Google has begun shipping Gemini 4 — officially named Gemini 4 Argon — starting with a September 2026 release that gave the model a 1-million-token output limit, according to a report published September 24, 2026 by The Information and relayed by Investing.com. The rollout followed the model entering “post-training” — the refinement phase that follows initial training — after Google DeepMind’s new senior leader said he wanted it out well before the calendar turns. The timing matters: Google’s newest publicly available model, Gemini 3.8 Flash, currently trails both OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 on independent intelligence benchmarks, and rivals have spent September shipping flagship after flagship while Google raced to get its next headline release out the door.
The disclosure comes from Koray Kavukcuoglu, in his first media interview since taking over as Google DeepMind’s top leader. He told The Information that Gemini 4 is in the early stages of post-training and that he expects the model to launch “much earlier” than the end of 2026, though he stopped short of naming a date. That single line has been enough to send the story racing across tech media, from Techmeme’s aggregation feed to Ground News’s coverage summary and a dedicated writeup from BolNews, all published within a 24-hour window ending September 25.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Google DeepMind Confirms Gemini 4 Has Entered Post-Training
Post-training is the stage where a large language model’s raw, pre-trained weights get shaped into something usable: reinforcement learning from human feedback, safety tuning, instruction-following polish, and the thousands of smaller adjustments that turn a base model into a shippable product. Reaching that stage is a meaningful milestone, not a launch. It means the core architecture and pretraining run for Gemini 4 are effectively locked, and Google’s researchers are now working on reliability rather than raw capability.
What the report did not say at the time was just as important. Kavukcuoglu did not confirm a release date, did not name variants such as a Gemini 4 Pro or Gemini 4 Flash, and did not disclose benchmark scores, context-window length, modality support, or pricing. As of September 18, 2026, a review of Google’s public model catalog by Eggstriker found no Gemini 4 entry anywhere in Google’s official listings — confirming that, six days before the Kavukcuoglu interview surfaced, Gemini 4 still did not formally exist as a product. That changed within weeks: by October 2026, Google DeepMind had published a dedicated Gemini 4 page identifying Argon as its newest frontier model, closing the gap between the September tease and an official listing. Any claim about Gemini 4’s benchmark scores or feature set that predates that October listing should still be treated as speculation dressed up as reporting.
Why the Timing Matters for Google
Google has spent 2026 in an odd position for a company that invented much of the transformer architecture underpinning the entire AI industry: it has been shipping upgrades to an aging model family while competitors roll out entirely new generations. Gemini 3.8 Flash, which reached general availability on September 2, 2026, is still built on the Gemini 3-series foundation first introduced earlier in the year. Meanwhile OpenAI moved to its sixth-generation Astra model in early September, and Anthropic shipped two new flagship releases — Claude Fable 5.1 and Claude Opus 5.5 — inside the same three-week window.
That compressed release cadence is the backdrop for why Kavukcuoglu’s comments landed as hard as they did. BolNews framed the story explicitly as Google “closing the AI gap” after falling behind on flagship releases, and the framing tracks with what independent benchmark trackers have been showing for weeks: Google’s most capable public model is no longer the most capable model on the market, and the company knows it.
Where Gemini 3.8 Flash Stands Today
Gemini 3.8 Flash remained Google’s highest-numbered publicly shipping model for most of the fall. According to the Tech Times AI release timeline, a third-party tracker of frontier model launches, Gemini 3.8 Flash scores 40.9 on its intelligence index, 76.3 on its coding benchmark, runs at roughly 267,305 tokens per second in throughput testing, and is priced at $1.50 per million tokens. Those numbers positioned it as a fast, inexpensive workhorse model — but not as a frontier-leading one, which is exactly the gap Gemini 4 Argon was built to close: CNBC reported in October 2026 that Argon tied OpenAI on a cybersecurity benchmark and led the field on software-engineering results, the first time a Gemini model has topped that particular category.
