The week of $2/$10 models showed pricing moving faster than access

Zvi Mowshowitz's weekly roundup connects several distinct events rather than forcing them into one grand story. Its most useful common thread is concrete: Gemini 4 Argon and GPT-6.1 Sol met at $2 per million input tokens and $10 per million output tokens, while actual access remained uneven.

Argon has a price and benchmarks, but ordinary developers still lack access

Google announced Gemini 4 Argon for long-running work in software engineering, knowledge work and cyber defence. The initial rollout is limited to selected defenders through the Fairwind Program, with wider availability promised later, starting with paid API customers and Google AI Ultra subscribers.

The introductory price is $2 per million input tokens and $10 per million output tokens, with Google quoting a 95% discount for cached input. After the introductory period, the rates are due to rise to $4 and $20. Mowshowitz is therefore right to separate an attractive price card from a product he can test himself.

A shared price turns model selection into a workflow question

According to the roundup, GPT-6.1 Sol carries the same $2/$10 price and is positioned near the more expensive GPT-6 Astra. Mowshowitz still prefers Claude Opus 5.5 for his own work. The piece is therefore best read as a map of different trade-offs, not an independent leaderboard with one winner.

When two frontier offers meet at the same token rate, comparing token prices in isolation stops being useful. Teams will care more about tokens consumed per completed task, tool use reliability, latency and whether the model is available in the region and plan they actually use.

Vendor charts still stand in for experience on unfamiliar repositories

Google reports 77.9% on DeepSWE v1.1, 51.3% on AutomationBench and 91.7% on LVBench for Argon. These are concrete results, but ordinary developers could not compare them with their own workloads at announcement time. The limited rollout creates an information asymmetry: the vendor has both the numbers and the model, while the audience mostly has the numbers.

The same caution applies to the roundup format itself. Mowshowitz provides a useful guide and first-hand experience with some models, not one controlled experiment comparing Argon, Sol and Opus under identical conditions.

Bills for completed work and open access will set the order

The next relevant signal arrives when Google opens Argon to paid API customers and Google AI Ultra subscribers. Independent evals on real codebases can then show how much of the claimed performance survives different tools, permissions and long-running tasks.

It is also worth watching whether the introductory rate drives real deployment before the move to $4/$20. Price per million tokens is an invitation. A supplier decision emerges from the cost of a completed and reviewed task.

Lilith's verdict

The window says $2/$10, but Argon's door still opens only for names on the guest list. The model market becomes comparable when customers receive the price, access and their own bill for completed work at the same time.