Hands-on verdicts pour in: Gemini 4 Argon puts Google back in the frontier race

After Google DeepMind officially released Gemini 4 Argon, October 14 saw a wave of hands-on testing and reviews, with the consensus among most reviewers being 'Google is back.' Since Gemini 3 Pro last November (and the 3.1 Pro in February), Google's flagship had gone nearly a year without a major upgrade; Argon is seen as putting Google back in head-to-head competition with OpenAI and Anthropic.

Confirmed

• Ethan Mollick, after deep pre-release testing, said Gemini 4 is a massive leap over its predecessor and the first non-OpenAI/Anthropic model to pass the Fable/Mythos/Astra bar. Google's Logan Kilpatrick responded with thanks
• Every's team (Dan Shipper et al.), six people testing for a week across coding, writing, design, and knowledge work, found it surprisingly good: writing is the standout — DeepMind hired professional writers to polish the prose and remove the AI voice, and it's seen as a new writing partner; but it's constrained outside the Google ecosystem and feels more like a backup option
• Output capability: user screenshots and multiple sources indicate a maximum output length of 1,048,576 tokens; default output cap is 32,768 tokens (some say 65,536 default synchronous output), with longer outputs auto-split into multiple segments (requires long decode). Roughly 8x the 128k-output GPT/Claude frontier models
• Matt Wolfe posted a hands-on video covering the million-token long outputs, coding comparisons, and simultaneous updates to Antigravity and the Gemini App
• Google executive OfficialLoganK confirmed Gemini 4's real performance surprised many — it was expected to be a benchmark-chaser but actually competes head-on with frontier models, and said this is just the start of the 'Gemini 4 era'
• Developer hrishioa tested it: the model itself is good, works well paired with Gemini Flash, but tool call accuracy still falls short and it still misfires in demos
• Developer Kieran Klaassen's quick verdict after testing: 'Google is back'

Unconfirmed

• Wes Roth's testing found the model strong but behaving oddly; he suspects its reasoning may hide shortcuts and reproduced Google AI Studio's official demo — a personal observation without a definitive conclusion
• bindureddy's first results on the new LiveBench 2.0 benchmark: Claude Fable 5.1 leads with an 88.0 overall score (at $4.526 per-task success cost, the most expensive on the board); Gemini 4 Argon trails Fable and Astra but still leads open-source models; the benchmark claims to be hard to game and closer to real usage, but its credibility remains to be verified
• User imjustnewatai said they expected Gemini 4 to be a benchmark-chaser but after testing found it can match the first tier — no data provided, and model names mentioned in the post are not officially confirmed

Why it matters

• Gemini 4 Argon is Google's first major flagship upgrade in nearly a year, directly shaping the three-way frontier race (OpenAI, Anthropic, Google)
• If the million-token output capability holds, it far exceeds current mainstream flagship models (8x), opening up room for very long documents, large codebases, and similar use cases
• One relay of developer hammermt's week-long testing view: Google has long had internal 'AGI belief,' but low agent usage previously let Claude pull ahead; Argon 4 is Google's first agent model actually used internally, hinting at potentially stronger productization momentum ahead

2026-10-14 ~ 2026-10-14 · 29 related posts

Full story(18 episodes)→

• Episode 1: Google Unveils Gemini 4 Argón with Focus on Deep Reasoning(2026-10-07, 2 posts)
• Episode 2: Gemini 4 Argon's Delayed Rollout Fuels Speculation of Weekend Release(2026-10-07, 4 posts)
• Episode 3: Gemini 4 Codenamed Argon Leaks Across Google Platforms Ahead of Launch(2026-10-09, 10 posts)
• Episode 4: 'Gemini 4 Argon' Hype Sweeps Reddit, But It's All a Joke(2026-10-10, 3 posts)
• Episode 5: Google reportedly tests Gemini 4 "Carbon" internally, coding nears Opus 5.5(2026-10-10, 19 posts)
• Episode 6: Gemini 4 "Argon" rumors surge as model option surfaces in Antigravity IDE(2026-10-11, 6 posts)
• Episode 7: Gemini 4 Argon Leaks Across Google Surfaces as Launch Nears(2026-10-12, 20 posts)
• Episode 8: Google Reportedly Launches Gemini 4 Argon(2026-10-13, 4 posts)
• Episode 9: Google Launches Frontier Model Gemini 4 Argon Worldwide(2026-10-13, 189 posts)
• Episode 10: Hands-on verdicts pour in: Gemini 4 Argon puts Google back in the frontier race(2026-10-14, 29 posts)
• Episode 11: Gemini 4 Argon Hands-on: Million-Token Output in One Go(2026-10-14, 7 posts)
• Episode 12: Gemini 4 Argon Launches, Refactoring 6M Lines of Chrome Code in Weeks(2026-10-14, 9 posts)
• Episode 13: Gemini 4 Argon benchmarks leak: tops knowledge work and long-context, mixed in coding(2026-10-14, 6 posts)
• Episode 14: Google Publishes Gemini 4 Argon Safety Report; Situational Awareness Details Spark Discussion(2026-10-14, 12 posts)
• Episode 15: Gemini 4 Argon Generates Playable 3D Games in One Prompt(2026-10-14, 2 posts)
• Episode 16: Gemini 4 Argon Arrives on GitHub Copilot(2026-10-14, 2 posts)
• Episode 17: Leaked Gemini 4 Argon safety evals show top-tier cyber offense and stealth beyond human baselines(2026-10-14, 10 posts)
• Episode 18: Neel Nanda: Gemini 4 Argon markedly harder to monitor(2026-10-14, 2 posts)

