For the first time in the industry’s history, Anthropic passed OpenAI in enterprise AI market share. The Wall Street Journal reported on June 11 that OpenAI is weighing drastic token price cuts in response. GPT-5.6 Luna’s $1 per million token input price — announced June 26 — appears to be that response starting to land.
Here is what the market shift means, what price cuts are likely to look like, and what to do before the next announcement.
How the Market Flipped
The clearest read on where enterprise money is flowing comes from Ramp’s AI spending index, which tracks actual software purchasing across its business-card customer base.
May 2026 Ramp AI Index: Anthropic 34.4%, OpenAI 32.3% — the first time Anthropic has led.
A March 2026 Ramp figure drove the point harder: Anthropic captured 73% of all enterprise spending among companies buying AI tools for the first time. That is the new-customer pipeline, the leading indicator for where total market share goes next.
Revenue followed the same direction. Anthropic went from roughly $1 billion annualized in January 2025 to $47 billion annualized by May 2026 — a 47x increase in sixteen months. OpenAI grew in the same period from $13 billion to approximately $25 billion annualized — still large, but roughly 2x growth while Anthropic posted 47x.
The valuation gap inverted on a similar timeline. Anthropic’s valuation surpassed OpenAI’s in May 2026 — $965 billion versus $852 billion — ending a streak of OpenAI valuation leadership that dated to the company’s founding.
The mechanism behind Anthropic’s growth is not complicated: Claude Code drove “mind-blowing growth” among developers and pushed Anthropic into its first profitable quarter since founding in 2021. The $200/month Max plan converted a large base of individual developers into recurring revenue, and enterprise Claude Code licenses added API token billing on top of seat fees.
OpenAI’s Response
The Wall Street Journal reported on June 11 that OpenAI is “preparing to slash the prices it charges for AI tokens” with cuts aimed explicitly at winning customers from Anthropic. Sam Altman acknowledged that “costs had become a huge issue” for enterprise customers. No specific cut amounts, model targets, or timing were announced at time of reporting — discussions were “still in flux.”
GPT-5.6, which launched on June 26 as a limited preview for approximately 20 government-approved organizations, provides the first concrete pricing signal:
| Model | Role | Input per 1M | Output per 1M |
|---|---|---|---|
| Sol | Frontier — agentic, coding, biology | $5.00 | $30.00 |
| Terra | Balanced — everyday work, 2x cheaper than GPT-5.5 | $2.50 | $15.00 |
| Luna | Fast, cheap — high-volume routine tasks | $1.00 | $6.00 |
Luna at $1/$6 per million tokens is a real repricing signal. DeepSeek V4-Flash is $0.14/M input, so Luna is not price-competitive with open-weight alternatives — but it is dramatically cheaper than GPT-5.5’s $5/$30 and positions OpenAI to compete on cost-sensitive tiers it was previously ceding.
Three likely next moves from OpenAI, in rough probability order:
- GPT-5.5 cuts 30-50% once GPT-5.6 reaches general availability — lowering the everyday-tier cost to pressure Anthropic on Sonnet 4.6 and Haiku 4.5 pricing
- Enterprise volume discounts protecting headline per-token pricing while reducing effective rates for large-volume customers
- No further headline cuts before IPO — protecting gross margin narrative for S-1 investor presentations (OpenAI filed its S-1 confidentially on June 8; Altman told staff a public listing within 12 months)
The IPO timing is the complicating factor. A sustained price war compresses margins at a moment when both companies are preparing public market narratives. Anthropic is targeting an October 2026 IPO. Infrastructure costs are rising: inference is projected to go from $8.4 billion in 2025 to $14.1 billion in 2026. Pricing down while costs go up is a difficult story for either company’s S-1.
The most likely outcome is a carefully staged sequence: GPT-5.6 GA (coming weeks) with Luna’s $1 price broadly available, followed by GPT-5.5 cuts framed as “making the previous generation accessible” while the new flagship holds pricing.
What the Efficiency Shift Means for Your Budget
The tokenmaxxing era — where developers burned tokens as a proxy for productivity — is well documented. The shift to efficiency is already visible in the same enterprise data.
Lindy, an AI startup, switched 100% of its traffic from Claude to DeepSeek, citing cost discipline over capability preference. Uber, which burned through its entire 2026 AI budget in four months, implemented spending tiers capped at $1,500/month per engineer. Finance teams that were invisible when AI spend was small are now active participants in model selection decisions.
This is the new baseline: CFO scrutiny is standard, ROI measurement is expected, and “we use the best model” is no longer a complete justification for a bill that grew 4x in a quarter.
Price cuts from OpenAI do not change this dynamic. If anything, cheaper tokens make the ROI case easier to quantify — and harder to avoid quantifying.
Builder Audit: Four Things to Do Before the Next Announcement
1. Baseline your current API costs by model and workload. When OpenAI announces cuts, you want to calculate the savings immediately — not reconstruct three months of billing history. Export your usage dashboard now and tag it by workflow type.
2. Avoid locking annual contracts at current pricing. OpenAI’s discussions are “still in flux” but directionally clear. Any contract signed in the next 60 days should include a “most-favored pricing” clause that passes future price reductions through automatically — or use quarterly terms.
3. Audit which workloads actually need frontier pricing. Luna at $1/$6 and DeepSeek V4-Flash at $0.14/$0.28 cover a large share of production workloads that are currently running on $5/$30 models because nobody made the routing decision explicitly. Make it explicit now, before prices change and the urgency disappears.
4. Benchmark Sol and Fable 5 against your actual tasks, not published scores. Sol (GPT-5.6) and Fable 5 post different benchmark results — Sol’s TerminalBench is 88.8%, Fable 5’s SWE-Bench Pro lead is 22 points over GPT-5.5. Neither tells you which model performs better on your codebase, your retrieval pipeline, or your document workflows. When GPT-5.6 reaches general availability, benchmark both on your highest-value tasks. Published scores are starting positions, not decisions.
The price cut is coming. The exact timing and magnitude are uncertain, but the direction is not: OpenAI has lost the enterprise market lead it held for three years and has clearly signaled repricing as its response. Builders who have baselined their costs, negotiated flexible contracts, and routed workloads appropriately will capture the savings immediately when the announcement lands. Those who have not will get the same price — and spend a month figuring out what to do with it.