Gemini 4 Argon turns context into a metered budget

Google's newest frontier model now has a memory dial. Inside an insiders build of Antigravity, Google's agentic coding platform, Gemini 4 Argon reportedly shows three context sizes: 256K ⁰², 512K and 900K tokens. TestingCatalog reported the selector on October 11, 2026 ⁰⁶. The biggest option holds about 3.5 ²⁵ times what the smallest holds.

The dial has a price on it. TestingCatalog reports that a 512K ¹¹ option observed in an Antigravity configuration burns about 1.3 times the quota per turn. The 900K ¹² option burns about 1.8 times. Google claims Argon scores 77.9% ⁰¹ on the DeepSWE coding benchmark, so a strong model sits behind every setting.

The evidence is thin, and we want to be upfront about that. Google has not announced these tiers. Lumina reports spotting them as apparent context options, not confirmed API limits, in the interface of a bundled Antigravity CLI build and posted the finding on X on October 10. The quota figures are interface hints, and Google has published no pricing to back them.

Context Becomes a Metered Budget Line

The mechanism is metered context. A context window is how much text a model can read in one turn, measured in tokens. Argon's selector takes that window, which used to be one fixed number, and makes it a choice with a cost per turn. Every task now buys a size.

The tiers sort into three parts. The 512K ²⁶ tier is the step-up: double the room for about 30% more quota. The 900K ²⁷ tier is the ceiling, with 3.5 times the room for about 80% more quota.

This isn't the first time the pattern has appeared. TestingCatalog notes that the Argon options resemble existing context settings for Gemini Flash models. One model with a dial could be an experiment. Two model families with the same dial look like a product decision.

The per-token math goes in a surprising direction. At 900K ²⁸, each available token costs about half the quota it does at 256K, because 1.8 divided by 3.5 is roughly 0.51. So bigger looks cheaper per token. That only holds if the extra tokens are doing useful work.

Inside the Antigravity Tier Selector

Think of the tiers as rental vans for a house move. The big truck costs 1.8 ¹² times the small van and holds 3.5 ²⁵ times the boxes. Nobody is happier with the truck if half the boxes are full of junk that belonged in the bin. On the reported hints, the multiplier applies to every turn at the chosen setting.

Google has not published what a quota unit costs, so here it is in round numbers. A 512K ³² turn then costs $1.30, and a 900K turn costs $1.80 ²⁹.

Retries are where the bill grows. One agent turn means the model reads its context and then acts. A failed turn at 900K ²⁹ still charges $1.80. An agent stuck looping on a broken test can rack up six failed turns, which comes to $10.80 before anything ships.

Quality is the bigger risk, though. Liu and colleagues' Lost in the Middle research found that models often use relevant information at the edges of a long prompt better than information buried in the middle. The pattern held in multi-document question answering and in key-value retrieval. Push 900K tokens ⁰² of repository into the window, and your most important interface file might land in that weak zone. The model will still answer fluently. That makes the miss harder to spot.

Repository size and relevant context are different numbers. A focused 256K ⁰⁷ packet can beat a raw dump of the whole codebase. Build it from interfaces and tests first, then add recent diffs and error traces. The agent can only reason well over what it can actually find, and every irrelevant file you add dilutes the files that matter.

Latency is still an open question. Google has not published response times for any of the three tiers. Reading nearly a million tokens per turn will not be quick, and a slow turn in an interactive editor costs developer attention too.

It's unclear whether the 1.8x ¹² figure tracks Google's real compute cost or just rations capacity during a limited rollout. TestingCatalog flags another unknown: how the context dial interacts with Argon's reported separate effort setting in Antigravity's CLI. Lumina's findings show both controls side by side in the same model entry. Nobody outside Google knows yet whether a high effort setting at 256K ⁰⁷ beats a low effort setting at 900K.

We think the right default is boring. Run 256K ²⁶, and make the agent earn every step up with evidence that it needs more room. Most coding tasks touch a handful of files. Pay ceiling prices only for the tasks that touch the whole building.

Where the bigger window quietly costs more

2031: Agents That Book Their Own Tier

The dates say a lot about packaging. Google launched Antigravity on November 18, 2025 ¹⁴, according to its developer blog. It announced Argon on September 30, 2026 ²², giving a set of trusted cyber defenders access through its Fairwind Program. On October 9, 2026 ¹⁶, a Google Cloud release note said new Gemini Code Assist Standard and Enterprise subscriptions could no longer be purchased, and pointed buyers to Antigravity.

The tier selector surfaced a day later. So a paid coding product closed to new buyers, and within 48 hours its replacement gained a metered dial. We read that as Google building a menu it can sell through the Gemini Enterprise bundle. Enterprise customers can then be charged by workload intensity, while the model's advertised maximum stays the same.

Raw capacity is growing faster than anyone's ability to use it. Google's announcement says Argon's output limit jumped from 64K tokens ¹⁹ to 1 million, a 15.6x ³⁰ increase in one generation. That is an output limit, and Google has not published an official input window. If ceilings keep growing at anything near that pace, raw size stops being the scarce resource well before 2031.

What stays scarce is quota and wall-clock time. So does the model's ability to find the right fact. Some napkin math shows why routing wins. Picture a team running 1,000 ³¹ agent turns a day, all at 900K, for 1,800 quota units. Instead, route 80% ³² at 256K, and split the rest with 15% at 512K and 5% at 900K. The small tier then costs 800 units, the larger two add 195 plus 90, and the bill comes to 1,085 units.

That mix saves about 40% ³³ every day the router runs. And the risk is lopsided. A router costs a weekend to build. Skipping it means paying up to 1.8 ¹² times on turns that never needed the room.

By 2031, we expect agents to pick their own tier per turn, the way databases pick a query plan. Two things would change that call. One is official Argon API pricing from Google that flattens the multipliers. The other is task-level tests showing that 900K ¹² lifts pass rates enough to cover its quota.

Wire Staged Escalation Into Your Agent Loop

The build is a small router that starts every task at 256K ⁰² and escalates only on evidence. A router here is a few lines of code that choose a context setting before each model call. It works on any model with sized context options, so you can ship it now and point it at Argon when access opens.

Step one: log every task. Record the tier, input tokens, output tokens, turn count, wall-clock time and whether the tests passed. Keep the log in a plain spreadsheet or a single database table. Without this log, every later step is guesswork.

Step two: budget the window before retrieval. Retrieval is the step where your agent pulls files into the prompt. Start with roughly half the window for source files and interfaces and about a fifth for tests. Give the remainder to docs, diffs and system instructions, then tune the split against your pass rate.

Step three: define the escalation triggers. Escalate when the agent hits an unresolved reference, finds contradictory evidence or fails the same tool call twice. Before stepping up, fetch only the missing dependencies and retry at the same tier. Many failures clear at 256K ⁰⁷ once the right file is in the prompt.

Step four: escalate in order. Use 512K ¹¹ for cross-module refactors and dependency migrations. Use 900K ¹² only when the failure looks like missing global context and you have ruled out weak planning. Write the reason for each escalation into the log from step one.

Step five: measure cost per successful task. Divide the total quota spent by the number of tasks that passed. Run the same 20 tasks at each tier and compare the results. If 900K ²⁸ does not raise the pass rate by more than its 1.8x ¹² cost, keep it off by default.

Move a key interface file from the top of the prompt to the middle and rerun the task. If the answer changes, your context ordering is a bug, so pin critical instructions and interfaces at the start or end. Rerun the 20-task comparison every week, because Google can change the tiers at any time while Argon remains a test build.