Sending work to GPT-6 Luna saves more than OpenAI's halving of Sol's price
OpenAI priced GPT-6 Sol at $2/$10 and Luna at $0.10/$0.50 per million tokens on September 22, half its GPT-5.6 promotional rates. Moving a job from Sol to Luna cuts its token rate by 95%, a bigger saving than the halving for any team whose work Luna can handle.
The Investor · Invest desk
What happened
- Under GPT-5.6, OpenAI charged $4 and $20 per million input and output tokens for Sol and $0.20 and $1.20 for Luna, according to its announcement.
- GPT-6 raises default cache hit rates and discounts cached input-token reads by 90%, with breakpoints and a diagnostics tool to control which prompt prefixes are reused.
- On Zapier's AutomationBench, OpenAI says Sol at xhigh effort scored 33.2% at $0.27 a task, ahead of Claude Opus 5's 26.9% at about nine times the cost per task.
- OpenAI says Sol makes about half as many factual errors as GPT-5.6 Sol on an internal test built from user-flagged mistakes, a sample it says is not typical usage.
- Anthropic released Claude Opus 5.5 the same day, saying it costs 40% less for typical workloads, and AI Weekly put OpenAI's launch about 90 minutes later.
Why it matters
- What doesn't add up contradiction Luna's output rate falls about 58%, not the 50% both reports describe, so Luna jobs that write more tokens than they read save more than the advertised half.
- What it costs cost Cached Sol input at $0.20 per million costs twice Luna's uncached rate, so a context-heavy agent's bill depends on its cache hit rate almost as much as on its tier.
- Who has to decide decision Deciding which jobs go to Luna falls to each team's own accuracy tests, since OpenAI's published scores cover Sol and its case for Luna is a qualitative claim of useful performance.
Next story · Product · Sep 30
OpenAI's 10-cent GPT-6.1 Sol price covers only cached inputConfirmed5 publishers
Both of Sol's rates fell by exactly half [4], so the per-token saving holds for any mix of input and output. Teams budget per task, though, and every cost figure OpenAI published is a per-task number taken at xhigh or max reasoning effort [6][7]. If GPT-6 Sol spends the same tokens on the AutomationBench task that GPT-5.6 Sol would have, that $0.27 task would have cost $0.54 at the old rates [21]. If it spends twice as many, the bill does not move [22]. The reports do not include token counts per task, and Gizmodo reported that customers have been complaining about token costs that are steep and unpredictable [16].
The halving is measured against GPT-5.6's promotional prices [3]. A budget written on those rates was already written on a discount. OpenAI's guidance that the new models are drop-in upgrades [17] invites teams to swap the model name without re-measuring what a task consumes.
Luna's evidence is thinner than Sol's. Every benchmark score in the two reports belongs to Sol [6][7]. For Luna, Crowdfund Insider reports only that OpenAI presents it as delivering useful performance at a much lower price than earlier mid-tier models [12], and the company aims it at summarization, information extraction and short-answer questions [11]. A team moving those jobs down a tier is relying on its own tests of Luna's accuracy on its own documents.
Caching changes the tier comparison for agent work that re-sends a long prompt each turn. Reasoning effort and tool availability can now change mid-conversation without breaking the cache [14], so an agent can raise effort for one hard step and keep its cached prefix. GitHub reports that over recent months OpenAI's caching changes cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests [15].
The competitor comparisons need a discount of their own. OpenAI took rival scores from public reports and used Fable 5 figures where Fable 5.1 numbers were unavailable [8]. The Opus 5 results Sol beat belong to a model Anthropic followed with Opus 5.5 on the same day [10].
In my view the budget line that moves most is the share of work a team can push to Luna, with the Sol halving second [18]. That view fails if little of a team's work fits Luna's brief of well-defined, high-volume jobs [11], or if Luna misses often enough on extraction and summarization that retries and escalations to Sol consume the tier discount [18].
What to watch
- Token counts per task for GPT-6 Sol at xhigh and max effort against GPT-5.6 Sol, from OpenAI or from customers' invoices.
- Any published Luna benchmark scores, or customer accuracy reports on Luna for extraction and summarization work.