A separate comparison published by Valletta Software, citing the independent Artificial Analysis Intelligence Index, places Gemini 3.8 Flash at 41 points — trailing Claude Opus 5.5 (58) by 17 points and sitting well below GPT-6 Astra and Claude Fable 5.1, both listed at 53. Google, in other words, entered the fall with “no top-tier entry” on that particular leaderboard, in Valletta’s own phrasing. Gemini 4 Argon arrived at a price that matched OpenAI’s introductory rate — $2 per million input tokens and $10 per million output tokens when it launched in October 2026 — though Google has already disclosed those rates are temporary, with list pricing set to rise to $4 per million input tokens and $20 per million output tokens. That launch pricing, and the planned increase, is the clearest evidence of how aggressively Google needed to position Gemini 4 against the gap its predecessor left on the leaderboard.
The September 2026 Model Wave: A Timeline
September 2026 has been an unusually dense month for frontier AI releases. Three major vendors shipped new flagship or near-flagship models within the first three days of the month alone, according to tracking from Digital Applied’s release tracker and confirmed independently by Evertune’s AI model tracker. The table below lays out what shipped, when, and from whom — Gemini 4 is conspicuously the gap in the lineup.
| Model | Vendor | Release Date | Status vs. Gemini 4 |
|---|---|---|---|
| Claude Fable 5.1 | Anthropic | Sept 1, 2026 | Shipped 24 days before Gemini 4 report |
| Gemini 3.8 Flash | Sept 2, 2026 | Google’s current flagship (interim) | |
| GPT-6 Astra | OpenAI | Sept 3–4, 2026 | Shipped 21 days before Gemini 4 report |
| Claude Opus 5.5 | Anthropic | Sept 22, 2026 | Shipped 2 days before Gemini 4 report |
| Gemini 4 | Unconfirmed (targeted before year-end) | In post-training as of Sept 23–24, 2026 |
Google was technically first out of the gate among the three majors on September 2 with Gemini 3.8 Flash, but that model is an incremental update to an existing family, not a new generation. OpenAI’s GPT-6 Astra and Anthropic’s back-to-back Fable 5.1 and Opus 5.5 releases represent full generational jumps, which is why the comparison stings more than the calendar alone suggests.
How the Current Flagships Compare on Benchmarks and Price
Independent benchmark data gives the clearest read on where Google needs Gemini 4 to land. A coding-focused comparison from Morph put Claude Opus 5.5 at 66.4% on the Terminal-Bench 4.0 agentic coding benchmark, ahead of GPT-6 Astra at 57.9% and Claude Fable 5.1 at 55.8%. Gemini 3.8 Flash was not included among the top scorers in that particular ranking. Pricing tells a similar story of Google occupying the budget tier rather than the frontier tier.
| Model | Intelligence Index | Terminal-Bench 4.0 | Price (per 1M tokens, in/out) |
|---|---|---|---|
| Claude Opus 5.5 | 58 | 66.4% | $4 / $20 |
| GPT-6 Astra | 53 | 57.9% | Not confirmed in cited source |
| Claude Fable 5.1 | 53 | 55.8% | $10 / $50 (cached reads $0.25) |
| Gemini 3.8 Flash | 41 | Not ranked in cited comparison | $1.50 blended |
Read together, the two tables explain the urgency behind Kavukcuoglu’s comments. Google is not just behind on a single metric — it is behind on the industry’s most-cited general intelligence index by 17 points against the current leader, while also lacking a model in the premium price tier where Anthropic and OpenAI are now competing directly. A Gemini 4 that merely matches Gemini 3.8 Flash’s efficiency without closing the intelligence gap would not change Google’s competitive position; a Gemini 4 that closes the gap would reset the leaderboard entirely, given Google’s scale advantages in compute, distribution, and existing Workspace and Android integration.
