What an LLM Actually Is, Explained Through the Two Models OpenAI Just Shipped

Published: September 29, 2026 | Category: AI | By Mahesh

If you have never sat down and had someone explain what actually happens when you type a question into ChatGPT, that gap in understanding is worth closing before diving into any specific product launch, because it changes what a headline like "OpenAI ships two new models" actually means. A large language model, the technology behind ChatGPT, Gemini, Claude and nearly every chatbot you have used this year, is not a database that looks up answers, and it is not a search engine wearing a friendly interface. It is a statistical system trained on enormous volumes of text to predict, one fragment at a time, what word is most likely to come next given everything written before it. Ask it a question, and it is not retrieving a stored answer; it is generating, token by token, the sequence of words its training makes most probable as a response, informed by patterns it absorbed from books, articles, code and conversation at a scale no human could read in a lifetime.

That single mechanism, next-token prediction at enormous scale, is responsible for everything from drafting an email to writing working code to holding a coherent conversation across dozens of turns. It also explains why these systems can be simultaneously impressive and occasionally wrong in ways that feel strange: they are not reasoning the way a person does, they are producing the statistically likely continuation of a pattern, refined through training to be useful, accurate and aligned with what a person actually wants far more often than not. Understanding that is the difference between treating a chatbot's output as gospel and treating it as a genuinely capable but fallible collaborator, and it is the right lens for looking at what OpenAI actually shipped this week.

Two Models, One Very Deliberate Trade-Off

On September 22, 2026, OpenAI released GPT-6 Sol and GPT-6 Luna, filling out the GPT-6 family it opened three weeks earlier with the flagship GPT-6 Astra [1]. Sol is positioned as a mid-tier model built for interactive and agentic coding, the kind of multistep work that involves building a feature, reviewing a pull request, debugging an error and analyzing a dataset in the same session. Luna sits below it, a lightweight, high-volume model meant for fast, narrow tasks: summarizing a document, extracting a data point, answering a quick factual question [2]. Both carry a 1.05 million token context window, meaning either can hold roughly the equivalent of a very long novel's worth of text in working memory during a single conversation [3].

The pricing is the headline, and it is genuinely steep. Sol costs $2 per million input tokens and $10 per million output tokens through the OpenAI API; Luna costs $0.10 and $0.50 respectively. Both figures are exactly half of what OpenAI had been charging for the GPT-5.6 Sol and GPT-5.6 Luna models they replace [4]. For context, Sol's rate roughly matches what Anthropic charges for Claude Sonnet 5, while Luna undercuts Anthropic's cheapest model, Claude Haiku 4.5, by a factor of ten on both input and output pricing [5]. If you build products on top of AI models rather than just using a chat interface, a halving of your per-token cost is not a marginal update. It can be the difference between a feature being commercially viable at scale and not.

"Token prices only tell part of the story. A model can have a lower rate but use more reasoning tokens. OpenAI's charts report cost per task, which captures both effects."

MetricNexus, independent analysis of OpenAI's own launch benchmarks [6]

The Honest Part of This Story: Capability Didn't Uniformly Improve

This is where the launch gets genuinely interesting, and where most coverage of it stops short. OpenAI's own published benchmark charts, when read carefully rather than skimmed for the headline number, show something that undercuts the simple "newer is better" story every AI company's launch post implies. On two of the three coding and computer-use benchmarks OpenAI itself published, GPT-5.6 Sol's best score is actually higher than GPT-6 Sol's [7]. On the DeepSWE v1.1 software engineering benchmark specifically, GPT-5.6 Sol at maximum effort still outperforms GPT-6 Sol at every effort level OpenAI tested, though at more than double the cost per completed task [8].

That is not a knock against the release so much as a genuinely useful correction to how these launches usually get reported. The real gain here is economic, not purely intellectual: on four of five benchmarks OpenAI published, Sol's cost per completed task fell by 58 to 59 percent, even on the charts where its raw accuracy score dipped slightly [9]. In practical terms, that means a team running thousands of coding tasks through the API each day gets roughly the same finished output for close to half the spend, which is arguably a more consequential improvement for a business than a few extra percentage points on a benchmark leaderboard, even if it makes for a less exciting headline.

Luna tells a more flattering story on pure capability. At maximum reasoning effort, Luna scores 66.6 percent on DeepSWE v1.1, edging past Sol running at high effort while costing a small fraction of the price, and landing within seven points of the far more expensive flagship Astra model [10]. On Zapier's AutomationBench, a benchmark measuring how well a model handles multistep workflow automation, Luna improved roughly 5.4 points over its predecessor, and the gap on that same test widens to 41.4 percent versus 32.0 percent for Sol compared with GPT-5.6, one of the largest jumps anywhere in the launch [11]. The pattern that emerges is that OpenAI's cheapest model, run at its most aggressive reasoning setting, is now doing meaningful work that used to require a considerably more expensive one, a genuinely useful finding for any developer deciding which tier to default to for a given task.

