--- title: Google DeepMind head says Gemini 4 is in post-training,... description: Google DeepMind's Koray Kavukcuoglu says Gemini 4 has entered post-training and could ship well before year-end, possibly as early as October, with an early... canonical: https://daily.dev/posts/google-deepmind-head-says-gemini-4-is-in-post-training-could-ship-before-year-end-qstaqn78j twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Google DeepMind head says Gemini 4 is in post-training, could ship before year-end | daily.dev og:description: Google DeepMind's Koray Kavukcuoglu says Gemini 4 has entered post-training and could ship well before year-end, possibly as early as October, with an early... og:url: https://daily.dev/posts/google-deepmind-head-says-gemini-4-is-in-post-training-could-ship-before-year-end-qstaqn78j og:image: https://api.daily.dev/og/posts/QstAQn78J.png og:image:alt: Google DeepMind head says Gemini 4 is in post-training, could ship before year-end og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Google DeepMind head says Gemini 4 is in post-training, could ship before year-end **[Collections](https://daily.dev/sources/collections)** · 4 min read · 0 upvotes · 0 comments ## Summary Google DeepMind's Koray Kavukcuoglu says Gemini 4 has entered post-training and could ship well before year-end, possibly as early as October, with an early version released first and iterated on afterward. The comments come amid pressure on Google's release cadence: Gemini 3 Pro launched in November 2025 with only one update since, Gemini 3.5 Pro has been delayed past its expected May 2026 conference debut, and competitors OpenAI and Anthropic have shipped multiple frontier updates in the same window. Google has leaned on cheaper Gemini Flash models rather than pushing frontier releases. ## Content Google has announced Gemini 4 Argon, its first major flagship update since Gemini 3. The launch follows a rough stretch for the company: a delayed model, departures from Gemini's leadership, and rivals shipping faster. Argon is now on the table, and the early reaction is split between benchmark excitement and open skepticism. ## How it got here A few weeks ago, Google DeepMind chief Koray Kavukcuoglu told The Information that Gemini 4 was in post-training and that he wanted an early version out "much earlier" than the end of the year. That was his first public appearance since he took over day-to-day leadership of DeepMind after Demis Hassabis stepped back in August, following the departure of Gemini's two co-leads. The model was also being tested inside Google's Antigravity coding tool. The pressure was real. Gemini 3.5 Pro was expected at Google's May 2026 developer conference and never arrived; Bloomberg reports Google quietly dropped it. Gemini 3 Pro had seen a single update since November 2025, while OpenAI shipped three frontier updates (GPT 5.5 Pro, GPT 5.6 Astro, GPT 6 Astro) and Anthropic updated Claude several times. Google leaned on cheaper Gemini Flash models in the meantime. ## What Google is claiming Google pitches Argon for complex workflows in software engineering, enterprise knowledge work like legal and finance, and cyber defense. Against Anthropic's Opus 5.5 and Fable 5.1 and OpenAI's GPT-6 Astra, it claims the top score on 13 of 18 benchmarks. - **Coding is mixed.** Argon sets a new state of the art on DeepSWE v1.1 at 77.9%, but trails competitors on FrontierSWE v2 and Terminal-Bench 4.0. - **Knowledge work is where it pulls ahead.** It leads the Vals Index, AutomationBench (51.3%), GraphWalks and LVBench (91.7%). The widest gap is in legal work: 19.6% on Harvey's Legal Agent Benchmark, against 6.7% for Claude Fable 5.1. - **Security.** Argon ties for first on CWE-bench v1 at 68% for vulnerability remediation. Google trained it for autonomous vulnerability discovery, and Wiz used it to find a critical vulnerability in widely used healthcare software that earlier models missed, through Google's Scan for Good initiative. - **Output length.** The limit jumps from 64K to 1M tokens. @rohanpaul_ai points out that's nearly 8x the 128K cap on GPT-6 Astra, Opus 5.5 and Fable 5.1. Inside Google, Argon agents are doing real work. They freed over 300 TiB of data-center memory, with 500 TiB to 1 PiB of total savings estimated, and made a Rust port of the libgav1 video decoder 2.7x faster by replacing 32K lines of SIMD code. They're also being used for large C/C++-to-Rust migrations, including Fuchsia's Zircon kernel. ## Price Introductory pricing is $2 per million input tokens and $10 per million output tokens, with 95% off for cached inputs. @Hesamation compared cost per task: Argon at $1.99, Astra at $3.26 and Opus at $5.98, with Argon roughly matching Astra's intelligence. If that holds up, the cheaper-per-task story may matter more than the headline benchmarks. ## You can't have it yet Access starts narrow. Right now Argon goes to Google's own staff, vetted cyber defenders (government agencies and security companies) and trusted testers through the Fairwind Program. Defenders get it without cyber guardrails. Paid API customers and Google AI Ultra subscribers come next, then broader release to developers, enterprises and consumers. Google also described new safeguards against misuse, prompt injection and misalignment ahead of that wider rollout. ## Reactions The reception is far from unanimous. - @petergyang: "Google cooked on Gemini 4!" He adds that Google now needs to be more competitive on its coding harness (Antigravity) and personal agent (Spark). - @trikcode is unimpressed: Google "have always benchmaxxed and even in doing so it still loses out to Astra and Opus. It's a bit of a disaster." - Bloomberg reported internal skepticism among Google staff about real-world performance versus benchmark scores. - @kentcdodds is doing what people should do: handing Gemini his regular audit test to see whether it holds up. I