GPT-6 Astra vs Sol vs Luna: Price, Release Dates and Codex Usage

The short answer

GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, GPT-6.1 Sol costs $2 and $10, and GPT-6 Luna costs $0.10 and $0.50. That makes Sol one-fifth of Astra and Luna one-twentieth of Sol. Astra shipped on September 3, 2026, Sol and Luna on September 22, and the 6.1 upgrade to Sol on September 29. In Codex on a Plus plan, OpenAI estimates 5 to 45 local messages per five hours on Astra, 15 to 160 on GPT-6.1 Sol, and 350 to 3,000 on Luna.

The practical reading is a ladder with a clear job for each rung. Use Luna for narrow, repeated work with a clear goal: field extraction, classification, structured summaries. Use GPT-6.1 Sol as the default for coding agents and long-running work, since OpenAI says it nearly matches Astra on agentic coding, computer use and professional work. Keep Astra for the tasks where a wrong answer costs more than the extra $8 per million input tokens, and for the first planning pass on hard problems.

Everything below comes from OpenAI’s pricing page, model pages, launch posts and help-center article, checked on October 28, 2026, plus Artificial Analysis for the one independent cost-per-task comparison. Performance claims are labeled vendor-reported. The deep benchmark tables live in two earlier guides, the GPT-6 Astra complete guide and the GPT-6.1 Sol guide, so this page summarizes and links instead of repeating them.

Release timeline: Astra, then Sol and Luna, then 6.1 Sol

People search for the GPT Astra release date and the Astra release order because the rollout was staggered, and the first announcement was not the same day as general availability. OpenAI announced Astra on Thursday, September 3, 2026. Its launch post says Astra was “rolling out today to a limited set of organizations” and would reach all ChatGPT Plus, Pro, Business and Enterprise users, the API, Azure and AWS Bedrock “over the coming days.” Press coverage from Reuters and TechCrunch carries the same September 3 date.

The staggering is documented in Sam Altman’s post of September 4: Astra was then available to all Pro, Enterprise and Business Premium users in Work and Codex, and in the API, with Plus and Business to follow. OpenAI’s help center describes the launch as a “broader Astra launch delay” and lists three compensating resets: banked resets on September 3 and September 4 for eligible subscribers, and an automatic reset on September 7. The sibling guide on what to do when Codex runs out of usage covers how resets work, so this page does not repeat it.

Eighteen days later, on Tuesday, September 22, OpenAI added the cheaper tiers. The Sol and Luna announcement says GPT-6 Sol and GPT-6 Luna launched in ChatGPT Work and Codex that day for Plus, Pro, Business, Enterprise and Edu users, with Free and Go users getting Luna in the desktop app. Neither was in Chat at launch. The API model IDs are gpt-6-sol and gpt-6-luna. The Verge published its report at 7:13 PM UTC the same day.

The third step came on Tuesday, September 29, the day of OpenAI DevDay. GPT-6.1 Sol is an upgrade to GPT-6 Sol with the same $2 and $10 token prices but a cheaper cached rate, $0.10 against $0.20. It is live in Work, Codex and the API as gpt-6.1-sol, and still not in Chat. OpenAI said an Ultrafast version of 6.1 Sol, with up to 8x faster token generation in Codex, would follow “in the coming days.” The DevDay roundup lists what else shipped that day.

• July 30, 2026: GPT-5.6 Luna price cut by 80% to $0.20 input and $1.20 output (OpenAI).
• September 3: GPT-6 Astra announced; limited organizations first, then phased rollout.
• September 4: Astra open to Pro, Enterprise and Business Premium in Work and Codex, and to the API.
• September 22: GPT-6 Sol and GPT-6 Luna launch in Work, Codex and the API.
• September 29: GPT-6.1 Sol launches at DevDay at the same token price as GPT-6 Sol.
• October 26: GPT-6.1 Terra launches in Work, Codex and the API at $1 input and $5 output per million tokens, between Sol and Luna.
• October 27: Artificial Analysis publishes Intelligence Index v5 with Astra and 6.1 Sol near the top.

OpenAI’s own launch video for developers is embedded below. It covers Astra only, since it predates Sol and Luna. The main launch film and the full DevDay keynote are separate OpenAI uploads.

The master comparison

The table below is the one to bookmark. Prices are OpenAI’s Standard rates per million tokens from the pricing page and each model page. The Codex column is OpenAI’s own estimate of local messages in a five-hour window for Plus and Standard Business, not a quota. The last column is this guide’s recommendation, not OpenAI’s.

