OpenAI cuts the input price of its Luna model line by 90% in 54 days

Working backwards from the July cut gives the starting price. An 80% reduction that left GPT-5.6 Luna at $0.20 per million input tokens and $1.20 per million output tokens means it cost $1.00 and $6.00 the day before ^[1][12]. GPT-6 Luna arrived on September 22 at $0.10 and $0.50 ^[4]. So over 54 days the input price of OpenAI's Luna line fell 90% and the output price fell about 92% ^[13][14]. Terra, the higher-end model, went from $2.50 and $15 to $2 and $12 ^[2][15].

The mid-tier has settled on a single price. Claude Sonnet 5 and GPT-6 Sol both charge $2 in and $10 out ^[17]. Anthropic lists its more expensive Claude Opus 5.5 at $4 and $20, and CryptoBriefing says efficiency gains cut the effective bill by around 40% ^[5].

CryptoBriefing argues the cuts are steep enough to make developers ask whether self-hosting open models is still worth the trouble ^[11]. The report includes no GPU, power or staffing costs, so it cannot settle a breakeven. It does fix the other side of the comparison. A self-hosted model justified against July's $1.00 and $6.00 Luna is now up against $0.10 and $0.50 ^[12][4]. That competition sits in the cheap tier. According to the report, enterprises now send summarisation, classification and basic customer interactions to cheap proprietary models and keep premium models for specialised work ^[6].

The two American labs are not the only source of cheap tokens. DeepSeek V4 Flash costs roughly $0.14 and $0.28, a price CryptoBriefing says undercuts American counterparts by a wide margin ^[8]. Against Sol that holds: Sol charges about 14 times as much for input and 36 times as much for output ^[19]. Against GPT-6 Luna it fails on input, where DeepSeek charges 40% more, and only DeepSeek's output is cheaper, by 44% ^[16]. The report adds that some Chinese models approach American mid-tier capability at five to nine times lower cost ^[9].

Usage is the other test of the thesis. Open-source models still command significant volume share, particularly on OpenRouter, according to the report ^[7]. If that share holds through the months after September 22, developers are picking open models for reasons that per-token price does not capture.

The sellers' motive is the main risk to the buy case. Both labs are widely expected to pursue IPOs, and the report describes the cuts as a bet that higher volume will eventually make up for thinner margins on each call ^[10]. OpenAI and Anthropic are giving up margin this year to buy usage ^[10]. A team that retires its own inference hardware on these prices also gives up the fallback it would need if they rise after a listing.

In my view the per-token case for self-hosting routine workloads is weaker than it was in July. A build case still quoting spring API prices is off by a factor of ten on input ^[13]. The counter-case is the IPO: prices set to lift adoption before a listing are the least durable number in the comparison ^[10]. I would be wrong if open-source volume on OpenRouter holds ^[7], or if the cheap tiers get more expensive once the labs are public.