TL;DR: OpenAI cut GPT-5.6 Luna from $1.00/$6.00 to $0.20/$1.20 per million tokens on July 30, 2026. At a heavy 1M-tokens-a-day coding workload, that moves the breakeven on a $2,268 used RTX 4090 from about 3.4 years to roughly 25 years. Buying a GPU to save money on API bills is now dead math — the reasons left to buy are privacy, control, and everything a GPU does that an API can’t.

	GPT-5.6 Luna API	Used RTX 4090 (local)	RunPod RTX 4090 rental
Upfront cost	$0	~$2,268 (eBay avg, Aug 2026)	$0
Cost at 1M tokens/day	~$0.29–0.40/day	~$0.15/day electricity	$0.34/hr while running
Output speed	~154–190 tok/s	~70 tok/s (Qwen3.6-27B Q4)	~70 tok/s (same card)
Breakeven vs Luna	—	~25 years at 1M tok/day	Never, for pure LLM inference
The catch	Your code leaves the house	Capital sunk in one card	Setup time + storage fees

Honest take: If your only reason to buy a $2,268 GPU was replacing an OpenAI subscription, don’t — Luna at $0.20/M ended that argument. If you already own a 24GB card, keep running it: your marginal electricity cost (~$0.25–0.34 per million output tokens) still undercuts Luna’s $1.20/M output rate. And if privacy is why you run local, nothing about a price cut changes that.

What actually changed on July 30

OpenAI shipped the GPT-5.6 family — Sol, Terra, and Luna — on July 9, 2026. Three weeks later, on July 30, it cut Luna’s API price by 80% and Terra’s by 20%, citing efficiency gains across its serving stack. CNBC framed it as cost-sensitivity among enterprise customers; VentureBeat called it a price war, with Chinese labs like DeepSeek setting the floor.

The new rate card, per OpenAI’s announcement and the model docs:

Per 1M tokens	Before Jul 30	After Jul 30
Luna input	$1.00	$0.20
Luna output	$6.00	$1.20
Luna cached input	$0.10	$0.02 (90% discount on input rate)
Terra input / output	$2.50 / $15.00	$2.00 / $12.00
Sol input / output	$5.00 / —	unchanged

A few spec-sheet facts that matter for the math below: Luna has a 1.1M-token context window, 128K max output, vision input, tool calling, and structured outputs. Cache writes bill at 1.25× the input rate, cached reads at 10% of it — designed for exactly the write-once-read-many pattern coding agents produce.

You’ll see some coverage quoting Luna’s old output price or assuming $0.80/M. The current published rate is $1.20/M output. We verified it against OpenAI’s model documentation and three independent pricing trackers on August 7, 2026; every number in this article uses $0.20/$1.20.

The billing trap to know before you celebrate

Here’s a problem you will actually hit if you point a long-context agent at Luna: any single request whose input exceeds 272,000 tokens is billed at 2× the input rate and 1.5× the output rate — the whole request, not just the overage. An agent that naively stuffs a 300K-token repository dump into one call pays $0.40/M on input and $1.80/M on output for that request, silently. The fix is unglamorous: keep individual requests under the 272K line by chunking or summarizing context, and lean on prompt caching so repeated context bills at $0.02/M instead of $0.20/M. Every serious coding harness (Cursor, Cline, Claude Code-style agents) already manages context this way; if you’re writing your own loop, you have to.

What a local GPU actually costs per million tokens

The honest comparison needs the local side priced the same way — per million tokens, with sources.

The hardware. A used RTX 4090 averaged $2,268 on eBay in early August 2026 per Best Value GPU’s tracker, with ResalePrices putting the fair asking range near $2,500 on active listings. (Launch MSRP was $1,599 in October 2022; the used market has been above MSRP all year — see why NVIDIA skipping 2026 consumer GPUs keeps used prices high.)

The model. The strongest coding model that comfortably fits a 24GB card is Qwen3.6-27B at Q4_K_M (16.8GB), which we benchmarked at ~70 tok/s on an RTX 4090 in May 2026.

The electricity. At 70 tok/s, one million output tokens takes 14,286 seconds — just under 4 hours of generation. Charging the full 450W board power against those hours (worst case), that’s 1.79 kWh, or $0.34 per million output tokens at the EIA’s 18.83¢/kWh US residential average (April 2026). Power-limit the card to 315W — which keeps ~94% of performance per Tom’s Hardware’s testing — and it drops to about $0.25 per million output tokens. All of that is our arithmetic, flagged as such; your kWh rate moves it (California residents: multiply by ~1.5).

So a card you already own generates tokens for $0.25–0.34/M in electricity against Luna’s $1.20/M output rate. Owners should keep running local. The purchase decision is where the math collapses.