Three details are easy to miss. All five models share a 1,050,000-token context window with up to 922,000 input tokens and 128,000 output tokens, so context size does not separate the tiers. Knowledge cutoffs do differ: April 30, 2026 for Astra and 6.1 Sol, April 20 for GPT-6 Sol, May 18 for GPT-6 Luna and February 16 for GPT-5.6 Luna. And the Standard column is only one of several price modes, covered next.

What GPT Astra costs, and what the rest cost

GPT Astra pricing is $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens at Standard processing, according to the Astra model page and the pricing page. Searches for “gptastra cost” or “GPT Astra cost” land on that same number, but the headline is not the whole bill. Four rules change what a real request costs.

Long prompts. A request with more than 272,000 input tokens is billed at 2x the input and cache rates and 1.5x the output rate for the whole request, not just the tokens above the line. For Astra that means $20 input, $2 cached and $75 output. The multiplier applies identically to Sol and Luna, so the ratio between tiers does not change; the absolute price of crossing the line does. Section three of the worked examples shows a 10% longer prompt costing 2.15x as much.

Caching. Cached input is priced at 10% of the uncached rate on Astra and Luna and at 5% on GPT-6.1 Sol. Writing to the cache is billed at 1.25x the input rate for 30-minute retention, which is $12.50 per million for Astra, $2.50 for Sol and $0.125 for Luna, and 2.5x for 12-hour retention. OpenAI’s practical guide to building with GPT-6 says to put stable instructions and reference material before changing task details and to count cache writes and long-context rates when estimating a full workflow. It also says cached input costs “up to 95% less” than uncached input, which is the Sol figure.

Batch, Flex and Fast. Batch and Flex are priced at 50% of Standard, so Astra drops to $5 input and $25 output. Fast mode is priced at 2x and, for Astra, OpenAI says it delivers up to 2x the speed of Standard. Regional processing adds a 10% premium where available. The mode multiplies the tier rather than replacing it, so Batch Sol at $1 input and $5 output lands on the same numbers as Standard GPT-6.1 Terra; compare tier and mode together before concluding one model is cheaper than another.

The Sol upgrade. GPT-6.1 Sol did not change the headline token price of GPT-6 Sol. It halved the cached rate from $0.20 to $0.10, which OpenAI describes as 95% below standard input pricing and 50% below GPT-6 Sol’s cached input pricing. For agent loops with a stable prefix, that is the lever that matters, and it is why the cached column in the master table is worth reading before the input column.

Worked cost examples at list price

The numbers below use only the official Standard prices above and assumed workload shapes, so they are arithmetic illustrations, not measurements. Reasoning tokens bill as output, so real output counts run higher than visible text on high effort settings.

Example 1, one agent session. The session sends 600,000 input tokens, of which 500,000 are served from cache and 100,000 are not, and produces 40,000 output tokens. On Astra that is $1.00 for uncached input, $0.50 for cached input and $2.00 for output, $3.50 in total. On 6.1 Sol it is $0.20, $0.05 and $0.40, $0.65 in total. On Luna it is $0.01, $0.005 and $0.02, $0.035 in total. Without caching, the same session costs $8.00, $1.60 and $0.08.

Example 2, one million classification calls. Each call has a 2,000-token prompt, of which 1,500 tokens are a cached shared prefix, and a 150-token answer. Across a million calls that is 500 million uncached input tokens, 1.5 billion cached and 150 million output. Astra costs $14,000, 6.1 Sol $2,650 and GPT-6 Luna $140. GPT-5.6 Luna, the older tier, costs $310. For short structured answers, Luna is the only tier where the bill stays a rounding error.

Example 3, the 272K line. A single request with 272,000 uncached input tokens and 5,000 output tokens costs $2.97 on Astra. The same request at 300,000 input tokens falls above the long-context line and costs $6.375, 2.15x as much for 10% more input. On Sol the pair is $0.594 and $1.275, and on Luna $0.0297 and $0.0638. Trimming a context to fit under the line is often worth more than switching a tier.

Example 4, a routed batch of 1,000 tasks. Each task has 200,000 input tokens, 160,000 of them cached, and 20,000 output tokens. All Astra costs $1,560. All 6.1 Sol costs $296. All Luna costs $15.60. Now route by difficulty: 70% of tasks to Luna, 25% to Sol, 5% to Astra. The total is $162.92, of which Astra’s 5% share accounts for $78, or 48%. Routing cut the bill to 10% of all-Astra, and the hardest 5% of tasks still make up nearly half of what is left.

