Gemini 4
Where it stands: Google now has two Gemini 4 models: Argon, which shipped September 30 to trusted cyber defenders and then opened worldwide to developers and subscribers in mid-October, and Carbon, announced October 28 for cybersecurity and long-horizon work. Carbon claims the top Text Arena score of 1528 and a second-place Agent Arena showing, but remains limited to a subset of Fairwind cyber defenders, with Perplexity saying it will reach Pro and Max users once capacity allows. Watch for Carbon's wider access and outside checks of its numbers.
Latest
- 2026-10-30(~4:19 PM ET, Arav Srinivas — x.com/AravSrinivas): Perplexity's early evals and first distribution commitment for Carbon — "Strongest model that we've seen so far for long-form research and answer quality, with a new SoTA on the Wide-And-Deep-Research (WANDR) benchmark," coming soon to all Perplexity Pro and Max users inside Computer "once we have sufficient token capacity." A named distributor commit and a fresh benchmark claim for the Carbon rollout; vendor claim, no independent check. Watch: the Computer rollout landing, WANDR methodology surfacing.
AI image - 2026-10-30 (~1:52 PM ET, Hassabis RT of Mercor — x.com/demishassabis): Gemini 4 Carbon debuts at #1 on Mercor's APEX-Agents — 84.2% Pass@1, 90.4% mean, the board's only model above a 90% mean, ahead of Gemini 4 Argon (82.4/87.4) and Claude Sonnet 5.5 (75.5/83.7). Vertical splits: management consulting 91.9% Pass@1 (#1), investment banking 86.9% (#1), corporate law 73.8% (#2 — Argon leads at 76.7%). APEX-Agents scores multi-step tasks from investment banking, management consulting and corporate law. Another board debut for Carbon following Arena's, again amplified by its own team rather than independently written up. Watch unchanged: independent checks, access widening beyond Fairwind, pricing/API availability.
- 2026-10-28 (~1:00 PM ET, DeepMind blog + announcement posts — deepmind.google/blog): Gemini 4 Carbon announced — the second Gemini 4-generation frontier model, two weeks after Argon, positioned around state-of-the-art cybersecurity and long-horizon capabilities. Claims: #1 on Artificial Analysis' v5.0 Intelligence Index at 59 with Opus 5.5 and GPT-6 Astra behind (theverge.com/tech); 29% fewer output tokens per task than Opus 5.5 and a >10x claim over any publicly available model on high-quality security-patch generation, with GPT-6 Astra about 20% cheaper at $6.25 (wccftech — wccftech.com). Access gated to a subset of trusted cyber defenders in the Fairwind program ("expand access as soon as possible") — announced, not broadly shipped, echoing the Argon ship pattern. Watch: access widening beyond Fairwind, independent checks of the 59 on a closed-benchmark board, pricing/API availability.
Images in this thread
Claims to watch · 1 (0 pending, 0 slipped, 1 hit, 0 missed)
- Google DeepMind, Sep 24, 2026: Gemini 4 launch "much earlier" than end of year review by hit Source
October 2026 · 14 updates
- 2026-10-28 the closed v5.0 board's leader gets a name — Google's Gemini 4 Carbon, claimed at 59 on Artificial Analysis' v5.0 index with Opus 5.5 and GPT-6 Astra behind (theverge.com/tech). Against the Tuesday-night v5 relays (best open-weights entry Kimi K3.1 Pro at 46, weights unreleased), the top-to-open gap now reads thirteen points if the vendor claim stands. Watch: whether any released-weights model closes ground in a later v5 revision; Kimi K3.1 Pro's actual weight release.
- 2026-10-28 (Hassabis RT of Arena.ai — x.com/demishassabis): Gemini 4 Carbon debuts on Arena's boards — #1 Text Arena (1528, ahead of Gemini 4 Argon at 1520), #2 Agent Arena (+16.2% net improvement across ~21.9K real-world agentic sessions, ahead of GPT-6 Astra and Fable-5.1) at a $1.80 median cost per task, 72% below the board's #1 Claude Opus 5.5 (Max), and #4 Code Arena: WebDev (1768 — 2 pts behind Claude Sonnet 5.5 xHigh, 89 ahead of Argon). Signal standouts: #1 steerability (+18.6%), #3 confirmed success (+18.7%). First third-party board placements for Carbon one day after the closed-benchmark announcement — the independent-check watch gets its first datapoint, though the numbers arrive via Carbon's own team's amplification of Arena's post, not an independent writeup. Hassabis restates positioning: "state-of-the-art across cybersecurity and long-horizon capabilities." Watch unchanged: access widening beyond Fairwind, broad replication of the #1s, pricing/API availability.
- 2026-10-21 (~12:45 AM ET, Hassabis RT of Logan Kilpatrick — x.com/demishassabis): the post-ship thread's first usage-scale figure — "~25 trillion tokens" processed "in just the first 72 hours since Gemini 4 Argon launched," demand "incredible and still under appreciated." Vendor-adjacent, no third-party measurement in the reporting. Timing wrinkle noted in the record: the "first 72 hours" framing does not obviously reconcile with the thread's September 30 Argon ship date given the RT's October 21 timestamp — original-post date or alternate launch beat unresolved. Watch: recurring token-volume disclosures as a comparable series, independent corroboration.
