GPT-6 Sol and Luna: 50% Cheaper, Not Always Better

OpenAI cut the price of using its models in half on Tuesday, and it did so about 90 minutes after Anthropic quietly shipped an updated Opus 5.5. The two moves landed on the same day as GPT-6 Sol and GPT-6 Luna, cheaper siblings to the GPT-6 Astra flagship OpenAI launched three weeks earlier. Sol now costs $2 per million input tokens and $10 per million output tokens, half of GPT-5.6 Sol's promotional rate. Luna drops to $0.10 and $0.50. The headline is a price cut. The more interesting number is what OpenAI's own benchmark charts show once you read past the parts it chose to highlight.

Table of contents

• The new prices
• What OpenAI says the models can do
• What the fine print shows
• The baseline is temporary
• Where this sits against Claude and Gemini
• Why cost-per-task is becoming the real scoreboard
• Frequently asked questions

The new prices

Both models arrived on September 22, according to sqmagazine and Benzinga. GPT-6 Sol runs $2 per million input tokens and $10 per million output tokens, down from GPT-5.6 Sol's $4 and $20. GPT-6 Luna runs $0.10 and $0.50, down from roughly $0.20 and $1.20 for GPT-5.6 Luna. OpenAI credits the cut to better caching and more efficient inference, and Benzinga reports paid ChatGPT Work and Codex accounts get both models immediately, while free and Go users get Luna in the desktop app.

Both models share a 1.05 million token context window, a 128,000 token maximum output, and OpenAI's standard agent tools: function calling, web search, file search and computer use. Their knowledge cutoffs sit a few weeks apart, April 20 for Sol and May 18 for Luna, according to Kingy AI's specs breakdown.

What OpenAI says the models can do

OpenAI's pitch is efficiency, not a new capability ceiling. The company says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol on its internal factuality test, which it built from de-identified real conversations where users flagged errors. On its own AutomationBench chart, Sol at its highest reasoning setting scores 33.2% at 27 cents per task, a result OpenAI compares favorably against GPT-6 Astra at low effort, which scores 30.3% at nearly four times the cost per task.

Kingy AI's review of the same chart set adds a second data point: on Agents' Last Exam, Sol at maximum effort scores 56.4%, edging past Claude Opus 5's reported best of 55.9%. Framed that way, the story is a genuine efficiency win. Framed a different way, using the numbers OpenAI put on the same pages but did not lead with, the picture gets more complicated.

What the fine print shows

Two independent write-ups, from Kingy AI and DigitalApplied, read past OpenAI's headline comparisons and into the full chart data. Both found the same pattern: on DeepSWE and OSWorld 2.0, two coding and computer-use benchmarks, GPT-6 Sol's best recorded score trails GPT-5.6 Sol's best score, and both trail Claude Opus 5. DigitalApplied's phrase for it is direct: on two of three coding and computer-use charts, the predecessor model still wins.

None of this means GPT-6 Sol is a weaker model in any simple sense. It means OpenAI optimized this release around cost per task rather than peak score, and its own comparison charts show that trade-off plainly once you look at every bar rather than the one the launch post points to. That is worth knowing before switching a production workload over on price alone.

The baseline is temporary

The 50% figure is measured against a price that will not last. GPT-5.6 Sol's $4 and $20 rate is promotional, and DigitalApplied notes OpenAI guarantees it only "at least through November 21, 2026." What GPT-5.6 Sol costs after that date, and whether GPT-6 Sol's discount looks as large against a reset baseline, is not stated anywhere in OpenAI's pricing page.

This is not the first time OpenAI has moved prices on a promotional model rather than a fresh one. It cut GPT-5.6 Luna's price by 80% and GPT-5.6 Terra's by 20% back in July, and it trimmed GPT-5.6 Sol's API rate by more than 20% in August, before this week's deeper cut arrived alongside the GPT-6 label. The pattern suggests OpenAI treats pricing as a lever it pulls between model launches, not only at them.

Where this sits against Claude and Gemini

DigitalApplied's price comparison lines Sol up against Anthropic's published rates: Sol's $2 input and $10 output match Claude Sonnet 5 on every line both companies publish, including a $2.50 short cache write and a $0.20 cache read on Anthropic's side. Luna at $0.10 and $0.50 undercuts Claude Haiku 4.5's $1 and $5 by a factor of ten.

The timing with Anthropic's own update adds a wrinkle. sqmagazine reports OpenAI's announcement landed roughly 90 minutes after Anthropic shipped a new version of Opus 5.5, turning the day into what the outlet called a head-to-head launch window. Neither company has framed it as a coordinated response to the other, and nothing in the public record shows one release caused the other. It may simply be that both companies ship on similar internal cycles in the same crowded month. What is clear is that a buyer comparing prices this week is comparing two sets of numbers announced hours apart, not months apart, which is new even by this industry's pace.

Why cost-per-task is becoming the real scoreboard

Benzinga's framing is the useful one here: the announcement reflects a shift in how AI vendors compete, away from a single top benchmark number and toward cost per completed task. That shift makes sense once agents are doing the buying. A person asking one question barely notices per-token pricing. An agent running thousands of steps across a long task multiplies every token rate into a real bill, and a model that is slightly weaker but dramatically cheaper can win on total cost even when it loses on a leaderboard.

The risk in that shift is that headline price cuts are easy to publish and full benchmark tables are easy to skip past. GPT-6 Sol and Luna are a fair test case: a genuine price improvement, sold with charts that reward a quick read and reward a slow one differently. Anyone deciding whether to switch is better served by the slow read.

Frequently asked questions