FULL STORY

Gemini 4: From Secret Arena Tests to Official Launch

After anonymous checkpoints surfaced on LMArena and leaked benchmarks circulated, Google officially unveiled Gemini 4 Argon on October 1. The model earned strong reviews and topped multiple leaderboards, though some questioned the credibility of its benchmark results.

2026-09-26 ~ 2026-10-05 · 19 episodes · 204 posts

Episode 1 · Suspected Gemini 4 Pro Spotted in Anonymous Arena Testing (2026-09-26, 8 posts)

Between 09-26 and 09-27, multiple users reported spotting an unreleased checkpoint suspected to be Google Gemini 4 Pro on LMArena, possibly tested anonymously under the placeholder name "gemini-3.8-flash." Leaks claim it outperforms Claude Opus 5.5 and GPT-6 Astra at roughly 1/5 the cost with 2M context. None of this is officially confirmed, but it has triggered intensive community blind testing.

Confirmed

  • On 09-26, @cedricchee found Google had deployed a new Gemini 4 Pro checkpoint on Arena competing against GPT-6 Astra and Fable, with Gemini 4 Pro currently leading; @markk posted "don't underestimate Google DeepMind," previewing a strong comeback (unverified).
  • The same day, WorldofAI summarized the week's leaks: the Barium-B checkpoint reportedly beat Claude Opus 5.5 and GPT-6 Astra on new benchmarks, with release timing predictions; the roundup also mentioned Kimi K4 and a mystery model.
  • On 09-27, @CodeByPoonam relayed @TimJayas's testing: a model routed under "gemini-3.8-flash" on LMArena performed like the rumored Gemini 4 Pro — a mechanical hummingbird prompt produced high-quality output close to Fable, clearly better than current flagships.
  • @CodeByPoonam also compiled 10 community generation examples; @Wanderspor relayed a Reddit roundup claiming leaked specs of 1/5 competitor cost and 2M context, which would upend frontier-model economics if released in October.

Unconfirmed

  • Whether the anonymous model is truly Gemini 4 Pro: @aiforsuccess's comparison screenshots (relayed by @CodeByPoonam) showed it "completely crushing" Opus 5.5, but the poster cautioned it cannot be 100% confirmed; @markk's claims are also unverified.
  • A same-prompt comparison against Claude Opus 5.5 High concluded Gemini 4 "has not yet reached" that level, contradicting the "crushing" narrative.
  • No official data exists on rankings, pricing, or release timing.

Why it matters

  • If true, Google would field a highly competitive next-gen flagship against OpenAI (GPT-6 Astra) and Anthropic (Opus 5.5), potentially reshaping frontier-model economics with drastically lower costs — driving intense community attention, with some calling it potentially the biggest comeback in AI history.

Episode 2 · Google's Gemini 4 Enters Final Training Push, May Launch by Year-End (2026-09-26, 2 posts)

Google is preparing to launch Gemini 4 to close the gap with Anthropic and OpenAI. DeepMind chief Koray Kavukcuoglu confirmed the model has entered early post-training, with a possible launch before year-end.

Episode 3 · Leaked Gemini 4 Pro Benchmark Sparks Buzz Ahead of OpenAI DevDay (2026-09-28, 4 posts)

Unverified screenshots and benchmarks of Google's Gemini 4 Pro leaked on Reddit just as OpenAI's DevDay kicked off, claiming it beats GPT-6 Astra and Claude Opus 5.5; skeptics warn the scores are unverified and that benchmarks don't equal intelligence.

Episode 4 · Gemini 4.0 Pro and Multiple New Models Tipped for Release Within Weeks (2026-10-01, 2 posts)

Developer Bindu Reddy and Google DeepMind's Logan Kilpatrick hint that multiple notable new models, including Gemini 4.0 Pro, Fable 5.5, and Astra++, will arrive within the next 4-6 weeks.

Episode 5 · Google Unveils Gemini 4 Argon with 1M-Token Output, Defense-First Rollout (2026-10-01, 118 posts)

On October 1, Google CEO Sundar Pichai pre-announced the next-generation frontier model Gemini 4 Argon amid widespread rumors, followed by formal announcements from the Google DeepMind account, CEO Demis Hassabis, executive Logan Kilpatrick, and the official blog "Introducing Gemini 4 Argon." The model targets complex workflows, software engineering, enterprise knowledge work, and cyber defense; it was delivered to the US government on launch day and opened to trusted cyber defenders via the Fairwind program, with broader availability to follow.

Confirmed

  • Pichai said Argon reaches frontier levels on complex workflows, cyber defense, and software engineering, with heavy internal Google usage from coding to quantum computing; the model carries frontier safety guardrails and was delivered to the US government today.
  • Official specs: the output token limit jumps from 64K to an industry-leading 1 million, enabling long-chain, multi-step problems in a single run; Sam Witteveen called this an order-of-magnitude leap over existing 64K/128K limits, with longer thinking time.
  • Pricing: introductory rates of $2 per million input tokens and $10 per million output tokens, per Logan Kilpatrick, opening to cyber defenders first before broader rollout.
  • The official blog confirms the model was built jointly by Google DeepMind and Google Research; official claims include significant benchmark gains and a new SOTA on real-world long-horizon software engineering tasks.
  • Third-party data: LMArena reports per-task costs 33% below GPT-6.1 Sol and up to 70% cheaper than other top models.

