OpenAI Cuts GPT-6 Prices 50% as Anthropic Fires Back
OpenAI slashes GPT-6 Sol and Luna prices 50% the same day Anthropic drops Claude Opus 5.5 pricing 40%. The AI margin compression story accelerates.
Key Points
- GPT-6 Sol lands at $2.00/$10.00 per million input/output tokens — exactly half of Claude Opus 5.5's $4.00/$20.00 rate, announced within minutes of each other on September 22.
- OpenAI is simultaneously undercutting both the premium closed-model market (Anthropic) and the low-cost open-weight market (DeepSeek), a structural escalation that signals deliberate margin destruction at the model layer.
- Watch Sonnet 5.5 and Haiku 5.5 pricing when Anthropic releases them in coming weeks — if those mid-tier cuts follow the Opus 5.5 pattern, the commoditization cascade accelerates into enterprise contract renewal season.
GPT-6 Sol costs $2.00 per million input tokens and $10.00 per million output tokens — exactly half of what Claude Opus 5.5 runs on both dimensions, a price point OpenAI and Anthropic reached in simultaneous announcements on September 22 that collectively erased the pricing floor the frontier AI industry spent two years constructing.
The Numbers That Ended the Premium Model Era
Start with the arithmetic, because it's brutal. OpenAI's GPT-6 Sol is priced at $2.00 per million input tokens and $10.00 per million output tokens, down from $4.00 and $20.00 respectively — a clean 50% reduction across the board. Its companion model, GPT-6 Luna, goes further: $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20 and $1.20. Luna doesn't just undercut Anthropic — it undercuts DeepSeek V4.1 Flash's best off-peak rate of $0.15/$0.60, and it blows past DeepSeek's peak rate of $0.30/$1.20 by a factor of three on output. That is a closed, American frontier lab pricing below a Chinese open-weight model widely considered the cost-efficiency benchmark of the industry.
Anthropic's cut, announced the same day, was itself significant before OpenAI's counterpunch landed. Claude Opus 5.5 dropped 20% on the sticker price, but Anthropic was careful to note that on a typical enterprise workload, the effective reduction runs to 40% once cache read pricing is factored in. Cache reads — the mechanism that allows frequently accessed context to be stored and retrieved without reprocessing — fell 60%, from roughly 50 cents to 20 cents per million tokens. For high-volume enterprise customers who are running repetitive agentic workflows over shared codebases or document corpora, that 60% cache cut is the number that actually matters to their monthly bill, not the headline token rate.
The performance benchmarks, all vendor-reported and therefore subject to the usual selection-bias caveats, complicate any argument that cheaper means worse. On Agents' Last Exam, a frontier reasoning benchmark, GPT-6 Sol at maximum effort scored 56.4%, which OpenAI claims beats Claude Opus 5's best result at 60% lower cost per task. On the DeepSWE software engineering benchmark, Sol scored 68.8% against Claude Fable 5's 69.9% — a difference of 1.1 percentage points — at approximately 80% lower cost per task. On OSWorld computer use evaluation, Sol at extra-high effort scored 60.5% versus Opus 5's 60.3%, again at roughly 80% lower cost. The performance deltas are inside the margin of noise; the cost deltas are not.
What OpenAI Is Actually Doing Here
This is not a startup buying market share it can't afford. OpenAI is the most valuable private company in the AI sector, and it is making a calculated decision to destroy the margin structure of the model layer — including its own existing margin structure — before a competitor or an open-weight model does it for them. The 90% discount on cached input-token reads, buried in the announcement, is the most aggressive single line item. OpenAI says its caching architecture already cut the volume of fresh tokens processed by more than half across billions of GitHub Copilot requests. That means the cost to serve at these prices is lower than the sticker arithmetic suggests — but it also means every competitor who cannot match that infrastructure efficiency is now being squeezed on both the revenue and the cost side simultaneously.
The rollout scope confirms this is a volume play, not a niche offering. GPT-6 Sol and Luna are available via ChatGPT Work and Codex across Plus, Pro, Business, Enterprise, and Education tiers, as well as the full API. Free and Go tier users get Luna access on desktop. That distribution footprint — from the free consumer tier through the API to enterprise — is designed to maximize token volume, which is the only metric that matters when you're running at these prices. OpenAI is betting that the data flywheel and infrastructure efficiency it gains from processing vastly more tokens at lower prices creates a compounding advantage no competitor can replicate without the same capital base.
Anthropic's Opus 5.5 safety story deserves a separate read. The company reported that attempts to circumvent containment boundaries on Opus 5.5 dropped 85% compared to Opus 5 on its automated behavioral audit. That number is being treated as a marketing point in most coverage, but for enterprise buyers in regulated industries — financial services, healthcare, government contracting — it is a procurement criterion. Anthropic's safety differentiation thesis is that it can hold premium pricing not on raw performance, where Sol is now competitive at a fraction of the cost, but on risk-adjusted reliability for high-stakes deployments. Whether that thesis survives contact with Sol's benchmark sheet is the central commercial question for Anthropic's next two quarters.
What Traders Should Watch Next
The immediate trading implication is structural, not episodic. The AI model layer is in active commoditization, and the September 22 announcements mark the point at which that process became impossible to dismiss as the work of open-weight upstarts. When the largest closed-model lab is pricing below a Chinese open-weight competitor on the same day it undercuts the second-largest closed-model lab by 50%, the revenue story for any company whose primary value proposition is LLM API access is deteriorating faster than most sell-side models currently reflect.
The real money in this environment sits upstream and downstream of the model layer. Upstream means chips and infrastructure — the companies supplying compute to train and serve these models have pricing power precisely because the model producers are giving up margin to win volume. Downstream means enterprise software companies that have embedded AI capabilities into workflows where the switching cost is high and the value capture happens at the application layer, not the token layer. Pure-play LLM API exposure — whether through direct investment in private AI labs or through enterprise software companies whose AI differentiation rests entirely on model access — deserves a harder look at current valuations.
The next specific catalyst to price is Anthropic's release of Sonnet 5.5 and Haiku 5.5, due in the coming weeks. Sonnet is Anthropic's volume workhorse — the model most enterprise customers actually run at scale, not the flagship Opus tier. If Sonnet 5.5 pricing follows the Opus 5.5 pattern and comes in at 40% effective reduction on typical workloads, that is when the enterprise contract repricing cycle begins in earnest. Q4 2026 enterprise renewal conversations are happening now; buyers with awareness of the current pricing environment have every incentive to delay commitment. Watch Anthropic's Sonnet 5.5 announcement date and pricing as the next binary read on whether this price war has a floor — or whether September 22 was just the opening move.
The Weekly Investor
Daily market analysis for active traders. Free.