Why This Particular Week Matters

The timing here is not incidental. OpenAI's launch landed roughly ninety minutes after Anthropic shipped Claude Opus 5.5 on the same day, and nineteen days after OpenAI's own flagship GPT-6 Astra debuted [12]. That kind of near-simultaneous release from the two largest labs in the field is no longer a coincidence worth remarking on for its rarity; it has become close to the expected rhythm of this industry in the second half of 2026, where a competitor's roadmap decisions visibly shape the timing of a rival's own announcements. xAI's Grok 4.7, priced at $2 input and $6 output per million tokens below a 200,000-token prompt threshold, and rising to $4 and $12 above it, sits in a similar competitive band to Sol, giving developers a genuine three-way choice at roughly comparable price points for the first time this generation [13].

OpenAI has also been notably explicit about alignment work accompanying this release, a detail that reads differently against the backdrop of a difficult summer for the company on the safety front. The emphasis on alignment testing alongside the performance and pricing figures is not incidental messaging; it is a direct response to public scrutiny the company faced over model behaviour earlier this year, and it signals that safety evaluation has become as much a part of a model launch's marketing as its benchmark scores, at least for OpenAI specifically [14].

What This Actually Means If You Use These Tools

For anyone building a product on top of an AI model rather than just chatting with one, the practical takeaway from this launch is to stop defaulting to the most expensive tier out of habit. Luna at maximum effort is now genuinely competitive with much pricier models on several real-world tasks, and testing your specific workload against it before assuming you need Sol, let alone the full flagship Astra, could meaningfully cut costs without a corresponding drop in output quality. For coding-heavy workloads specifically, the benchmark data suggests a sensible split: use Luna as a fast first-pass filter for routine diffs, linting and mechanical fixes, and reserve Sol or Astra for the architectural decisions and substantive patches that actually need deeper reasoning.

For everyday ChatGPT users rather than developers, the more relevant fact is simply that OpenAI's paid tiers now route more of your usage through cheaper, faster models by default, which should translate into snappier responses for routine questions without materially changing what the assistant can do for anything genuinely complex. The underlying trend across this entire industry, prices falling roughly in half every few months while raw capability improves more unevenly, is worth internalising as the actual shape of progress here: not a single dramatic leap in intelligence every release, but a steady compounding of what a given dollar of AI usage can actually buy.

What to Watch Next

Two things are worth tracking over the coming weeks. First, whether OpenAI eventually ships a GPT-6 Terra to complete a three-tier structure matching its GPT-5.6 predecessor lineup; as of this launch, no such model has been announced, and the GPT-6 family currently stops at three tiers total [15]. Second, watch third-party benchmark trackers like Artificial Analysis for independent output-speed measurements on Sol and Luna, since neither had a confirmed tokens-per-second figure published at launch, which matters for latency-sensitive applications in a way raw accuracy scores do not capture. The pattern worth keeping an eye on across the rest of 2026 is whether this half-price, split-benchmark trade-off becomes the industry's default release shape, or whether the next major launch from any of the big three labs breaks that pattern with a genuine, uniform capability jump rather than another round of cheaper-but-mixed results.

Sources

1. CodersEra, "GPT-6 Sol and Luna: Complete Guide, Pricing and Benchmarks," September 22, 2026. codersera.com
2. Eigent, "GPT-6 Sol and Luna Explained: Pricing and Uses," September 2026. eigent.ai
3. CodersEra, "GPT-6 Sol and Luna: Complete Guide" (see source 1).
4. PowerDrill, "GPT-6 Sol and Luna: What's New, Pricing, and Alternatives (2026)," September 24, 2026. powerdrill.ai
5. DigitalApplied, "GPT-6 Sol and Luna: API Prices, Benchmarks and Trade-offs," September 22, 2026. digitalapplied.com
6. MetricNexus, "GPT-6 Sol and Luna: Benchmarks, Pricing, and Cost Cuts," September 2026. metricnexus.ai
7. DigitalApplied, "GPT-6 Sol and Luna: API Prices, Benchmarks and Trade-offs" (see source 5).
8. MetricNexus, "GPT-6 Sol and Luna: Benchmarks, Pricing, and Cost Cuts" (see source 6).
9. MetricNexus, "GPT-6 Sol and Luna: Benchmarks, Pricing, and Cost Cuts," cost-per-task figures (see source 6).
10. Vellum, "GPT-6 Sol and Luna Benchmarks Explained," September 2026. vellum.ai
11. MetricNexus, "GPT-6 Sol and Luna: Benchmarks, Pricing, and Cost Cuts," AutomationBench figures (see source 6).
12. CodersEra, "GPT-6 Sol and Luna: Complete Guide," launch timing (see source 1).
13. CodersEra, "GPT-6 Sol and Luna: Complete Guide," Grok 4.7 pricing comparison (see source 1).
14. Eigent, "GPT-6 Sol and Luna Explained," alignment and safety context (see source 2).
15. CodersEra, "GPT-6 Sol and Luna: Complete Guide," GPT-6 Terra status (see source 1).

Article by Mahesh | Depth Grid