The breakeven math, shown in full

Take the workload the upgrade argument always assumes: a developer pushing 1M tokens a day through a coding agent, at a realistic 4:1 input-to-output ratio (800K in, 200K out).

# roi.py — breakeven on a used RTX 4090 vs GPT-5.6 Luna, Aug 2026
luna_day   = 0.8 * 0.20 + 0.2 * 1.20      # $/day, uncached
local_day  = 0.067 + 0.087                 # generation kWh + 24/7 idle @ 20W
gpu_price  = 2268
days = gpu_price / (luna_day - local_day)
print(f"Luna: ${luna_day:.2f}/day  Local power: ${local_day:.2f}/day")
print(f"Breakeven: {days:,.0f} days ({days/365:.1f} years)")

$ python3 roi.py
Luna: $0.40/day  Local power: $0.15/day
Breakeven: 9,257 days (25.4 years)

Run the same script with the July 29 prices ($1.00/$6.00) and Luna costs $2.00/day — breakeven lands at 1,229 days, about 3.4 years. That was a defensible purchase for a heavy user who’d keep the card five years. The price cut multiplied the payback period by seven, past any plausible hardware lifetime.

It gets worse with caching. An agent whose system prompt and repo context hit the cache 75% of the time pays roughly $0.29/day at the new rates. And it’s not just OpenAI: DeepSeek V4-Flash prices at $0.09/$0.18 per million, which works out to about $0.11/day for this workload — less than the local card’s electricity. When the API undercuts your own power bill, “buy a GPU to save on tokens” isn’t a marginal case anymore; it’s over.

Scenario (1M tok/day, 4:1)	Daily cost	Breakeven on $2,268 RTX 4090
Luna, old pricing (pre-Jul 30)	$2.00	~3.4 years
Luna, new pricing, uncached	$0.40	~25 years
Luna, new pricing, 75% cached input	$0.29	~44 years
DeepSeek V4-Flash	$0.11	Never (cheaper than local power)
Local RTX 4090 (power only)	$0.15	—

To break even inside three years at the new uncached Luna rates, you’d need to push roughly 6 million mixed tokens a day, every day — about 1.2M output tokens, or five solid hours of flat-out generation at 70 tok/s, sustained for three years. A single 4090 tops out near 6M output tokens per day running 24/7 single-stream, so the volumes that rescue the ROI live at the edge of what one card physically does. (Batch serving with vLLM raises the aggregate ceiling, but if you’re batch-serving for three years you’re running a business, not a home lab — different article, different math.)

Renting doesn’t rescue it either: at RunPod’s $0.34/hr Community rate for an RTX 4090, the ~4 hours to generate 1M output tokens costs about $1.35 — slightly more than Luna’s $1.20/M output price, before you count setup time. Renting a 4090 on RunPod still makes sense for image generation, fine-tuning, and trying hardware before buying — just not as a cheaper LLM endpoint than Luna.

The quality check flips the script too

The last refuge of the ROI argument would be quality: “sure, Luna is cheap, but a lightweight cloud model can’t match a tuned local 27B.” In July 2026, that’s no longer true.

Benchmark	GPT-5.6 Luna	Qwen3.6-27B (best 24GB local)	GPT-5.6 Terra
SWE-bench Pro	62.7%	53.5%	63.4%
Terminal-Bench	84.7% (v2.1)	59.3% (v2.0)	87.4% (v2.1)
Price per 1M in/out	$0.20 / $1.20	electricity	$2.00 / $12.00

(Terminal-Bench versions differ between the two eval runs — 2.1 vs 2.0 — so treat that row as directional rather than a controlled head-to-head. The SWE-bench Pro gap is the cleaner signal.)

Luna scores within two points of Terra on SWE-bench Pro at a tenth of the price, and clearly ahead of the strongest coding model that fits a 24GB consumer card. Artificial Analysis clocks Luna’s API at 154–190 tokens per second depending on reasoning effort — two to nearly three times what the 4090 manages on the 27B. The budget cloud tier is now faster and stronger than the best single-card local stack. DeepSeek V4-Flash complicates the picture slightly — it beats Luna on two agentic benchmarks (AutomationBench, Toolathlon) while losing on three others — but that’s a fight between two APIs, not a case for buying silicon.

What’s left of the local case (more than you’d think)

None of the above touches the reasons most home labs actually exist:

Privacy is not a price point. Your code, your health questions, your unreleased product docs — none of it transits anyone’s infrastructure when the model runs on your card. Our privacy audit of what actually stays local goes deep on this, and 2026’s incident list (Hugging Face breach included) keeps making the argument for you.

