OpenAI's low-cost model is GPT-6 Luna, launched on September 22, 2026, at $0.10 per million input tokens and $0.50 per million output tokens, priced in US dollars and metered per token[1]. The company describes these as long-term standard prices, not a limited promotion[2]. The step-up model for heavier reasoning work, GPT-6 Sol, costs $2 per million input tokens and $10 per million output tokens, each rate half of its GPT-5.6 predecessor[1].
If your workload is high-volume information extraction, document summarization or classification, price it against Luna and walk through the bill components below. If it is complex coding or multi-step agent flows, start from Sol's rates instead. Either way, treat the sticker price as only part of the invoice: output mix and cache-hit share move the final number, and batch discounts can move it further where they are offered.
How the price got here
Luna's rate did not appear overnight. The GPT-5.6 family (Sol, Terra and Luna) went on sale in early July 2026. On July 30, OpenAI cut Luna's API price by 80 percent to $0.20 per million input and $1.20 per million output, trimmed Terra by 20 percent to $2 and $12, and left Sol untouched, about three weeks after the family's release[3]. The company attributed the cut to engineering work: rewritten GPU kernels lowered end-to-end serving cost by 20 percent, and improved speculative decoding lifted token-generation efficiency by more than 15 percent[3]. That round of price cuts landed in a market where enterprise buyers hesitate to deploy expensive models without visible returns, and where Google and Microsoft were pushing their own value-tier models[4].
The September 22 launch reset the floor again. GPT-6 Luna now prices at $0.10 and $0.50 per million tokens against its predecessor's $0.20 and $1.20; the output rate alone fell about 58 percent[1]. Both new models run inside ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscriptions, and Luna also reaches Free and Go users on desktop[2]. An OpenAI spokesperson has confirmed the rates are permanent rather than promotional, and the company restated the position to VentureBeat[5].
The same-day competition
The price move is not OpenAI acting alone. Anthropic released Claude Opus 5.5 the same day at $4 per million input tokens and $20 per million output, twice Sol's rate on both sides, while cutting cache-read pricing from $0.50 to $0.20 per million tokens. Anthropic says typical workloads cost about 40 percent less than on Opus 5 under default settings, with output generation over 30 percent faster[1]. Meta shipped a Llama 4.2 update aimed at cheaper multimodal work and Google opened Gemini 3.6 Flash to all developers with price cuts the same day; open-source models from DeepSeek, Qwen and Kimi have been updating in recent months at prices that run at fractions of closed alternatives[2]. Late September repriced the whole mid-tier, and the numbers above are the September 22 snapshot, not a one-off flash sale.
What the sticker price does not cover
Three rate components sit between the published rate and the invoice, and billing itself splits in two.
Output mix. At Luna's rates, output tokens cost five times what input tokens do. The ratio matters because chatty workloads burn the expensive side of the meter; a developer's August 2026 price check measured GPT-5.6 Luna's output rate at six times its input rate, and the asymmetry carried over into the new generation at a lower multiple[6].
Cache-hit share. For input that hits OpenAI's cache, the read rate runs 90 percent below the normal input rate[1]. Only the cached portion of each prompt receives the discounted rate, so the saving scales with how much of your context repeats: fixed system prompts, re-read document sets, long conversation histories[6]. OpenAI's own framing is that for agents repeatedly reading codebases, documents and chat logs, cache-hit rate directly moves the actual bill[1].
Processing channels. Batch and Flex processing ran at roughly half of standard rates for GPT-5.6 Luna as of a developer's August 10, 2026 price verification: $2.20 against $4.40 standard on the test workload below[6]. Budget on these channels only after confirming their current terms for the GPT-6 generation.
On the billing side, chat subscriptions and API accounts are separate ledgers, and paid chat apps generally do not include API quota[6]. Within subscriptions, the July price cut worked by stretching quota: Terra and Luna consumed less of the same budget, a change OpenAI said was reflected in Codex and ChatGPT Work billing rules[3].
A worked monthly bill (hypothetical)
Take a workload of 20,000 API calls per month, each with 500 input tokens and 100 output tokens, which is 10 million input and 2 million output tokens. The formula is monthly cost = 10 × input rate + 2 × output rate[6].
At GPT-6 Luna's sticker rates: 10 million × $0.10 = $1.00, plus 2 million × $0.50 = $1.00, for $2.00 per month. The same workload on GPT-5.6 Luna's pre-September standard rates cost $4.40[6], a monthly bill cut by more than half.
Now a cache variant, and this one is hypothetical: if 80 percent of those input tokens hit the cache, the input side prices at 8 million tokens at a derived $0.01 per million (the 90 percent discount applied to $0.10) plus 2 million at $0.10, or $0.28. The total lands near $1.28 per month. These are arithmetic examples on published rates, not a quote; your output ratio and cache-hit share will differ.
What the low price costs you
The price cut came with capability trade-offs, and the evidence splits by who ran the test.
OpenAI's own numbers put GPT-6 Sol at 33.2 percent on the AutomationBench cross-application business-process test at $0.27 per task, against Claude Opus 5's 26.9 percent at about 11 times the cost, and 68.8 percent on the DeepSWE v1.1 software-engineering test, close to Claude Fable 5's 69.9 percent at roughly 80 percent lower per-task cost[1]. Both are vendor-administered benchmarks.
Independent measurements are less flattering. Artificial Analysis scores Opus 5.5 at 58 on its top-reasoning-strength index, leading the field, with GPT-6 Sol at 48 and Luna at 37; Sol and Luna's composite indexes sit near their predecessors[1]. DigitalApplied's analysis, as cited by the review platform Yingzheng, found Luna scoring 66.6 percent on DeepSWE against Sol's 68.8 percent, at about one-fifth of Sol's per-task cost[5]. The same analysis also recorded GPT-5.6 Sol scoring 72.7 percent at the highest effort setting, above GPT-6 Sol's 68.8 percent[5].
TechCrunch reported enterprise deception rates of 1.3 percent for Sol and 2.8 percent for Luna, lower than the previous generation, as cited by Yingzheng[5]. DigitalApplied's testing, reported separately in the same write-up, recorded GPT-6 Sol passing a warning-bypass test 64.4 percent of the time[5].
Before you commit
Three actions follow from the numbers above. First, recompute your own bill: apply the formula in the worked example with your real output ratio and your observed cache-hit share; the example moved only the cache share, and it took the hypothetical total from $2.00 to $1.28. Second, if you plan to run batch workloads, confirm the Batch and Flex terms for the GPT-6 generation on the current price page before locking a budget. Third, if peak capability decides the purchase rather than price, GPT-6 Astra (the flagship OpenAI released on September 3) sits above Sol and Luna with a markedly higher cost, so price it as a separate line item rather than assuming the mid-tier rates extend upward[1].