Unconfirmed

  • Yuchen Jin claims Argon beats GPT-6 Astra and Claude Opus 5.5 across benchmarks while questioning possible benchmaxxing; Bindu Reddy is skeptical of Google's benchmark record but sees the 5x-cheaper-than-Astra pricing as the highlight. Detailed official benchmarks have not been released.
  • Polymarket gives Google a 93% chance of holding the top model by year-end—clearly meme-flavored, not official.

Why it matters

  • Argon deepens the binding of frontier model releases to national-security evaluation, with safety guardrails and defender-first access.
  • The 1M-token output limit plus aggressive $2/$10 pricing dramatically lowers the barrier for long-horizon autonomous and agentic workflows.
  • It continues Google's rapid cadence—three frontier models in a month by one Reddit user's count—landing the same day as OpenAI's DevDay, intensifying the frontier race.

Timeline

  • October 1: Pichai's early announcement; DeepMind, Hassabis, Kilpatrick, and Google AI publish specs and pricing; official blog goes live; model delivered to the US government and opened via Fairwind; same-day third-party cost comparisons from LMArena.

98 more related posts →

Episode 6 · Gemini 4 Leaks Flood Out as Bloomberg Reports Internal Struggles (2026-10-01, 35 posts)

From October 1-2, a wave of leaks around Google's rumored Gemini 4 / Gemini 4 Argon pointed to strong benchmark results, but Bloomberg reported the same days that Google insiders found real-world performance falling short of scores, especially in coding, and that the model is very large with high running costs that could pressure margins. The community erupted into a debate over "benchmaxxing vs. real capability." None of the information is official.

Confirmed

  • Bloomberg reported: Alphabet is preparing to launch Gemini 4; the model performs well on industry benchmarks, but multiple insiders with direct access said employee real-world usage fell short of expectations, particularly struggling on some coding tasks. Internally, opinions are split on whether it has caught up with OpenAI and Anthropic. Sources spoke anonymously.
  • Bloomberg further disclosed (relayed by firstadopter): Gemini 4 is reportedly a very large model with costly inference that could pressure Google's margins; experts also noted Google's tendency toward benchmark optimization.
  • Google employee Kuba tweeted proudly about working on Gemini; developer sunjiao123sun quoted him, saying that without Kuba, Gemini 4 Argon would not have scored 91.9 on VibeCodeBench.
  • LuminaBench posted that Gemini 4 Argon scored 53 on the AA Intelligence Index, tied with GPT-6 Astra, sitting near the Pareto frontier on the cost curve.

Unconfirmed

  • A Reddit leak claimed Gemini 4 Argon topped 12 of 18 benchmarks against Fable 5.1, Opus 5.5 and GPT-6 Astra, including 19.6% on autonomous legal work; authenticity unverified.
  • Blogger haider claimed Gemini 4 reaches SOTA on knowledge tasks with coding roughly at Astra/Fable 5.1 level — an unverified leak.
  • teortaxesTex speculated from pricing that Argon may be a mid-tier model targeting Claude Sonnet (internal Sol tier).
  • DrSingularity claimed Gemini 4 Argon "has launched," with no official or third-party confirmation.

Why it matters

  • The contrast between Bloomberg's insider account and the flood of benchmark rumors made "do benchmarks reflect real capability" the core debate. Users like EstablishmentFun3205 questioned whether this is real progress or benchmaxxing; Angaisb and ishuagra02 cited past gaps between Google's official scores and actual performance for Gemini 3 / 3.1 Pro; koltregaskes and Neurogence urged waiting for public availability before judging; ChrisGPT called it a step in the right direction despite struggles.
  • Cost is another key dimension: Bloomberg's report of a very large model pressuring margins versus LuminaBench's near-Pareto-frontier claim together point to the tradeoff between flagship capability and inference cost. If the benchmark data holds, Gemini 4 Argon could reshape the frontier-model landscape, and the tiering rumors suggest Google is building a multi-tier product line for the mid-to-high-end market.

15 more related posts →

Episode 7 · Gemini 4 Argon Shines in Long-Horizon Coding Benchmarks (2026-10-01, 2 posts)

Gemini 4 Argon delivers strong results on the AA Coding Agent Index and scores 55.0% on the FrontierSWE v2 long-horizon coding benchmark, far ahead of Gemini 3.x's 20%.

Episode 8 · Viral "Gemini 4 Argon" blog page turns out to be a hoax (2026-10-01, 6 posts)

On October 1, a blog.google URL circulating as a "Gemini 4 Argon" page briefly fueled speculation that Google was about to launch Gemini 4. The claim does not hold: Gemini 4 does not exist, and the page appears to be a prank or meme. Some posters further dragged "GPT-6 Astra" into the comparison, amplifying rumors despite no official information on either next-generation model.

