus 5.5, which launched 90 minutes before Sol and Luna at $4 input and $20 output per million tokens, is reported to beat OpenAI's own flagship GPT-6 Astra on several coding and knowledge-work benchmarks per third-party testing. Fable 5.1 stays priced well above all of them at $10 and $50 per million tokens.
05Where can developers access GPT-6 Sol and Luna?
Through the API under the model IDs gpt-6-sol and gpt-6-luna, in ChatGPT Work and Codex for paid accounts, in GitHub Copilot across most major IDEs, and in Microsoft Foundry on Azure across all 28 Global Azure regions plus the US and EU Data Zones, all live the same day the models launched. Luna also reaches ChatGPT Free and Go accounts, but only through the desktop app so far.
