OpenAI Halves API Prices — And Changes the Economics of AI Forever

OpenAI’s Price Cut Is Not a Discount. It’s a Siege.

OpenAI dropped two new models on September 22 — GPT-6 Sol and GPT-6 Luna — and slashed their API prices by 50% compared to the GPT-5.6 generation. Sol costs $2 per million input tokens and $10 per million output. Luna, its lighter sibling, sits at $0.10 in and $0.50 out. For context, the GPT-5.6 equivalents ran $4 and $20 for Sol, and $0.20 and $1.20 for Luna.

This is the most aggressive pricing move in the frontier AI market to date, and it changes the math on everything.

The Real Target: Anthropic’s Enterprise Share

Anthropic launched Claude Opus 5 on September 12 and Claude Opus 5.5 on September 22 — the same day as OpenAI’s announcement. The timing is not coincidental. This is a competitive pincer movement.

OpenAI’s internal benchmarks claim Sol scores 33.2% on AutomationBench, a workflow evaluation suite, while costing 9% of what a task with Claude Opus 5 would cost. That 9% figure is the most important number in this press release. It tells enterprise buyers something dangerous: the cheapest viable model may already be good enough, and OpenAI just made it dramatically cheaper.

For companies that have been hedging between OpenAI and Anthropic, the calculus shifts overnight. If Sol can handle most production workflows at a fraction of Claude Opus 5’s price — and at a parity that still beats Opus 5 on raw score — the rationale for dual-vendor deployments weakens considerably.

Who Wins, Who Loses

Enterprises win. Companies running high-volume inference — customer service bots, code assistants, document processing — will see their AI bills drop by roughly half for the models that actually power most workloads. Luna at $0.10 per million input tokens is effectively commoditized pricing. At that rate, inference becomes a line item, not a strategic budget concern.

Anthropic loses margin. Opus 5 was positioned as the premium alternative to GPT-5.6, commanding a price premium for better alignment and reasoning. Now OpenAI is claiming parity on benchmark performance at 9% of that cost. Anthropic will need to justify its pricing or adjust — neither option is easy.

Chinese model makers face a harder ceiling. Byx and DeepSeek have been the existential worry for OpenAI’s moat. Cutting prices to these levels raises the bar for any competitor trying to compete on inference cost alone. But if capabilities truly plateau at the GPT-6 level, Chinese models may not need to beat OpenAI on price — they just need to be good enough, locally deployed, and politically acceptable.

The open-weight community gains little. These are proprietary models behind OpenAI’s walled garden. The price cut benefits the platform, not the ecosystem. There is no signal that weights will open or that third-party providers will match these rates.

The Plateau Signal Hidden in Plain Sight

The safety documentation reveals something more interesting than the pricing. Sol and Luna are classified as “High” risk for cybersecurity and bio/chemical capabilities — the same rating as GPT-5.6 Sol and Luna. They did not reach the “Critical” tier assigned to GPT-6 Astra, the flagship model.

That distinction matters. OpenAI is explicitly tiering its GPT-6 family: Astra remains the frontier, while Sol and Luna serve the mass market. The capabilities gap between them is deliberate. OpenAI is signaling that for most applications, the top-tier model is unnecessary — and that the second tier is now priced so low that buying the top tier makes economic sense only for the hardest problems.

This is a capability plateau dressed as a product strategy. The real frontier — Astra — has not been discounted. The models being cut in price are the ones good enough for routine work.

The Missing Narrative: Safety Discourse Without Substance

Sam Altman publicly endorsed Dario Amodei’s September 12 essay calling for AI pace-limiting and independent evaluator access. OpenAI announced a misalignment reporting framework on September 16 and a push for U.S.-led international AI standards on September 21. All of this created the impression of a company engaging seriously with governance.

But the Sol and Luna system cards contain none of it. No mention of pace-limiting. No reference to external evaluators. No acknowledgment of the misalignment framework. Astra’s documentation included results from the UK AI Safety Institute; Sol and Luna do not.

This is a calibrated silence. The governance narrative belongs to Astra, the flagship that carries the reputational risk. The volume models — Sol and Luna — are commercial products, and OpenAI is treating them as such. The separation is clean and intentional.

What Changes Next

The pricing floor has moved. Any company building on top of frontier models — API wrappers, agent platforms, vertical AI tools — now faces a lower cost basis for their primary inference layer. Margins compress, but so does the barrier to launching new AI products. Expect a wave of new entrants within six months.

Anthropic will respond. Whether through price cuts, capability leaps with Opus 5.5, or a repositioning around safety differentiation remains unclear. But sitting still is not an option.

OpenAI’s next move is likely another model tier — perhaps a Luna-light at even lower cost — or a push to bundle these models into per-seat enterprise pricing that locks customers into the platform rather than the API.

The economics of AI are being rewritten this week. The question is not whether prices will go lower. They will. The question is whether OpenAI’s pricing power buys it enough time to stay ahead of the capability curve before the next competitor proves otherwise.