What “Post-Training” Tells Us — and What It Doesn’t
It’s worth being precise about what a post-training disclosure actually signals, because it has been widely misread across social media in the hours since the report broke. Post-training does not mean a model is close to a consumer release in the way a “beta” or “preview” label would. It means the expensive, months-long pretraining run — the part that determines a model’s raw capability ceiling — is complete, and remaining work focuses on alignment, safety filtering, tool-use reliability, and instruction-following behavior. Historically, the gap between a model entering post-training and reaching general availability has varied widely across the industry, from a few weeks to several months, depending on how much reinforcement learning and red-teaming a lab deems necessary.
That ambiguity is precisely why Kavukcuoglu’s “much earlier than the end of 2026” framing, while notable, leaves enormous room for interpretation. It could mean October. It could mean December 20. Readers searching for a specific Gemini 4 release date will not find one in the current reporting, and any article claiming otherwise is not drawing from verified sourcing.
A New Era for Google DeepMind’s Leadership
The interview itself is notable beyond its Gemini 4 content. Kavukcuoglu recently became Google DeepMind’s top leader, and this marks his first public media appearance in that capacity. CNBC later reported that Google employees had already been testing Gemini 4 internally for weeks by September 2026, with some staff given unlimited access well before the public knew the model existed — and that same month, Google entered the model into trusted-cyber-defender testing through its internal Fairwind Program. Leadership transitions inside frontier AI labs tend to coincide with shifts in release philosophy — a new head choosing to speak candidly about internal timelines, as Kavukcuoglu did here, is itself a signal that Google wants to reset the public narrative around its AI roadmap after a stretch where Gemini updates felt incremental relative to OpenAI and Anthropic’s generational jumps.
It also lands in the same week Google pushed further into AI hardware. Reuters reported on September 24 that Alphabet’s Google had opened pre-orders for Googlebook, a new premium laptop category for Android users built around Gemini AI tools. Two major Gemini-adjacent announcements in the same 24-hour stretch is unlikely to be coincidental — it reads as a coordinated push to keep Gemini in headlines while the flagship model itself is still months from release at the earliest.
How Gemini 4 Would Need to Compare to GPT-6 Astra
OpenAI has already set the bar Google will be measured against. GPT-6 Astra reached general availability in early September and immediately became OpenAI’s reference point for professional and computer-use benchmarks, listed alongside GPT-5.6 Sol, Claude Fable 5.1, Claude Fable 5, Claude Opus 5, and Gemini 3.8 Flash on OpenAI’s own comparison materials. That Google’s current model is the explicit baseline against which a rival lab measures its own flagship is an uncomfortable detail for Google, and it underscores why the timing of a Gemini 4 launch matters more than usual: every month Gemini 4 stays in post-training is another month Gemini 3.8 Flash serves as the industry’s shorthand for “the model Google hasn’t caught up with yet.” Readers who want a detailed look at how the current generation stacks up can see our GPT-6 Astra vs. Gemini 3.8 vs. DeepSeek V4.1 comparison, which lays out the 83x throughput gap between the fastest and slowest of the three.
How Gemini 4 Would Need to Compare to Claude Opus 5.5
Anthropic’s Claude Opus 5.5, released September 22, presents an even sharper target. Anthropic says the model performs at roughly Claude Fable 5.1’s quality level on most tasks while costing 40% less to run than Opus 5 at default settings — a value proposition that directly threatens Google’s traditional pitch of Gemini as the budget-friendly frontier option. If Gemini 4 launches without matching Opus 5.5’s combination of benchmark performance and cost efficiency, Google risks losing its price-to-performance argument at the same time it’s already lagging on raw intelligence scores. Our recent coverage of how Claude Opus 5.5 matches Fable 5.1 at 60% less cost breaks down exactly how Anthropic engineered that positioning.