No caps, no deprecations, no repricing risk. The July 30 cut went the right direction, but the same lever moves both ways, and models get deprecated on OpenAI’s schedule, not yours. A GGUF on your NVMe drive runs identically in 2030.

The GPU does things an API doesn’t. ComfyUI image and video generation, LoRA fine-tuning on your own data, embedding pipelines, game streaming on the weekend. The API replaces exactly one of the card’s jobs — the rent-or-buy breakdown covers how to split those workloads sensibly.

Offline is a feature. Nothing about $0.20/M helps on a plane, behind an air gap, or during an outage.

What the price cut did kill is buying hardware with a spreadsheet that says it pays for itself. It won’t — not at $0.20/M, not against $0.02/M cached reads, and not while DeepSeek is selling tokens below your electricity rate. If you’re using Luna as a BYOK backend in a coding tool instead, our sister site aicoderscope.com covers that setup side.

Verdict by buyer

• You were about to spend $2,300 on a used 4090 purely to cut API spend: don’t. Put $50 into a Luna API key, add caching, and revisit in a year. The $400/month GPU-bill problem is now mostly a configuration problem, not a hardware one.
• You already own a 24GB card: change nothing. Your marginal cost per output token is a fifth of Luna’s, the card is sunk cost, and your data stays home.
• You want local AI for privacy/control/image gen: the math in this article is irrelevant to you, and that’s the point — buy the card for what an API can’t do. A used RTX 3090 at roughly half the 4090’s price remains the value entry to the same 24GB tier.
• You need frontier-quality output at volume: the price war is your friend. Luna, Terra, and DeepSeek V4-Flash are racing each other down; local was never going to win that race on cost, and now it doesn’t have to — run local for the private 80%, burst to cloud for the rest.

FAQ

Did OpenAI really cut Luna 80%, or is this a promo rate? It’s the published API rate as of July 30, 2026 — $0.20/M input, $1.20/M output — confirmed by OpenAI’s announcement and CNBC. No end date was announced. Rates can change in either direction; that risk cuts both ways in this article’s math.

Is GPT-5.6 Luna actually good enough for coding? On public benchmarks, yes: 62.7% SWE-bench Pro and 84.7% Terminal-Bench 2.1 put it ahead of anything that fits a 24GB consumer card, and within ~2 points of Terra at 10% of the price. Test it on your own repo before committing either way.

What’s the cheapest way to run 1M tokens a day now? DeepSeek V4-Flash at $0.09/$0.18 per million (~$0.11/day at a 4:1 mix) is currently cheaper than the electricity to generate the same tokens on a 4090 you already own. Luna with aggressive caching lands around $0.29/day. Local wins on privacy, not price.

Does the math change with a cheaper card? Directionally no. A used RTX 3090 (~$1,050–1,250 in recent months) halves the capital but also runs the 27B at ~40 tok/s, raising electricity per token. Breakeven vs uncached Luna still lands past a decade at 1M tokens/day.

When does buying a GPU still pay for itself? Sustained volumes around 6M+ mixed tokens per day against uncached Luna rates, workloads APIs can’t serve (fine-tuning, image/video gen, air-gapped environments), or any workload where the data can’t leave the building. Otherwise treat the GPU as a privacy and capability purchase, not a savings plan.

Recommended Gear

• RTX 4090 — the fastest 24GB consumer card; ~70 tok/s on Qwen3.6-27B Q4_K_M
• RTX 3090 — same 24GB tier at roughly half the used price, ~40 tok/s on the same model

Sources

• OpenAI cuts prices for two of its GPT-5.6 AI models — CNBC
• Advancing the price-performance frontier with GPT-5.6 — OpenAI
• AI price wars: OpenAI cuts GPT-5.6 Luna prices by 80% — VentureBeat
• OpenAI reaches one billion active users and cuts GPT-5.6 Luna and Terra prices — gHacks
• OpenAI releases GPT-5.6: a three-tier model family — MarkTechPost
• GPT-5.6 benchmarks: Sol, Terra & Luna scores — AnyCap
• GPT-5.6 Luna — Intelligence, Performance & Price Analysis — Artificial Analysis
• DeepSeek-V4-Flash vs GPT-5.6 Luna comparison — LLM-Stats
• GPT-5.6 Luna model page (context window, caching, long-input billing) — Requesty
• GPT-5.6 pricing: Sol, Terra and Luna rates explained (prompt caching) — eesel AI
• RTX 4090 price tracker US, August 2026 — Best Value GPU
• RTX 4090 used GPU fair asking range — ResalePrices
• Electric Power Monthly, residential rates — U.S. EIA

Last updated August 7, 2026. Prices and API rates change; verify current rates before purchasing.