- Per-task cost comparisons between GPT-6 Sol and Anthropic's Claude Opus 5.5, the model released the same day.
Information, not investment advice. This story is general information about companies and markets, drawn from cited reporting. It is not a recommendation to buy, sell or hold any security; check the cited sources before acting.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+20
- Incentives65
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
On September 22, 2026, OpenAI released GPT-6 Sol and GPT-6 Luna, placing them beneath the flagship GPT-6 Astra model released earlier in September.
- [2]
GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens; GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens.
- [3]
API rates for both models are 50 percent lower than the promotional prices of their GPT-5.6 counterparts.
- [4]
OpenAI said Sol drops from $4 to $2 per million input tokens and from $20 to $10 per million output tokens, while Luna falls from $0.20 to $0.10 on input and from $1.20 to $0.50 on output.
- [5]
OpenAI attributed the price reduction to improvements in caching and inference and said it is passing the savings to customers.
- [6]
On Zapier's AutomationBench 1.0.6, Sol at xhigh effort scored 33.2% at $0.27 per task, ahead of Claude Opus 5 at max effort (26.9%), at roughly a ninth of Opus 5's cost per task.
- [7]
On DeepSWE v1.1, Sol at max effort scored 68.8%, within 1.1 points of Claude Fable 5's 69.9%, at about 80% lower cost per task; on OSWorld 2.0 offline, Sol at xhigh effort scored 60.5% against Opus 5's 60.3% at medium effort, at about 80% less per task.
- [8]
OpenAI noted that competitor scores came from publicly available reports and that Fable 5 figures stood in where Fable 5.1 numbers were unavailable.
- [9]
OpenAI said Sol makes about half as many mistakes as GPT-5.6 Sol on an internal evaluation drawn from de-identified ChatGPT conversations in which users had flagged a factual error, and cautioned that those conversations are not representative of typical usage.
- [10]
Anthropic released Claude Opus 5.5 the same day, which it says costs 40% less for typical workloads than its predecessor; by AI Weekly's account, OpenAI's launch came roughly 90 minutes after Anthropic's.
- [11]
Sol is positioned for recurring complex work such as coding, debugging, building features, data analysis and agentic workflows; Luna is aimed at high-volume, well-defined jobs including document summarization, information extraction and short-answer questions.
- [12]
Luna is presented as delivering useful performance at a much lower price point than earlier mid-tier models.
- [13]
OpenAI said it has raised default cache hit rates for GPT-6, with discounts of 90% on cached input-token reads, and added explicit breakpoints plus a diagnostics tool to control and inspect which prompt prefixes are reused.
- [14]
Reasoning effort and tool availability can now be changed mid-conversation without breaking the cache.
- [15]
GitHub reports that over the past several months the caching changes cut the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models.
- [16]
Customers, individual and corporate, have been complaining about steep and unpredictable token costs, Gizmodo reported.
- [17]
OpenAI's guidance for organizations using previous Sol and Luna versions is that the new models are intended as drop-in upgrades.
- [18]
GPT-6 Luna costs one-twentieth of GPT-6 Sol per token on both input and output, so moving a job from Sol to Luna cuts its per-token rate by 95%.
- [19]
Luna's output rate falls about 58%, from $1.20 to $0.50 per million tokens, more than the 50% cut described.
- [20]
With the 90% cached-read discount, cached Sol input costs $0.20 per million tokens (twice Luna's uncached $0.10) and cached Luna input costs $0.01.
- [21]
If GPT-6 Sol used the same tokens as GPT-5.6 Sol, the $0.27 AutomationBench task would have cost $0.54 at the old Sol rates.
- [22]
If a GPT-6 Sol task uses twice the tokens its GPT-5.6 Sol equivalent did, its cost is unchanged.
Sources
2 independent publishers whose own reporting we read for this story.
- crowdfundinsider.comOpenAI Expands GPT-6 Lineup with More Economical Sol and Luna Models
1 article · September 26, 2026
- pivotnews.aiOpenAI halves API prices with new GPT-6 Sol and Luna models
1 article · September 25, 2026