Historical Context: Gemini’s Release Cadence Since Launch
Google’s Gemini family has moved through numbered iterations at a faster clip than its competitors’ naming conventions might suggest, with successive Flash and Pro variants rolling out across 2025 and 2026 rather than waiting for full-number generational leaps. That iterative approach worked well when the competitive field was less crowded — each incremental Gemini release could credibly claim to be “the best available” for stretches of weeks or months. It has worked less well in 2026, as OpenAI and Anthropic began shipping full-generation jumps (GPT-6, Claude Opus 5, Claude Fable 5) on a compressed schedule that made Google’s point-release strategy look reactive rather than proactive. Gemini 4, if and when it ships, would represent Google’s first true generational jump since the original Gemini 3 launch, making it the most consequential Gemini release in over a year by that measure alone.
Market Impact: What a Gemini 4 Launch Would Mean for Alphabet
Alphabet has other levers besides raw model quality — Google Search’s AI Mode, Android’s Gemini integration, and Workspace all give the company distribution advantages that OpenAI and Anthropic can’t easily replicate. But distribution only carries a lagging product so far. Enterprise buyers evaluating frontier models for coding, agentic workflows, and reasoning tasks increasingly benchmark before committing spend, and a 17-point intelligence-index deficit is the kind of gap procurement teams notice. A Gemini 4 that closes that distance would give Google’s sales teams a credible story to tell enterprise customers currently defaulting to GPT-6 Astra or Claude for frontier workloads; continued delay compounds the risk that habits solidify around competitors’ tooling and API ecosystems in the meantime.
The pressure is not limited to enterprise accounts. Consumer-facing comparisons — the kind readers actively search for — increasingly frame Gemini as the value option rather than the capability leader, a positioning that’s fine for cost-sensitive use cases but weak for the flagship-model narrative Google needs to compete for mindshare against ChatGPT and Claude. For a full breakdown of how the three assistants currently compare on reasoning, see our ChatGPT vs. Claude vs. Gemini vs. Grok reasoning comparison.
Competitive Landscape at a Glance
Zooming out, the frontier AI race now has three labs shipping on overlapping but distinct cadences. OpenAI has leaned into rapid-fire sub-releases within the GPT-6 family (Astra, Sol, and others), while Anthropic has run a tight two-model September push with Fable 5.1 and Opus 5.5 covering both flagship-quality and cost-efficient tiers simultaneously. Google’s approach — frequent Flash-tier updates punctuated by rarer full-generation jumps — has left it exposed during exactly this kind of compressed rival release window. Our coverage of how GPT-6 Sol and Luna undercut Claude Opus 5.5 on price shows how aggressively OpenAI has been pricing its own mid-tier models to pressure Anthropic — a dynamic Google will also have to navigate once Gemini 4 ships, since undercutting on price alone won’t be enough if the intelligence gap remains.
What Developers Should Do in the Meantime
For developers currently building on Gemini’s API, the practical guidance has shifted now that Gemini 4 Argon is live: teams can move onto it directly, since Google launched Argon in October 2026 at $2 per million input tokens and $10 per million output tokens — pricing Google has already said will rise to $4 and $20 respectively — while Gemini 3.8 Flash access remains unchanged for anyone mid-project. Teams already working with Gemini 3.8 Flash can continue building with it; those evaluating which frontier model to standardize on for new projects should weigh Argon’s October 2026 benchmark results, where CNBC reported it led on software-engineering tasks and matched OpenAI on cybersecurity evaluations, against Anthropic and OpenAI’s newest releases. For a hands-on look at working with Google’s current API, see our Gemini 3.8 Flash API setup tutorial.
Predictions: What Happens Next
- A launch window before December is plausible but not guaranteed. Kavukcuoglu’s “much earlier” framing suggests internal targets of October or November, but post-training timelines routinely slip when safety testing surfaces issues, and Google has strong incentive to avoid a rushed release that underperforms on launch-day benchmarks.
- Expect a tiered rollout rather than a single model. Given Google’s pattern with the Gemini 3 family, a Gemini 4 release will likely arrive with Pro and Flash variants either simultaneously or in close succession, mirroring how OpenAI and Anthropic now ship multiple tiers of each generation.
