The one-screen version: on October 13, Google made Gemini 4 Argon — its first frontier model since Gemini 3 — generally available to developers and businesses through the paid Gemini API in Google AI Studio and Android Studio, with Gemini app and Search AI Mode access for AI Pro and Ultra subscribers. We covered Argon on October 3 as the million-token model you couldn't buy yet — it shipped first only to vetted cyber defenders. The capability didn't change this week. The gate opened. And the price it opened at — $2 per million input tokens, $10 per million output, cached input 95% off — is the exact same headline rate as GPT-6.1 Sol and Claude Sonnet 5.5. Three labs, three flagship coding models, one number. That coincidence is the actual story, and it tells you where to point your attention.
1. The gate opened — on our own timeline#
Two weeks ago this column flagged Argon as a frontier capability with a turnstile in front of it: announced September 30 with a 1-million-token output ceiling, but reachable only through Google's Fairwind cyber-defense program, with paid API access promised "later." Later is now. Reuters reports the model is live for developers worldwide, with early enterprise users including Deloitte, LexisNexis, Ambience Healthcare, and Google's own Isomorphic Labs, and the Fairwind program expanded to hundreds of cybersecurity partners who get a version with the standard cyber guardrails relaxed.
The framing Reuters puts on it is worth keeping: this is Google trying to claw back ground after a rough stretch — a cancelled Gemini 3.5 Pro and a leadership shake-up at DeepMind. Read against that, shipping broad developer access rather than another gated preview is the news. The frontier labs have learned that a model nobody can call is not a product.
What it changes for you: nothing you must do today, but a default worth setting. Google says Argon is built for complex, multi-step work — coding, finance, law, cybersecurity, science — and one third-party index (Vals) ranks it first at launch. Treat that the way you'd treat any launch-day benchmark: a reason to test, not a reason to switch. The per-benchmark numbers circulating on aggregator sites already disagree with each other, which is the usual signal that the real ranking isn't settled yet.
2. Three models, one sticker — which means the sticker is noise#
Here is the part that isn't a coincidence anymore. Anthropic shipped Sonnet 5.5 at $2/$10. A day later OpenAI shipped GPT-6.1 Sol at $2/$10. Now Google opens Argon at $2/$10. Three independent labs, pricing their flagship agentic model at the identical rate, inside a few weeks.
When your three biggest competitors all choose the same price, price has stopped being a weapon and become a convention. The fight moved somewhere you can't see on the rate card.
That convergence is a gift to founders, because it collapses a question that used to be hard. You no longer pick a model by its headline cost — the headline is a tie. You pick on the three things the rate card hides:
- The harness. Argon lives in Antigravity (Ultra-tier) and AI Studio; Sol lives in Codex; Sonnet 5.5 lives in Claude Code. The agent wrapped around the weights touches your workflow all day — it moves your velocity more than a two-point benchmark gap ever will.
- Cache economics. All three now discount cached input heavily (Argon's cached reads are 95% off). A coding agent re-sends the same file tree and system prompt every turn, so your real bill is set by the cached-read rate and tokens-per-task, not the base price that happens to match.
- The clock. This is the one most people miss. Google's footnote lists a standard rate of $4/$20 — double the launch price — once the introductory period ends, and Google has not published the end date. The $2/$10 you're budgeting against is a promo. Model your costs at $4/$20 and the promo becomes upside instead of a surprise.
3. The coda: the model layer just finished commoditizing#
Two weeks ago we argued the model had stopped being the scarce input and the money had moved to deployment. A field-wide convergence on one price is the clearest confirmation yet. When three frontier labs can't differentiate on cost, the frontier model is a commodity input — priced, increasingly, like one.
Commodities reward buyers who stay liquid. The move is the same one we keep landing on, now with a third vendor to make it concrete: keep your inference swappable at the API boundary, so that when this week's benchmark leader is next week's second place, you move in an afternoon. Run the same ten real tickets through all three harnesses, measure tokens-per-task and cache hit rate on your own repository, and let your numbers decide — because the launch charts can't, and the price no longer will.
The one line to take into the week: there are now three frontier coding models at $2/$10, so stop shopping on the sticker and start shopping on the harness and the cache — and whichever you pick, budget it at $4/$20, because the price you see is an introductory one with a hidden expiry.