- 2026-10-21 DeepMind's Discovery team reports a Gemini 4 Argon-powered pipeline (led by student researcher Marco Santos) that generated 20 new potential Busy Beaver "Cryptids" for BB(6) and verified 33 machines, with the Lean 4 proofs open-sourced and an arXiv paper linked (x.com/GoogleDeepMind; arxiv.org/abs). A model-application readout inside the ship-window thread; watch whether the pipeline produces published results beyond the arXiv preprint.
- 2026-10-15 (~12:22 PM ET — x.com/demishassabis + ~6:52 PM ET rollout — x.com/demishassabis): two datapoints on the post-ship thread. (1) ARC Prize's Verified numbers for Gemini 4 Argon — ARC-AGI-3: 97.5% at $4.8K (provider adapter harness) vs 49.8% at $5.8K (standard harness); ARC-AGI-2: 91.3% at $0.32/task; ARC-AGI-1: 98.5% at $0.16/task; Google DeepMind "the second frontier AI lab to score above 95% on ARC-AGI-3." (2) Google Antigravity rolls Argon out to paid Google AI Pro subscribers — "scaled capacity deliberately," with a separate weekly-only Argon quota on top of the existing 3.8 Flash Gemini bucket. Watch: the ~48-point ARC-AGI-3 harness-gap explanation, the weekly quota's actual limits, whether the standard-harness 49.8% becomes the cited number anywhere.
- 2026-10-15 (~10:49 AM ET — x.com/GoogleAI): the Argon demo wave gains an agentic-world layer — GoogleAI: "See a fully-explorable 3D world build itself in minutes ⛩️," with DeepMind's @GeokenAI explaining how Gemini 4 Argon acts as a virtual builder in Antigravity to autonomously create a multi-tiered pagoda sanctuary. Development of the creative-demo thread; vendor demo framing with no artifact or reproducibility detail in the reporting. Watch: hands-on availability, Antigravity integration specifics.
- 2026-10-14 (~7:00 AM ET, MSN Austrian-Australian syndication): the access-widening watch's next datapoint — Gemini 4 Argon already live for Australian Google One subscribers (msn.com/en-au), after the Sep 30 Fairwind-gate ship and the Oct 13 developer/Pro widening. Only Australia named in this item, and the writeup is syndicated rather than Google primary. Development of the ledgered subscriber-widening rung. Watch: subscriber access in other markets, the free-plan trajectory (Oct 9 Flash-Lite tightening sits in tension with the widening), independent replication of the board #1s.
- 2026-10-14 the thread's independent-treatment watch gets its first datapoint — Scale Labs' third-party Humanity's Sixth Sense (HSS) benchmark puts Argon on top at 55.4%, "edging out GPT-6-Astra at 53.6%" (Hassabis RT, ~3:37 PM ET — x.com/demishassabis), described as testing reasoning beyond what's visible (hidden objects, next-state prediction, human behavior), with a category-level breakdown attached but not read in the reporting. Same-day race color rather than benchmarking: Hassabis RT of an Argon vs Muse Wonder game-development side-by-side — "Same prompt, two completely different games" (~7:25 PM ET October 13 — x.com/demishassabis) — Google's own leader amplifying a matchup against Meta's Muse agent.
- 2026-10-14 (~1:25 PM ET, Hassabis RT of Google DeepMind — x.com/demishassabis): the domain frame widens to science — "Gemini 4 Argon is already unlocking new scientific breakthroughs in quantum computing, nuclear fusion, protein design and beyond," with a 🧵 attached to the RT. First science-breakthroughs claim in this topic's capture; prior layers were benchmarks, robotics (Video-to-Sim-to-Real on Apollo 2), creative demos, the HSS third-party ranking, and the Oct 13 developer/Pro + global access widening. Vendor-adjacent, no artifacts in the reporting, attached thread unread. Watch: whether the named-field claims get paper-level evidence, and whether they ride the post-widening access the Oct 13 datapoints opened.
- 2026-10-13 (~1:00 PM ET pair): Gemini 4 Argon opens to developers and Pro users (Il Sole 24 Ore) — the thread's access-widening watch answered roughly two weeks post-ship — plus a same-day DeepMind post extending the domain frame to robotics. The thread's next watch (independent replication of the #1 board claims) is now testable from here. en.ilsole24ore.com/art deepmind.google/blog
- 2026-10-13 (second window, evening record pass): global availability confirmed on Google's own channels — Pichai (via Musk "Congrats!", ~2:39 PM ET): "available globally today for developers and enterprises in the paid tier of the Gemini API and the Gemini platform, for Ultra subscribers in Google Antigravity, and Ultra and Pro subscribers in the Gemini app and Search" (x.com/elonmusk) — the Sep 30 ledger's "developer/enterprise/consumer access undated" item converts to a date, same day as the Il Sole/robotics pair already ledgered. Same window: DeepMind offers 10,000 scientists the highest Antigravity usage tier free for 6 months (Hassabis RT of Jun Cheng, ~12:22 PM ET — x.com/demishassabis); GoogleAI demos Argon generating humanoid-robot training simulations from human-task video (cornhole demo, ~3:03 PM ET — x.com/GoogleAI). Watch unchanged: independent replication of the board #1s — now testable with open access.
