Gemini 4 Is in Post-Training, but Google Has Not Set a Release Date

The distinction matters to developers and enterprise IT teams weighing Google’s models against those available through Microsoft Foundry and other platforms. An executive’s timetable is worth watching; it is not yet a model that a team can price, test, or put into production.

What “post-training” tells us—and what it doesn’t​

Post-training is the work that follows a model’s initial training, refining how it behaves and responds. Reporting on Kavukcuoglu’s remarks says Google wants to release an early post-training output and keep iterating. It also reports internal testing involving Antigravity. Neither detail establishes who would get access to Gemini 4, whether it would appear in an app or API, or what safeguards and limitations an initial release might carry.

Google’s published Gemini API documentation offers a useful reality check. Its documented model list includes distinct Gemini 3 variants—among them Gemini 3.8 Flash, Live, and text-to-speech models—but no Gemini 4 entry. The API release notes document the September 22 release of Gemini 3.8 text-to-speech models, not a Gemini 4 launch. That does not rule out private testing; it means the interview should not be mistaken for a documented developer release.

It also cautions against treating “Gemini 4” as a single, inevitable launch across every Google product. Google documents separate model variants with different purposes and release dates. Until Google specifies an identifier and access channel, nobody can say which users or workloads would receive an early Gemini 4 output first.

Can Google catch up?​

The question makes for a good headline, but the available evidence cannot answer it. Kavukcuoglu’s remarks establish progress through development, not Gemini 4’s benchmark results, reliability, price, or performance against rival models. The submitted account attaches much more precise competitive claims and predictions to this story; those should not be used to judge a model that has not been publicly documented.

For a software team, “ahead” is also workload-dependent. A model that excels at one coding evaluation may still be the wrong choice for a production assistant if its tool calls are unreliable, its latency is unsuitable, or its operating cost is too high. That is an evaluation principle, not a prediction about Gemini 4.

What IT teams should do now​

Keep current model decisions tied to models that can actually be tested. For any planned comparison—including a future Gemini 4 evaluation—use the same representative tasks, scoring rules, data-handling requirements, latency targets, and cost assumptions across vendors. Record the exact model identifiers: Google’s existing API list shows why a family name alone is not precise enough for a reproducible test.

When Google publishes Gemini 4 details, the questions to check are straightforward:

• Access: Which app, API, or cloud platform offers it, and to whom?
• Status: Is it a preview or a generally available model?
• Fit: Does it improve results on your coding, support, or document workflows?
• Operations: What are its documented prices, limits, safeguards, and versioning terms?

For now, the news is meaningful but bounded: Google says Gemini 4 is in post-training and wants an early output sooner rather than later. The release date—and the evidence needed to decide whether it changes an enterprise AI purchase—are still to come.

Update: Gemini 4 Argon reportedly enters restricted access (September 30, 2026)​

Google has revealed Gemini 4 Argon and is providing access to trusted defenders through its Fairwind Program, according to XDA Developers’ September 30 report, which attributes the announcement to News from Google. This moves the story beyond the earlier post-training comments: a named model reportedly has a restricted access route, although no public release date has been announced.

According to XDA Developers, Google employees are already using Argon for everyday tasks, while Google is strengthening safeguards against cyber and chemical, biological, radiological, and nuclear misuse. That reported internal use and limited distribution should not be confused with availability through the consumer Gemini app or a generally accessible API.

XDA Developers also reports that Google intends to expand access to developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers. For Windows developers and IT teams, that identifies potential initial access channels—not a deployment timetable. The report does not establish public pricing, model identifiers, availability dates, or production terms, so enterprise evaluations still depend on further release documentation.

Update: Gemini 4 Argon pricing and output limits reported (October 1, 2026)​

According to Neowin’s October 1 report, Gemini 4 Argon will launch with introductory pricing of $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached inputs. The report says those input and output rates will double to $4 and $20 after the introductory period, but does not specify how long that period will last. These reported prices add a cost baseline that was missing from the earlier restricted-access announcement.

Neowin also reports that Google has increased Argon’s maximum output from 64,000 tokens to 1 million tokens, intended to support longer reasoning and complex tasks without splitting them across separate runs. For developers planning lengthy code migrations or document workflows, that is a potentially important capability, although the reported limit alone does not establish reliability, latency, or the cost of completing a particular task.

According to Neowin, Google reports scores of 77.9% on DeepSWE v1.1 for long-horizon software engineering and 68% on CWE-bench v1 for vulnerability remediation. Those vendor-reported results offer initial evaluation targets, not independent evidence of production performance. Broader availability remains pending, so Windows development teams and enterprise buyers still need access terms and hands-on testing before treating the announced pricing and capabilities as a deployment option.

Update: Independent Argon evaluations reportedly show mixed workload performance (October 3, 2026)​

Gemini 4 Argon now has reported independent evaluation results beyond the vendor benchmarks covered earlier. According to The Elec’s October 2 report, Artificial Analysis awarded Argon 53 on its Intelligence Index, tying GPT-6 Astra and Claude Fable 5.1 but trailing Claude Opus 5.5 and Sonnet 5.5. The outlet also reports that Argon placed third on Artificial Analysis’ Coding Agent Index, ahead of GPT-6 Astra. These provide additional comparison points, rather than confirmation of Google’s own performance claims.

The reported results vary substantially by workload. According to The Elec, Argon ranked first on AutomationBench, which tests multistep enterprise tasks using connected business tools, but placed 14th on GDPval-AA with 56%. For Windows development teams and enterprise evaluators, that contrast argues against treating strong coding results as evidence of equally strong performance across office workflows.

The Elec also reports Artificial Analysis measured Argon’s cost per task at $1.99 using promotional pricing and characterized its output as verbose. That figure is an evaluation-specific cost, not a production budget: teams would still need to measure token consumption and completion quality on their own tasks.