OpenAI Slashes GPT-5.6 Sol Prices in Half — But Only on OpenRouter and Vercel AI Gateway

OpenAI has opened a new front in the AI price war, and the location is as interesting as the discount itself. From August 17 through September 18, the company’s flagship GPT-5.6 Sol model is available at 50% off — but only for requests routed through OpenRouter and Vercel’s AI Gateway. Developers calling the same model through OpenAI’s own API or Azure still pay full list price. The move turns two third-party inference aggregators into privileged channels, and it signals exactly where OpenAI believes the next battle for developer wallets will be fought.

What exactly is discounted

GPT-5.6 Sol, released July 9 as the flagship of the GPT-5.6 family, normally lists at $5 per million input tokens and $30 per million output tokens. Through the promotion, OpenRouter and Vercel AI Gateway customers pay $2.50/$15 per million — with batch pricing dropping to $1.25/$7.50. OpenRouter’s live pricing API already reflects the cut, and the model’s listing page now carries a prominent “50% off” badge. Vercel’s changelog confirms the discount applies to all requests through AI Gateway during the month-long window, with the gateway mirroring provider pricing and adding no markup of its own.

This is not OpenAI’s first aggressive price move this summer. On July 30, the company cut GPT-5.6 Luna’s API price by 80% (to $0.20/$1.20 per million tokens) and dropped GPT-5.6 Terra by 20% (to $2/$12), framing it as “advancing the price-performance frontier.” Two weeks ago, OpenRouter layered a further 50% promotion on Terra and Luna, briefly bringing Luna down to $0.10/$0.60 per million. Sol — the reasoning and coding flagship — was conspicuously excluded from that earlier round. Its inclusion now, at half price, completes the sweep across the entire GPT-5.6 stack on third-party routers.

Why channels matter more than the discount

The unusual part is the exclusivity. Price cuts are routine in this market; channel-exclusive price cuts are not. SemiAnalysis called it “potentially a clever marketing play,” and the mechanics support that reading. OpenRouter publishes real-time token-usage rankings that the entire AI industry watches — model providers, investors, and enterprise buyers alike treat them as a de facto market-share tracker. A 50% discount that pulls price-sensitive traffic onto the platform during a heavily watched month effectively buys position on that chart.

The stakes are real. OpenRouter’s own “State of AI” research — an empirical study built on 100 trillion tokens of observed usage — documented how volatile developer loyalty is at the routing layer, where switching models is a one-line config change. Anthropic’s share of OpenRouter traffic, once dominant, has been eroding through 2026 as cheaper open-weight models from China and aggressive OpenAI pricing compressed everyone’s margins. Against that backdrop, subsidizing flagship-model traffic on the most-watched leaderboard is cheap advertising measured in tokens instead of billboards.

The early returns are visible. The Information reported on August 18 that “OpenAI makes gains against Anthropic among OpenRouter customers,” noting that OpenAI has found in OpenRouter a new place to fight Anthropic for developer wallets. OpenRouter’s discounted-models collection shows GPT-5.6 Sol already pushing more than 826 billion tokens through the platform — before the flagship discount even had a full day to compound.

Developers are not entirely cheering

The Hacker News discussion of the cut hit the front page with over 600 points and 400+ comments, and the mood was less celebratory than OpenAI might have hoped. A recurring complaint: OpenAI is cutting API prices in half while ChatGPT subscribers have seen their usage limits slashed, leaving paying consumer customers subsidizing a market-share offensive they cannot benefit from. Others noted the discount’s fine print — DigitalOcean, for instance, offers its own 50% Batch API discount outside OpenRouter’s listing, making “cheapest channel” a genuinely nontrivial calculation.

There is also healthy skepticism about durability. Promotional pricing that expires September 18 is not a structural cost reduction; it is a customer-acquisition spend. Developers who re-architect their stacks around $2.50 Sol and then face a reversion to $5 list price in the fall will have learned something about vendor lock-in the hard way. The Chris Strobl take circulating on LinkedIn — that this is “a chart play” — captures the cynicism: the goal is the usage graph, not necessarily the long-term economics.

The competitive read

For Anthropic, the pressure is direct. Claude models have been the default choice for coding and agentic workloads among developers, and OpenRouter is precisely the venue where that preference is measured daily. Anthropic recently made Sonnet 5’s introductory pricing permanent at $2/$10 per million tokens — undercutting even discounted Sol on input price, though Sol positions itself on throughput and latency (proponents cite roughly 4x faster generation and 5x lower first-token latency versus typical hosted alternatives). A month of half-price flagship access is designed to let developers re-run their benchmarks and re-rank their defaults.

For the broader market, the episode confirms that AI pricing power now lives at the routing layer. When a model provider will pay — in margin — to appear cheaper on a specific aggregator, aggregators have become distribution channels of the first order. Vercel positioning AI Gateway as a zero-markup mirror of provider pricing while hosting exclusive discounts is a quietly important precedent.

Whether the chart play works will be legible within weeks: OpenRouter’s public rankings will show whether Sol’s discounted token volume holds, grows, or evaporates when the promotion ends on September 18. Either way, the era of flat, channel-agnostic model pricing is over. The price you pay for the same model now depends on where you ask for it — and providers are betting real money that most developers will ask where it’s cheapest.