Codex usage per model

Codex on a ChatGPT plan does not bill per token. It draws from an allowance, and OpenAI’s Astra usage article states that the allowance “does not correspond to a fixed number of messages or tasks.” It then publishes estimated ranges of local messages in a five-hour window for Plus and Standard Business: GPT-6 Astra 5 to 45, GPT-6.1 Sol 15 to 160, GPT-6 Sol 15 to 150 and GPT-6 Luna 350 to 3,000. The article calls these estimates and notes that cloud tasks may use more allowance than local messages and that weekly limits may also apply.

The ratios are informative. At the low end, Sol’s 15 is 3x Astra’s 5. At the high end, 160 against 45 is about 3.6x. That is a smaller gap than the 5x API price ratio, which suggests the subscription allowance does not track token prices exactly. Luna is a different story: 350 against Sol’s 15 is about 23x at the low end and 3,000 against 160 about 19x at the high end, close to the 20x gap between Sol’s and Luna’s API prices. That reading is this guide’s own arithmetic, not an OpenAI statement.

Plan structure matters more than the model choice for heavy users. Work and Codex usage is included with Plus and Pro plans and Business Standard and Premium seats. Plus has a five-hour window and a weekly window, and you need allowance in both. Pro 100, Pro 200 and Pro 500 currently have no five-hour limit in Work and Codex, though each keeps an included allowance. Ultrafast, the lower-latency tier for Astra and GPT-6.1 Sol, is included only on Pro $500 at launch, where it draws included usage at 8x the Standard rate for the selected model and then credits at 6x. The Ultrafast and Pro plans guide has the details.

Two OpenAI statements change how to spend the allowance. First, switching models does not restore allowance in a shared usage pool; Astra, Sol and Luna draw from the same Work and Codex pool, so dropping to Luna stretches what is left but does not top it up. Second, reasoning effort is independent of the model, and OpenAI says “Astra at Low effort can outperform Sol at High effort.” Its suggestion for anyone getting good results from Sol at High is to try Astra at Low or Medium. That is a vendor claim worth testing on your own repository, since a low-effort Astra turn may cost less allowance than a high-effort Sol turn on the same task.

One caution about the help-center page itself. Its message-range table lists the GPT-6 models, but parts of its descriptive text still name GPT-5.6 Sol and GPT-5.6 Luna, and it says GPT-6.1 Sol is selectable “if available for your account.” Treat the model picker in your own Settings as the final word on what you can use, and Settings, then Usage, as the final word on what you have left.

GPT Astra vs Sol: what the extra money buys

The comparison people actually search for, GPT Astra vs Sol or Sol vs Astra, comes down to one number: five. Astra’s standard input and output prices are five times 6.1 Sol’s, and its cached input is ten times. OpenAI’s pitch for Sol is “near-Astra intelligence for a fifth of the price,” and its own wording for Astra is the model for “the best results.” The question is how large the remaining gap is for your work.

OpenAI’s numbers point to a narrow gap on agentic work and a visible one on the hardest science tasks. On OSWorld 2.0 offline at maximum reasoning effort, the Sol launch post says 6.1 Sol comes within 2.1 percentage points of Astra at roughly one-seventh the cost per task. On Terminal-Bench Science 0.1 at maximum effort, it reports $5.47 per task for 6.1 Sol against $23.80 for Astra and $23.21 for Claude Opus 5.5. These are OpenAI-reported results and OpenAI’s own advice, in the same post, is to use Astra for the most difficult scientific research tasks.

The independent view is closer to the vendor view than skeptics might expect. In Artificial Analysis Intelligence Index v5, published October 27, Claude Opus 5.5 at max effort leads at 58, GPT-6 Astra at max scores 57 and GPT-6.1 Sol at max scores 55. Artificial Analysis reports cost per Intelligence Index task of $10.04 for Opus 5.5, $6.24 for Astra and $1.27 for 6.1 Sol. Two index points separate Astra and Sol, and the measured cost ratio is about 4.9x, in line with the 5x list-price ratio. Artificial Analysis also reports Astra leading its Terminal-Bench-Science (63%) and CritPt v2 hard subset (72%) results, with 6.1 Sol at 69% on CritPt at max effort.

Read together, the evidence supports a simple rule. For most agent loops Sol delivers the large majority of Astra’s result at a fifth of the bill. For the narrow set of tasks where Astra clearly leads, such as hard scientific reasoning and the highest-stakes computer use, the premium buys a measurable margin. Astra also changes behavior: OpenAI’s developer guidance says it can feel more tentative about when to stop and needs a clear definition of done, and that skills written to constrain Sol or Luna may over-constrain it.

