Gemini 4 Argon Hits $2/$10 — Half Opus 5.5, a Fifth of GPT-6 Astra, and the AI Stack Just Re-priced

Alphabet announced Gemini 4 Argon on September 30, 2026 with two price points: $2 input / $10 output per million tokens introductory, then $4 / $20 after the introductory window expires. OpenAI's GPT-6 Astra — the model Google is most directly competing against — is priced at $10 / $50. Anthropic's Claude Opus 5.5, released September 22, costs $4 / $20, the same as Argon's standard price. Argon's introductory window sits at a fifth of GPT-6 Astra's list price and at half of Opus 5.5's, with a 95% cache discount on top.

Google DeepMind's own evaluation page shows Argon winning or tying GPT-6 Astra on 13 of 18 head-to-head comparisons in Google's release package, with deeper strengths on long-horizon coding and knowledge-work tasks. The reported numbers are real wins, but they are partial — Google self-reports several of the most flattering coding scores, including DeepSWE v1.1, where Argon posts 77.9% using a mini-swe-agent harness. Independent leaderboard timings, reported by YouTube benchmark aggregators citing third-party tests, give Argon 12 wins, 1 tie, and 5 losses against GPT-6 Astra.

Alphabet does not need frontier per-token margins the way OpenAI does. The Q2 FY2026 earnings release filed July 22, 2026 shows Google Cloud revenue up 82% YoY to $24.8B with $8.8B in segment operating income — three times the $2.8B a year ago. That is the segment that runs Gemini, runs the TPUs, and now sells TPUs to outside customers (Google recognized TPU-system revenue from external customers for the first time in Q2). Capex in Q2 doubled YoY to $44.9B, and full-year 2026 capex guidance was raised again in July to $195–205B. A $49.6B equity raise closed in June explicitly funds that build-out.

OpenAI and Anthropic both rely on rented compute. Microsoft's Azure is the primary host for GPT-6 Astra; Amazon's AWS is Anthropic's largest. Both models bill the labs for every token their frontier API customers spent. When Alphabet sets the headline rate at $2/$10, the bill for the same workload at Azure or AWS falls by a multiple of that ratio. That is a price war Alphabet can fight on gross-margin economics its competitors cannot match — Google Cloud is no longer a marginal segment subsidizing a frontier demo.

Argon is trained and served on Google's in-house TPUs, which means the price pressure it creates flows upstream to the silicon stack rather than getting absorbed by Alphabet's P&L alone. Broadcom co-designs Google's TPU silicon and reported AI revenue of $8.4B in Q1 FY2026, up 106% YoY, with the Anthropic-Google-Broadcom compute expansion adding an estimated 3.5 GW of TPU demand from 2027. The chips themselves are manufactured at TSMC, which raised 2026 capex guidance in mid-2026 to $60–64B — at least $4B above prior guidance — on AI demand. NVIDIA does not build the silicon that runs Argon but still captures the inference-GPU share for Azure and AWS-hosted frontier deployments.

GPT-6 Astra launched September 3, 2026 at $10/$50 — four weeks before Argon. That timing now looks like a misread: OpenAI priced as the frontier premium, and three weeks later Alphabet entered at the bargain entry. Anthropic cut Opus 5 by 20% on the way to Opus 5.5 (from $5/$25 to $4/$20) on September 22, an unusual pre-emptive move for a frontier vendor. Reuters also reported Alphabet's previous communications had shifted toward emphasizing cost advantage versus competitors, so the introductory $2/$10 is a continuation of that posture, not a one-off.

Microsoft and Amazon each host a frontier lab whose flagship just got repriced. Microsoft's Q2 FY2026 results showed Azure and other cloud services revenue up 39% with22–26 points of that growth from AI specifically — the segment most exposed if GPT-6 Astra demand deflates. Amazon's AWS carries the same exposure for Anthropic, with the additional complication that Anthropic's revenue run-rate surpassed $30B in 2026 — the faster it scales, the more per-token price matters to Amazon's hosting economics. Neither hyperscaler loses their customer, but both face a margin squeeze if the labs they host have to defend list pricing against a $2/$10 benchmark.

Argon changes the prior on what Alphabet is worth. Through most of 2024 and the first half of 2025, the bear case was that the frontier had moved past Google: OpenAI shipped on cadence, Anthropic won coding mindshare, and Gemini looked late. Google Cloud's Q2 result — $24.8B revenue, $8.8B operating income, a $514B backlog, plus the first quarter of TPU-system sales — already reframed the segment as a growth business rather than a side project. Argon adds the model layer on top of that stack, and at half the price of Opus 5.5 with comparable benchmark performance.

The risk is execution: only paid API customers and Google AI Ultra subscribers have access today; the model is rolled out to a vetted group of cybersecurity partners first through the Fairwind program, with no public-release timeline. Reuters reporting notes Google also dropped plans for an intermediate Gemini 3.5 Pro release, and CEO Sundar Pichai signed a voluntary AI safety accord with the U.S. administration a day before launch. Frontier lead is fragile — Microsoft and Meta ship their own models, and the next 12 weeks will determine whether Argon's pricing sticks as the new market rate or gets matched and erased.

Short-term (days to quarters), the question is whether Argon stays at $2/$10 once the introductory window expires. If Google lifts pricing to the standard $4/$20 within the next two quarters, the market will read that as the real reset — and as a margin-friendly outcome, since cached-input discount effectively halves the realized rate on most production workloads. If the introductory price holds longer, the pressure on Microsoft Azure-hosted GPT-6 and Amazon AWS-hosted Claude gets sharper, and the cheapest AI-adjacent trades are the ones that capture volume rather than list price.

Long-term (one to three years), the structural shift is which company owns the inference compute stack. Alphabet controls its own silicon (TPU, designed with Broadcom, manufactured at TSMC). Microsoft and Amazon buy from NVIDIA and increasingly from AMD and in-house silicon. Whoever runs the cheapest silicon per useful token by 2028 wins the bulk of the inference market, and the HBM4 supply from SK Hynix, Micron, and Samsung Electronics is the bottleneck that determines how fast each side can scale. Argon's price is the announcement; the silicon roadmap is the story.