GPT-6 Sol and Luna are out at half the price - what it changes for business automation

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22. They are the mid and low tiers under GPT-6 Astra, and they cost half as much as the GPT-5.6 models they replace. According to VentureBeat, Sol is now $2 per million input tokens and $10 per million output, down from $4 and $20. Luna is $0.10 input and $0.50 output, down from $0.20 and $1.20. OpenAI says these are permanent prices, not a launch promotion.

For anyone running automation at volume, that is the headline. The benchmark story is more mixed.

What is actually new

Three things changed.

**Price.** Cached input gets a 90% discount on both models, which works out to $0.20 per million on Sol and $0.01 on Luna. Most business agents resend the same system prompt, tool definitions and reference documents on every call. Cached reads are a large share of what those agents cost, so this discount matters more than the list price.

**Context.** Digital Applied lists a 1.05M-token context window and 128K maximum output. There is a catch in the fine print: requests over 272K tokens are billed at 2x input and 1.5x output. The window is big, but filling it is not cheap.

**Reliability on agent tasks.** VentureBeat reports Sol scoring 33.2% on AutomationBench at $0.27 per task, and a deception rate of 1.3% against 10.4% for GPT-5.6. AutomationBench is the closest public test to the work we build: multi-step business tasks across real tools. Astra still leads it at 41.4%.

What the numbers do not say

Read the benchmarks with the prices next to them.

A 33% AutomationBench score means Sol fails about two out of three of those tasks unaided. That is better than last month's options, not good enough to leave alone. The best model on the chart is still failing most of the time.

Sol is not better than its predecessor everywhere. Digital Applied's comparison has GPT-5.6 Sol scoring 72.7% on DeepSWE against GPT-6 Sol's 68.8%. The new model is cheaper per task ($2.74 against $6.46) but it is not stronger at coding. OpenAI has made the model cheaper to run and more honest. It has not made it smarter across the board.

Maximum reasoning effort is not always the best setting either. On AutomationBench, Digital Applied found Sol's max setting scored lower than the one below it while costing 24% more. If your vendor defaults everything to max, you are paying for nothing.

What it means for a business owner

If you already run agents on GPT-5.6, your bill for the same work roughly halves, with no rebuild. That alone is worth a model-string change and a regression test.

Luna is the more interesting model for most small businesses. At $0.10 per million input tokens, the unglamorous jobs get close to free: sorting inbound email, pulling fields out of invoices, tagging support tickets, summarising call notes into a CRM. These were already cheap. Now the model cost stops being part of the conversation, and the real cost is the build and the checking.

That changes which projects make sense. A workflow that handled 200 documents a day and only just paid for itself on older pricing now has room to spare. Tasks you ruled out on running cost are worth pricing again.

What it does not change is the design. The failure rate on multi-step tasks is still high enough that every agent touching money, customers or records needs a review step and a log. Cheaper tokens make it tempting to hand the model more autonomy. The benchmarks say not yet.

The honest caveat

Three reasons to go slower than the price cut suggests.

Safety testing is uneven. Digital Applied reports Sol showing "warning circumvention" in 64.4% of stress tests and unauthorised agent interaction in 11.3%, worse than both Luna and Astra. If Sol will run with write access to your systems, keep permissions narrow.

OpenAI has not published rate limits for the new models. If you run large batches, test throughput before you commit a production workload.

The release pace is the other problem. This is the second pricing reset from OpenAI in two months. A price you plan around today may not hold in November. Anything we build should be able to swap models without a rewrite, and yours should too.

What to do this week

If you run anything on GPT-5.6 Sol or Luna, take 50 real inputs from last month, run them through the GPT-6 equivalent, and compare the outputs side by side. Check accuracy before you check the bill. If quality holds, switch and keep the savings. If it slips on your specific task, you have lost an afternoon, not a quarter.

If you have a task you shelved because running costs did not add up, get the numbers out again. Price it at Luna rates with a human checking a sample. The maths may look different now.