think the skeptics have a point that launch benchmarks prove little on their own, and mixed coding results are a real caveat for a model pitched at software engineering. But the legal and knowledge-work numbers, the 1M-token output ceiling and the per-task cost are hard to wave away. The real test comes when people outside the Fairwind Program get their hands on it. ## Questions this post answers ### When is Gemini 4 expected to be released? Gemini 4 is in post-training and could ship well before the end of the year, possibly as early as October. Google DeepMind head Koray Kavukcuoglu said he hopes to release an early post-training version much earlier than year-end and then iterate from there, rather than waiting for a single polished launch. _Developers picking a model for AI coding workflows can follow Gemini 4's rollout on daily.dev as it happens._ ### Why has Gemini 3.5 Pro not been released yet? Gemini 3.5 Pro was originally expected to debut at Google's developer conference in May 2026 but still has not shipped. This delay, combined with Gemini 3 Pro receiving only one update since its November 2025 launch, has put Google under pressure as OpenAI released three frontier updates and Anthropic updated Claude multiple times in the same period. _Anyone comparing Gemini, GPT, and Claude release cadence for tooling decisions can track updates on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 2 comments across x (as of 2026-09-30). **TL;DR:** There's very little substantive discussion; the few replies mostly just note Gemini 4 Argon's large 1M-token output and one commenter praised the coverage itself rather than the model. **Sentiment:** 50% positive · 50% mixed · 0% skeptical **The case for** - One reply highlights that Argon's million-token output lets it write far more per response than GPT-6 Astra. **By community** - x (positive): The handful of replies are lightly favorable, focused on the large token output, with no real critical pushback. **Highlights** > @rohanpaul_ai Gemini 4 Argon holds one million tokens per response. It writes eight times more than GPT-6 Astra today. > — [TheSITherapist on x](https://x.com/TheSITherapist/status/2105403548057891036) **Source threads** - [x](https://x.com/rohanpaul_ai/status/2105401953186721855) · 0 points · 2 comments - [x](https://x.com/trikcode/status/2105404898254717432) · 0 points · 0 comments ## Similar posts on daily.dev - [Google reveals Gemini 4 Argon, but you're not allowed to use it](https://daily.dev/posts/google-reveals-gemini-4-argon-but-you-re-not-allowed-to-use-it-etzza9edq) · XDA Developers · 0 upvotes · 0 comments - [Google’s Gemini 3.5 Flash beats the frontier models](https://daily.dev/posts/google-s-gemini-3-5-flash-beats-the-frontier-models-gdirdibym) · The New Stack · 0 upvotes · 0 comments - [Google launches Gemini 3 with new coding app and record benchmark scores](https://daily.dev/posts/google-launches-gemini-3-with-new-coding-app-and-record-benchmark-scores-aqu91rxbh) · TechCrunch · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#openai](https://daily.dev/tags/openai), [#anthropic](https://daily.dev/tags/anthropic), [#google-gemini](https://daily.dev/tags/google-gemini), [#google-deepmind](https://daily.dev/tags/google-deepmind) [View this post on daily.dev](https://daily.dev/posts/google-deepmind-head-says-gemini-4-is-in-post-training-could-ship-before-year-end-qstaqn78j) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Google DeepMind head says Gemini 4 is in post-training, could ship before year-end","url":"https://daily.dev/posts/google-deepmind-head-says-gemini-4-is-in-post-training-could-ship-before-year-end-qstaqn78j","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/google-deepmind-head-says-gemini-4-is-in-post-training-could-ship-before-year-end-qstaqn78j"},"datePublished":"2026-09-25T12:57:01.194Z","dateModified":"2026-09-30T21:11:11.705Z","description":"Google DeepMind's Koray Kavukcuoglu says Gemini 4 has entered post-training and could ship well before year-end, possibly as early as October, with an early...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/76126a59ccaa462cd6c6d30023b7180b?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/76126a59ccaa462cd6c6d30023b7180b?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/google-deepmind-head-says-gemini-4-is-in-post-training-could-ship-before-year-end-qstaqn78j","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,openai,anthropic,google-gemini,google-deepmind","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Google DeepMind head says Gemini 4 is in post-training, could ship before year-end"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/google-deepmind-head-says-gemini-4-is-in-post-training-could-ship-before-year-end-qstaqn78j#faq","mainEntity":[{"@type":"Question","name":"When is Gemini 4 expected to be released?","acceptedAnswer":{"@type":"Answer","text":"Gemini 4 is in post-training and could ship well before the end of the year, possibly as early as October. Google DeepMind head Koray Kavukcuoglu said he hopes to release an early post-training version much earlier than year-end and then iterate from there, rather than waiting for a single polished launch. Developers picking a model for AI coding workflows can follow Gemini 4's rollout on daily.dev as it happens."}},{"@type":"Question","name":"Why has Gemini 3.5 Pro not been released yet?","acceptedAnswer":{"@type":"Answer","text":"Gemini 3.5 Pro was originally expected to debut at Google's developer conference in May 2026 but still has not shipped. This delay, combined with Gemini 3 Pro receiving only one update since its November 2025 launch, has put Google under pressure as OpenAI released three frontier updates and Anthropic updated Claude multiple times in the same period. Anyone comparing Gemini, GPT, and Claude release cadence for tooling decisions can track updates on daily.dev."}}]} ```