- 2026-10-13 (~12:20 PM and ~12:48 PM ET, Hassabis RTs): the first creative-demo wave for the publicly available Gemini 4 Argon — fofr's pelican-on-a-bike via fourier epicycles with time-synced Lyria 3.5 music prompted by Argon (x.com/demishassabis), and an Argon-in-Antigravity autonomously sculpted photorealistic Pagoda temple complex in Blender via "multi-agent visual critique loops" (x.com/demishassabis). Development: this rides on the same-day developer/Pro access widening; the Argon-to-Lyria pairing is the first in the reporting. The independent-replication watch (the thread's #1 board claims once access opened) remains unanswered — the demos are the evangelism layer, not benchmarks.
- 2026-10-03 (market layer): Polymarket prices Google at 65% (Anthropic 36%) as October's top AI model (ainvest) — a lead the piece itself calls two-day-old, built on Argon results "the public can't test" (Fairwind gate). The thread's watch — independent replication once access opens — now has a price attached to numbers nobody outside Fairwind can check.
- 2026-10-01 (corroboration day, access still gated): no widening — public access still runs through the Fairwind trusted-tester gate while the corroborating layer piles up. (1) Market: Alphabet +~2% premarket on the launch (Yahoo Finance). (2) Eval corroboration via X, all Hassabis RTs (vendor-adjacent — flag): Blueprint-Bench 2 #1 (Andon Labs: physical 3D understanding — floorplans from photos); Artificial Analysis hallucination 15% vs Grok 4.7 29%, GPT-6 Astra 45%, Opus 5.5 59%. (3) Product: Guided Vision in Gemini Live on Android — real-time camera descriptions ("read the fine print," ~3:47 PM ET). (4) Ops claim: Argon agents freed 300 TiB of memory across Google's data centers — deployment- scale automation datapoint, unverified. Watch unchanged: access widening, independent replication of the #1s once it happens.
September 2026 · 2 updates
- 2026-09-30 the ship lands — Gemini 4 Argon (blog ~4:01 PM ET; Verge 4:41 PM ET): rolling out first "to a set of trusted cyber defenders" via the Fairwind Program, citing the US government's voluntary process for pre-release model access; priced intro $2/M input, $10/M output, cached input 95% off; positioning = long-horizon software engineering, legal/finance knowledge work, cybersecurity defense. Board day-one (Hassabis RTs): Vals Index #1 for the first time (68.9%), Text Arena #1 (1525), WebDev #8 (1679), blended ~$8/MTok Pareto claim. Kavukcuoglu's Sep 24 window ("much earlier than end of year") hit six days after the claim — ledger updated to hit. The gate-shaped release (defenders-first, in the week of the AISI Astra cyber rate and the Accord) is the frame; developer/enterprise/consumer access undated ("as soon as possible"). Watch: access widening, whether the pre-release voluntary process is Accord machinery operating, independent replication of the #1s once access opens.
- 2026-09-30 (spec layer, from the blog primary): 1M output tokens (industry-leading, up from 64K); claimed boards — DeepSWE v1.1 SOTA 77.9%, AutomationBench #1 (51.3%), LVBench SOTA 91.7%, CWE-bench v1 tie-first 68%; comparatives printed on the blog are vs Google's own 3.8 Flash Cyber, not vs Astra/Fable. Rollout order per blog: Fairwind cyber defenders + trusted testers now → paid API customers + Google AI Ultra subscribers first when it widens ("rolling out soon", no date); intro pricing $2/$10 converts to $4/$20 standard after the intro period — pricing published end-to-end is a commercial-availability tell. Notable: Argon ships to trusted defenders without cyber guardrails (full frontier cyber capability) — the defender-access ask Anthropic's GLM-5.3 post made is now Google's default launch posture; guardrail sections: misuse/CBRN refusal, Gray Swan IPI robustness lead, CoT+action misalignment monitoring, sealed sandboxes.
Background
- The thread: Gemini 4 gets a public timeline for the first time — Koray Kavukcuoglu (day-to-day DeepMind leader since Hassabis moved to Alphabet Chief Scientist, Aug 5; deepmind-exodus) said in his first media appearance as the division's leader (The Information, via The Verge ~5:04 AM ET) that Gemini 4 is in its refinement stage and targets launch "much earlier" than end of year. Frame in coverage: Google "dawdling behind rival developers on flagship AI releases."
- Watch for: actual launch date vs the ledgered claim (review-by 2026-12-31), positioning vs GPT-6 Astra and Claude Fable 5.1, and whether the leadership formalization brings a public counter to the exodus narrative (the watch in deepmind-exodus).