Google ships Gemini 4 Argon: benchmark fine print, expiring intro pricing and API gotchas

Google moved Gemini 4 Argon to general availability this week after unveiling the model on 30 September, according to a hands-on analysis on dev.to. The piece highlights two pieces of fine print before anyone migrates workloads: the benchmark scores driving most coverage were measured at a reasoning setting that is not the default, and the headline API price is an introductory rate that third-party reporting expects to double.

What Google shipped

Per dev.to, Argon is callable through the API at introductory rates of $2 per million input tokens and $10 per million output tokens. Google AI Pro subscribers get the model's reasoning capabilities in the Gemini app and in AI Mode in Search, while Antigravity, Google's agentic coding environment that edits files, runs tests and iterates between prompts, remains exclusive to Ultra subscribers.

The model carries a 1,048,576-token context window shared across input, internal reasoning and output, with replies capped at 65,536 output tokens by default. Background requests can chain 64K continuation steps to reach roughly a million output tokens in total. Argon accepts text, images, video, audio and PDFs as input but replies in text only; image and audio generation are not supported.

The benchmarks were measured in High mode

Google claims Argon beats Claude Opus 5.5 and GPT-6 Astra on 13 of 19 benchmarks in its own comparison. Figures reported by VentureBeat include 77.9% on DeepSWE v1.1 (versus 74.2% for Opus 5.5 and 74.1% for Astra), 51.3% on Zapier AutomationBench (versus 42.5% and 41.4%), and 57.4% on Terminal-Bench 4.0, where Opus 5.5 scored higher at 66.4%.

The catch, per dev.to, is that those tables use Argon's High reasoning setting while the model ships with Medium as the default. Independent testing by Artificial Analysis puts Argon at 52.56 on its Intelligence Index in High mode, within half a point of Astra but below Claude Sonnet 5.5 (56.00) and Claude Opus 5.5 (57.62). One figure favours Argon: a 15% hallucination rate, the lowest of any model above 45 on the index, against 51% for Astra at its maximum setting.

Default-effort results are murkier. Cursor reports 52.8% on CursorBench, matching Opus 5.5; Scale AI ranks Argon first on its Sixth Sense test at 55.4%; and Arena's text leaderboard places it first at 1525 from 4,942 votes, explicitly marked preliminary. Dev.to also warns that all three flagship labs now call their default effort medium without the settings being equivalent: Anthropic's Terminal-Bench number for Opus 5.5, for instance, was measured at xhigh effort.

Introductory pricing that expires

Third-party reporting cited by dev.to says the rate doubles to $4 and $20 per million tokens after the intro period, but outlets disagree on timing: one points to 1 December 2026, another says the rate runs through the end of 2026, and Google has not confirmed a date. The budgeting consequence is concrete. Independent cost analysis puts Argon at about $1.99 per task on the Artificial Analysis index at intro rates, roughly 60% of Astra's cost, but about $3.98, around 120% of Astra, once the intro period ends.

API behaviour that trips up migrations

Dev.to flags several gotchas for teams porting code. Temperature, top_p and top_k are accepted but ignored, leaving thinking_level (low, medium, high) as the only meaningful control. Explicit cachedContents calls fail; only implicit caching is supported, on prompts of at least 4,096 tokens. Independent benchmarks measure a median of 43 to 45 seconds to first token on 10,000-token prompts, and Google's migration checklist recommends HTTP timeouts of at least 15 minutes for hard prompts. Output beyond 65,536 tokens requires background mode.

Mixed reception from early testers

A Facebook developer group thread with 80 comments found most hands-on testers disappointed, judging outputs less accurate than Opus 5.5 on multi-page websites and motion-graphics work, though a minority called Argon comparable or faster. Bloomberg reported that some Google employees felt the model underwhelms on coding and front-end work relative to its benchmarks, a characterisation Google called inaccurate. Sam Witteveen of Red Dragon AI told VentureBeat the model is "a really good model, but not obviously the best one available." Dev.to notes Argon's edge is concentrated in agentic coding and business-automation benchmarks such as DeepSWE, AutomationBench and CyberBench.

Why it matters

Argon shows how much model positioning now rests on configuration choices buyers never see. A several-point index gap hides behind a single High label, and an attractive price quietly reverses into a premium one within months. Anyone evaluating frontier models should pin down the reasoning setting behind each published score, budget against post-introductory rates, and treat early first-place leaderboards as preliminary until independent default-effort testing accumulates.