GPT-5.6 Luna cost vs GPT-6 Luna cost

GPT-5.6 Luna costs $0.20 per million input tokens, $0.02 cached and $1.20 per million output tokens, according to the GPT-5.6 Luna model page and the pricing page. That is after the cut OpenAI announced on July 30, 2026, when it said GPT-5.6 Luna “will cost 80% less” and Terra 20% less, effective that day. Subscription prices and quota budgets stayed the same, but OpenAI said Terra and Luna usage now consumes fewer credits.

GPT-6 Luna costs $0.10 input, $0.01 cached and $0.50 output, per the GPT-6 Luna model page. Against GPT-5.6 Luna’s current list price that is 50% lower on input and cached input and 58% lower on output. OpenAI’s September 22 post says Sol and Luna API prices are 50% lower than GPT-5.6 “promotional pricing.” The post does not say which GPT-5.6 price counts as promotional; the pricing page applies the word to GPT-5.6 Sol, whose promotional price it says runs at least through November 21, 2026. This guide compares Luna against the GPT-5.6 Luna prices still listed on the pricing page on October 28. On those, the input drop matches OpenAI’s 50% exactly and the output drop is larger. The July post’s 80% cut also implies the earlier GPT-5.6 Luna price was $1.00 and $6.00, but that is arithmetic from the stated percentage, not a published figure.

In dollars, the one-million-call classification job from the worked examples costs $310 on GPT-5.6 Luna and $140 on GPT-6 Luna, 55% less. Both models are still on the pricing page, both have the same context window, and GPT-6 Luna has a newer May 18, 2026 knowledge cutoff against February 16 for the older model. OpenAI also says Luna improved over its predecessor, with a 5.4-point gain on AutomationBench at high effort and 58% lower cost per task, a vendor-reported result. There is no price reason to start new work on GPT-5.6 Luna; the reasons to keep it are pinned snapshots and revalidation cost.

GPT Astra benchmarks, summarized

The full tables, with reasoning-effort settings and caveats, are in the Astra guide and the GPT-6.1 Sol guide. This section gives only the numbers that bear on the pick between tiers, and every figure marked vendor-reported comes from OpenAI.

• Computer use (vendor-reported): OpenAI’s Astra launch post reports GPT-6 Astra at 72.6% on OSWorld 2.0 offline, against 65.7% for GPT-5.6 Sol, in its latency simulations. At maximum effort, 6.1 Sol is within 2.1 points of Astra at about one-seventh the cost per task.
• Software engineering (vendor-reported): On DeepSWE v1.1, OpenAI’s Sol and Luna post lists GPT-6 Sol at max effort at 68.8% and GPT-6 Luna at max effort at 66.6%. OpenAI frames Luna at max as comparable to Claude Opus 5 and Fable 5 at medium effort at far lower cost per task.
• Scientific workflows (vendor-reported): On Terminal-Bench Science 0.1, 6.1 Sol costs $5.47 per task at max effort against $23.80 for Astra, and OpenAI says Astra, with the highest score of the models it tested, should be used for the most difficult scientific research tasks.
• Independent (Artificial Analysis v5): Intelligence Index of 58 for Opus 5.5, 57 for Astra and 55 for 6.1 Sol, all at max effort, with per-task cost of $10.04, $6.24 and $1.27.

Two cautions apply to any benchmark comparison across these models. Several of OpenAI’s charts compare models at different reasoning-effort settings, so a bar for one model at xhigh against another at medium does not isolate the model. And the scores come from OpenAI’s research environment or API, which OpenAI says may differ from production ChatGPT. The independent Artificial Analysis numbers cover Astra and 6.1 Sol and, in the sections reviewed for this guide, did not list Luna among its top-ten comparisons.

Which model to use: a decision guide

The guide below is opinionated and based on the published prices, OpenAI’s own positioning in its GPT-6 guide, and the cost arithmetic above. It is not a substitute for running your own evaluation set, which is the last step.

Use GPT-6 Luna when

The task is narrow, repeated and has a clear definition of correct: extracting invoice fields, classifying requests, producing structured summaries, making short edits, renaming symbols across files. OpenAI’s guide names the first three, extracting invoice fields, classifying requests and producing structured summaries, as Luna’s kind of work, and its reasoning-level advice lists small edits as routine. At $0.10 and $0.50 it also suits the cheap steps inside a larger agent loop, such as summarizing a tool result or checking that a test passed. In Codex on Plus it stretches the five-hour window the furthest by a wide margin.