• LiveBench 2.0: Gemini 4 Argon Trails Claude Fable and GPT-6 Astra but Beats Open Source — bindureddy · 2026-10-14
• [source] Every's week-long Gemini 4 review: a breakthrough if you live in Google's ecosystem — danshipper · 2026-10-14
• Vibe Check: Gemini 4 — Google is back — danshipper · 2026-10-14
• Vibe Check: Gemini 4 — Google is back — every · 2026-10-14
• Early vibe check on Gemini 4: 'Google is back', says developer — kieranklaassen · 2026-10-14
• Every's Vibe Check: Gemini 4 Argon puts Google back on the frontier — every · 2026-10-14
• After a week of testing, Argon signals Google is back in the frontier race — danshipper · 2026-10-14
• Every's vibe check: Gemini 4 Argon puts Google back on the frontier — every · 2026-10-14
• Gemini 4 Argon hands-on: million-token output, coding tests, and how Google stacks up — Matt Wolfe · 2026-10-14
• [source] Mollick: Gemini 4 is first non-OpenAI/Anthropic model to pass the Fable/Mythos/Astra bar — emollick · 2026-10-14
• Ethan Mollick: Gemini 4 is first non-OpenAI/Anthropic model to clear Fable/Mythos/Astra bar — zacharynado · 2026-10-14
• Mollick: Gemini 4 a Big Jump, First Non-OpenAI/Anthropic Model to Pass Key Bar — OfficialLoganK · 2026-10-14
• Gemini 4 praised as a big leap — but its generated UI is hilariously over the top — BraydonDymm · 2026-10-14
• Hands-on with Google's Argon: not frontier at coding but a solid, well-designed model — zacharynado · 2026-10-14
• A week with Argon 4: Google is back, its first frontier agent model with real internal usage — every · 2026-10-14
• Gemini 4 Argon supports up to 1M token outputs, 8x any frontier GPT or Claude model — daniel_mac8 · 2026-10-14
• [source] Gemini 4 Argon ships with up to 1M token max output, auto-continued past 32K — BeefHotSweetDipped · 2026-10-14
• Early tester says Gemini 4 competes with the frontier, defying benchmaxx expectations — imjustnewatai · 2026-10-14
• Hands-on with Gemini 4 Argon: strong on professional benchmarks, weak on agents and coding UX — 数字生命卡兹克 · 2026-10-14
• Gemini 4 hands-on: good model, but tool calls still fail validation in real workflows — hrishioa · 2026-10-14
• Gemini 4 Argon hands-on: powerful model with strange habits and vanishing safety shortcuts — Wes Roth · 2026-10-14
• After billions of tokens, developer finds Gemini 4 roughly on par with Opus 4.8 — hrishioa · 2026-10-14
• Google Exec Confirms Gemini 4 Competes at the Frontier, Calls It Just the Start — OfficialLoganK · 2026-10-14
• Billions of tokens later: Gemini 4 ≈ Opus 4.8, worse at writing and agentic loops — hrishioa · 2026-10-14
• Gemini 4 Argon hands-on: unified 1M context window, 77.9% on DeepSWE, prices double after Nov 30 — 卡尔的AI沃茨 · 2026-10-14
• Gemini 4 reportedly aces ThreeJS Boeing 747 benchmark that Gemini 3 failed — LysandreJik · 2026-10-14

3 near-duplicate retellings: danshipper · every · daniel_mac8