- Pricing will likely undercut Claude Opus 5.5 and GPT-6 Astra. Google’s historical Gemini Flash pricing strategy has consistently positioned the company as the low-cost frontier option, and there’s no indication in current reporting that this strategy will change for Gemini 4.
- Benchmark claims will be aggressively marketed on day one. Every major lab has used launch-day benchmark charts to claim leadership on at least one metric; expect Google to highlight whichever evaluations favor Gemini 4 most strongly, regardless of how it performs on the broader Artificial Analysis Intelligence Index.
- The intelligence-index gap will be the story to watch. Whether Gemini 4 closes the current 17-point deficit against Claude Opus 5.5 will determine whether this cycle’s narrative shifts from “Google is catching up” to “Google is back at the frontier.”
The Bottom Line
Gemini 4 is real, it is in active development, and Google’s own DeepMind leadership is now speaking about it on the record for the first time. That’s meaningfully more than existed in the public conversation a week ago. But “in post-training” and “shipping soon” are not the same claim, and readers should treat any specific release date, benchmark score, or pricing figure circulating outside this reporting as unverified until Google confirms it directly. What’s certain is the competitive backdrop: Google is trying to close a real, measured gap against Claude Opus 5.5 and GPT-6 Astra, and the company has now publicly signaled urgency about doing so before 2026 ends.
Frequently Asked Questions
When will Gemini 4 be released?
Google has already begun shipping it. Google DeepMind’s Koray Kavukcuoglu told The Information on September 23–24, 2026 that he hoped to launch Gemini 4 “much earlier” than the end of 2026, and the first version — Gemini 4 Argon, with a 1-million-token output limit — followed that same month, with Google DeepMind formally listing it as its newest frontier model in October 2026.
Is Gemini 4 out yet?
Yes. After entering the post-training refinement phase in September 2026, Gemini 4 shipped that same month as Gemini 4 Argon, and by October 2026 it appeared in Google’s official model catalog as the company’s newest frontier model, with pricing starting at $2 per million input tokens and $10 per million output tokens.
What is Google’s current most advanced AI model?
Gemini 3.8 Flash, which reached general availability on September 2, 2026, is currently Google’s highest-numbered publicly available model.
Will there be a Gemini 4 Pro?
No confirmed Gemini 4 Pro variant has been announced. Current reporting only confirms that a Gemini 4 base model has entered post-training; specific product tiers have not been disclosed.
How does Gemini 3.8 Flash compare to GPT-6 Astra and Claude Opus 5.5?
On the Artificial Analysis Intelligence Index cited by Valletta Software, Gemini 3.8 Flash scores 41 points, versus 53 for GPT-6 Astra and Claude Fable 5.1, and 58 for Claude Opus 5.5 — a 17-point gap between Google’s current flagship and the market leader.
Who is Koray Kavukcuoglu?
Kavukcuoglu recently became Google DeepMind’s top leader. His comments to The Information about Gemini 4’s post-training status mark his first media interview since taking on that role.
Should developers wait for Gemini 4 before building on Google’s AI APIs?
Not necessarily. Gemini 3.8 Flash remains fully available and priced at $1.50 per million tokens blended. Developers with active projects can continue building on it; those planning long-term architecture decisions may want to wait for confirmed Gemini 4 specs before committing.
What does “post-training” mean for an AI model?
Post-training is the refinement stage after initial pretraining, covering reinforcement learning from human feedback, safety alignment, and instruction-following tuning. Reaching this stage means a model’s core capabilities are largely set, but the timeline to public release can still vary from weeks to months depending on testing outcomes.
Related Coverage
- xAI Launches Grok 4.7 at $2/$6, Trails Claude by 7 Points [2026]
- Google Suncatcher Puts 4 TPUs in Orbit Oct. 1 [2026]
- ChatGPT vs Claude vs Gemini vs Grok: 20-Point IQ Gap [2026]
- ChatGPT vs Claude vs Gemini Memory: 2x Capacity Gap [2026]
- Gemini 3.8 Flash API: Full Setup Tutorial in 10 Steps [2026]