Use GPT-6.1 Sol when

The work is multi-step and you want it to run often: implementing a feature, a refactor across a codebase, a deep investigation, a long-running agent with stable tool definitions. This is the default for most coding and agent work. The 5% cached rate makes it unusually cheap for loops that reuse a large prefix, and OpenAI says it supports multi-agent workflows in the Responses API, which is currently in beta. Choose it over GPT-6 Sol unless you are pinned, since the token prices match and the cached rate is half.

Use GPT-6 Astra when

A mistake costs more than the premium: a subtle production bug, a security review, an unfamiliar scientific or mathematical problem, high-stakes computer use, or the planning pass that every cheaper model will follow. Consider Astra at Low or Medium effort before Sol at High, per OpenAI’s help center. Expect to define done explicitly, and budget for the five-times-higher price on output tokens, which dominate agent bills.

Route instead of choosing

The strongest cost result in the worked examples is routing. A plan-with-Astra, build-with-Sol, check-with-Luna split gave a $162.92 bill for 1,000 tasks against $1,560 for all-Astra and $296 for all-Sol. OpenAI’s July post on GPT-5.6 described the same pattern for Sol and Luna: use the larger model to resolve uncertainty and define the plan, then use the smaller one to implement well-specified changes and run tests. The routing thresholds are yours to measure; start with a rule such as “Astra only on tasks that failed twice on Sol.” For a wider comparison with non-OpenAI options, see cheapest AI coding models in 2026, and for the earlier generation, GPT-5.6 Sol, Terra and Luna explained.

A short evaluation before you commit

Take 30 to 50 real tasks from your own backlog. Run each on Luna, Sol and Astra at the same effort, then at one lower effort for Astra. Record pass rate, tokens and cost per passing task, not cost per call. If Luna passes 80% of a task class, route that class to Luna and send failures up. If Sol and Astra pass the same tasks, the extra $8 per million input tokens is not buying anything for that class. This is the only experiment that settles the question for your repository, and none of the figures in this guide were produced that way.

Limitations

Prices and availability change, and every figure here is as of October 28, 2026. The Codex message ranges are OpenAI estimates for local messages on Plus and Standard Business; cloud tasks, higher effort, Fast mode and weekly limits all change them. The worked examples assume prompt shapes and cache hit rates that your traffic may not match, and cache writes, which bill at 1.25x the input rate, are left out of examples 1 and 2 for clarity.

The release dates for September 22 and 29 and the OpenAI-side details come from OpenAI’s own posts and X accounts. The September 3 date is confirmed by OpenAI and by the dated Reuters and TechCrunch URLs, but this guide did not read the full text of those two articles. Benchmark figures are vendor-reported unless attributed to Artificial Analysis. No independent test of Luna appeared in the sources reviewed, and nothing in this guide was tested by the author.

FAQ

Verdict

Start every new agent on GPT-6.1 Sol, move the repetitive and structured steps down to GPT-6 Luna, and promote a task to GPT-6 Astra only after it has failed on Sol or when the cost of a miss is high. That ordering follows the prices, the Codex allowance ratios and OpenAI’s own positioning, and the independent index agrees that Sol and Astra are two points apart. Treat the five-times price gap as a budget line to spend where it changes outcomes, not a default. If you are on GPT-5.6 Luna, the move to GPT-6 Luna halves input cost and is the easiest saving in the lineup.

Sources

Sources checked October 28, 2026.

• OpenAI API pricing (Standard, Batch, Flex and Fast rates for every model above)
• Introducing GPT-6 Astra, Introducing GPT-6 Sol and Luna and Introducing GPT-6.1 Sol (availability, dates, prices, vendor-reported benchmarks)
• GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and GPT-5.6 Luna model pages (context, cutoff, cache and long-context rules)
• Managing usage with GPT-6 Astra in Work and Codex (Codex message estimates, plan limits, launch resets)
• Practical guide to building with GPT-6 and Advancing the price-performance frontier with GPT-5.6 (model selection, caching, July 30 Luna cut)
• Artificial Analysis Intelligence Index v5 (independent index and cost per task)
• Reuters, TechCrunch and The Verge (launch dates)
• X posts embedded above: @OpenAI, September 3, @sama, September 4, @OpenAIDevs, September 22, @OpenAI, September 29
• OpenAI on YouTube: Introducing GPT-6 Astra for developers, Introducing GPT-6 Astra and OpenAI DevDay 2026 keynote

Related reading: GPT-6 Astra complete guide, GPT-6.1 Sol guide, GPT-5.6 Sol, Terra and Luna explained, cheapest AI coding models in 2026, ChatGPT Ultrafast and Pro plans.