Confirmed

  • The link does lead to a page under the blog.google domain, but as koltregaskes noted when sharing it, the page body is only the Google blog's navigation directory, with no specs, release date, or capability details about Gemini 4.
  • A Reddit post shared the link with the "Wake up babe, it's Gemini 4" meme format; MrTimeHacker1 pointed out Gemini 4 does not exist and the page appears to be a hoax.

Unconfirmed

  • AIAndrew claimed that while studying manifolds and thermodynamics he found a possible "4 Argon" easter egg hidden in the Gemini logo — an unverified personal speculation with no supporting evidence.
  • svenai posted that "Gemini 4 Argon" would compete with "GPT-6 Astra"; neither model has been officially announced, and the claim is pure unverified X-circle rumor.

Why it matters

  • The episode shows how sensitive the AI community is to next-gen model news: an empty page went viral and spawned derivative claims such as a GPT-6 rivalry.
  • TheZvi joked about peeking under the surface for "something beyond human comprehension," and architsharma97 mocked the name for containing "two off-by-one errors in one model name," showing the event has become an in-group meme with no value for assessing Gemini 4's actual progress.

Episode 9 · Google's Gemini 4 Argon agents free over 300 TiB of datacenter memory (2026-10-01, 3 posts)

Google engineers say a team of Gemini 4 Argon agents autonomously analyzed fleet-wide telemetry and applied memory optimizations, freeing over 300 TiB of memory across Google datacenters. Engineers call it one of Argon's biggest wins in infrastructure optimization.

Episode 10 · Reported Gemini 4 Argon Beats Quantum Algorithm Baseline by 40% in Minutes (2026-10-01, 2 posts)

Reports say Google's Gemini 4 Argon optimized a bottleneck quantum subroutine, beating a published baseline by 40% in just minutes. The claim, relayed by Linus Ekenstam and others, remains unverified.

Episode 11 · Google Engineer Praises Gemini 4 Argon's Troubleshooting Skills (2026-10-01, 3 posts)

Google engineer rakyll, who got early access to the upcoming Gemini 4 Argon, praised its impressive troubleshooting in complex environments and said she uses it internally for development, security scanning and debugging.

Episode 12 · Polymarket Bets Surge on Google as Gemini 4 Rumors Swirl (2026-10-01, 2 posts)

Polymarket's 'top AI model by year-end' market shows Anthropic at 48%, Google at 35% and OpenAI at 8%. Odds on Google reportedly doubled after rumors of a Gemini 4 release with million-token outputs.

Episode 13 · Gemini 4 Expectations Split as Google Restart Pretraining (2026-10-01, 2 posts)

Opinions diverge on Google's Gemini 4 (codenamed Argon): some argue its real-world performance matters less than Google restarting full pretraining, while skeptics predict the public release will be overtaken by rivals within a month.

Episode 14 · Blogger Defends Google's TPU Edge as Gemini 4 Hype Builds (2026-10-02, 2 posts)

Ahead of Gemini 4's release, blogger haider pushed back against critics, arguing that Google's self-developed TPU compute infrastructure is an underrated moat. He added that even if Gemini 4 underperforms its benchmarks, having another model competing with OpenAI and Anthropic is itself a win.

Episode 15 · Rumors Claim Google Has Far Stronger Internal Models (2026-10-02, 2 posts)

Unverified posts on Reddit and X claim Google internally holds models far stronger than Gemini 4 Argon, allegedly from a DeepMind engineer. No official confirmation or details exist.

Episode 16 · Reported Google Gemini 4 Argon supports 1M-token outputs (2026-10-02, 2 posts)

Unconfirmed reports claim Google is rolling out Gemini 4 Argon, a deep-reasoning model for complex long-horizon workflows, capable of outputting up to 1 million tokens per run — far beyond the 64K limit of most frontier models.

Episode 17 · Gemini 4 Praised as Astra-Class at One-Fifth the Price (2026-10-03, 3 posts)

Google's Gemini 4 draws wide praise: investors note it delivers about 95% of Astra-class performance at one-fifth the price, while commentators say its strong benchmarks confirm a three-way race among frontier models and credit Google's willingness to skip incremental updates.

Episode 18 · Gemini 4 Argon Tops Multiple Leaderboards, Sparking Debate (2026-10-04, 3 posts)

Google's Gemini 4 Argon reportedly tops multiple leaderboards including Agent Arena at just 20% of Fable 5 High's cost, but critics call LMArena a slop benchmark and question whether Google has truly returned.

Episode 19 · Gemini 4 Argon Surfaces in Docs and Cloud Quotas Ahead of Launch (2026-10-04, 3 posts)

Google's next frontier model Gemini 4 Argon has appeared in Gemini API docs and as a 'gemini-eap' entry in Google Cloud quotas, with a public launch expected soon as Gemini 3.6/3.7 Flash face deprecation.