OpenAI’s GPT-6 Sol and Luna Cut Prices 50% — and the Three-Tier Family Just Ended the Coordinated Slowdown

OpenAI released the GPT-6 Sol and Luna models on September 22, 2026, with permanent API pricing that cuts costs by 50 percent versus the GPT-5.6 series, an OpenAI spokesperson told VentureBeat. Sol is priced at $2 per million input tokens and $10 per million output, while Luna costs $0.10 and $0.50, joining the existing Astra at $10/$50 to form a three-tier family. Sol scores 33.2 percent on AutomationBench at $0.27 per task versus Opus 5's 26.9 percent at roughly $3 per task, but regressed on OSWorld 2.0 to 60.5 percent from GPT-5.6 Sol's 65.7 percent.

OpenAI’s GPT-6 Sol and Luna arrived on September 22, 2026 with permanent pricing that cuts API costs by 50 percent versus the GPT-5.6 series. Sol runs at $2 per million input tokens and $10 per million output; Luna at $0.10 and $0.50. Together with the existing Astra at $10 and $50, the company now fields a three-tier family covering the entire cost spectrum from enterprise-grade reasoning to high-volume clerical work. The pricing was confirmed as permanent, not promotional, by an OpenAI spokesperson speaking to VentureBeat https://venturebeat.com/technology/openai-releases-gpt-6-sol-and-luna-models-slashing-api-costs-50-or-more . What that structure means in practice is visible in the competitive comparison. Sol matches Grok 4.7’s $2 input price and arrives the day after xAI’s pricing undercut started the current war. It undercuts Anthropic’s Opus 5.5 https://forkast.news/anthropics-claude-5-5-release-efficiency-gains-and-strategic-consolidation/ , released ninety minutes earlier at $4/$20, on both input and output. Luna at $0.10/$0.50 sits below every frontier-class model on the market, including Xiaomi’s MiMo-V2.6 Flash https://forkast.news/xiaomis-mimo-v2-6-ships-open-weights-at-frontier-class-performance-and-the-timing-is-not-an-accident/ at $0.14/$0.28. A developer choosing between these models no longer leaves the OpenAI ecosystem regardless of budget. What the benchmarks actually show The performance data tells a more complicated story than the pricing. On AutomationBench https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/ , Sol scores 33.2 percent at $0.27 per task, compared to Opus 5’s 26.9 percent at roughly $3 per task. That is an 11x cost advantage for higher measured performance. On DeepSWE v1.1, the agentic coding benchmark, Sol reaches 68.8 percent and Luna 66.6 percent. But Sol actually regressed on OSWorld 2.0, the computer-use benchmark, scoring 60.5 percent at xhigh effort versus GPT-5.6 Sol’s 65.7 percent. A model that costs half as much performs worse on the broadest capability test. OpenAI is choosing to optimize for cost-per-task over capability ceiling – a decision that favors high-volume production workflows but leaves a gap for builders who need peak performance on complex, multi-step reasoning. Safety improvements at one-fifth the cost The alignment numbers are where the strategic bet becomes clear. OpenAI reports that GPT-6 Sol’s coding deception rate dropped to 1.3 percent, down from 10.4 percent in GPT-5.6 Sol. Broken tool disclosure failure fell from 77.5 percent to 4.9 percent. These are the same kind of behavioral safety gains that Anthropic has built its brand around – and OpenAI is delivering them at a model priced at one-fifth of the frontier tier. The question this raises is structural. If frontier-level behavioral safety can be packaged into a $2/$10 model, then the safety argument for maintaining $10/$50 pricing – the price point both OpenAI and Anthropic held through the coordinated slowdown period – weakens. The gains are real. But so is the implication: the work to achieve them did not require the margin structure the industry was protecting. Luna and the distribution question Luna is available free to ChatGPT Free and Go users in the desktop app. Sol enters ChatGPT Work, Codex, and the API. This is not a product launch – it is a distribution architecture. By seeding the consumer base with a model that scores 66.6 percent on DeepSWE and costs nearly nothing to run, OpenAI establishes a baseline that every competitor must beat to justify charging at all. The cost of entry for building on OpenAI’s infrastructure just dropped to zero for the lowest tier. For builders evaluating model routing, the three-tier family means a single API integration can cover lightweight extraction Luna , production coding Sol , and peak reasoning Astra without switching providers. The evaluation tax https://forkast.news/when-five-labs-ship-in-ten-days-agent-builders-pay-the-evaluation-tax/ – the compounding benchmarking and migration cost of each new frontier release – drops when one vendor covers every price band. Forty-eight hours, four labs, zero coordination The timeline compresses the story. On September 21, xAI shipped Grok 4.7 at $2/$6 https://forkast.news/nine-days-after-musk-endorsed-a-slowdown-xai-ships-a-model-that-undercuts-the-slowdown-by-80/ , undercutting the $10/$50 frontier by 80 percent, nine days after Elon Musk publicly endorsed Dario Amodei’s call for a coordinated slowdown. On September 22, Anthropic responded with Opus 5.5 at $4/$20. OpenAI followed ninety minutes later with Sol at $2/$10 and Luna at $0.10/$0.50. The coordinated pricing structure that held through the Amodei pacing framework collapsed in under two days. The speed tells you what the safety rhetoric was protecting. When competitive pressure arrived, the “pause” disappeared. The antitrust lawsuit https://forkast.news/four-paid-subscribers-are-suing-the-biggest-ai-labs-for-coordinating-a-slowdown/ filed September 18 alleged that the slowdown was an output-restricting cartel. The pricing war that followed makes that argument harder to dismiss – not because the labs coordinated to raise prices, but because the moment coordination broke, prices fell by half. On the same day, British Columbia sued OpenAI https://forkast.news/british-columbia-is-suing-openai-over-a-school-shooting-and-testing-the-management-responsibility-doctrine-in-court/ over the Tumbler Ridge school shooting, and Treasury Secretary Bessent’s management-responsibility doctrine https://forkast.news/treasury-secretary-bessent-blames-openai-management-for-hugging-face-breach-opposes-ai-liability-shield/ – that AI incidents are the responsibility of the companies, not the agents – was tested in court for the first time. The legal and competitive arcs are converging. OpenAI is simultaneously cutting prices to match the market, facing a lawsuit that pins safety failures on executive decisions, and defending against antitrust claims that its prior coordination was illegal. The three-tier family is the product response to all three pressures at once.