FULL STORY

Gemini 3.8 Flash: From Leaks to Launch and Benchmarks

After leaks foreshadowed its release, Google launched Gemini 3.8 Flash and a Cyber variant on Sept 2; early tests were split between overfitting claims and strong multimodal results, amid complaints about rapid iteration.

2026-09-02 ~ 2026-09-04 · 6 episodes · 164 posts

Episode 1 · Gemini 3.8 Flash Surfaces in Shadow Rollout as Leaked Benchmarks and Pricing Circulate (2026-09-02, 45 posts)

According to the Wall Street Journal, Google plans to release Gemini 3.8 Flash (codename "Skimaki") as early as Wednesday, focused on coding, with insiders claiming major advances in agentic coding; the company has also abandoned the 3.5 Pro checkpoint. As of the posts, Google had not issued an official announcement, but the model has surfaced via shadow rollout in consumer apps, deployment environments, and the Google Cloud Console, with leaked benchmarks circulating widely and Artificial Analysis publishing third-party results.

Confirmed

  • WSJ reports Gemini 3.8 Flash brings major agentic coding advances, launching as early as Wednesday (relayed by @kimmonismus, @Hesamation, @xiaohu, etc.)
  • In A/B tests on Google's internal coding tool Jetski, Google engineers preferred Gemini 3.8 Flash over Anthropic's Opus model (@kimmonismus citing WSJ)
  • Multiple users observed the model in consumer apps (@markk, @IamNotGorbachev WebUI screenshots); @Able-Line2683's Free vs Pro/Ultra account comparison confirms it is being pushed to paid accounts only — a shadow rollout
  • @testingcatalog found Gemini 3.8 Flash on a Google Cloud Console quota page, a routine pre-launch step
  • @clmt says the model is released with major gains over 3.7 Flash in agentic coding and general knowledge work, and is available to all users on Antigravity (no Google announcement seen at the time)
  • WSJ says Google has officially dropped the 3.5 Pro checkpoint due to limited gains over the Flash line

Unconfirmed

  • @teortaxesTex relays leaked pricing matching 3.7 Flash: $0.75/Mtok input, $3.75/Mtok output, $0.075/Mtok cached input, with launch in 7 hours — unverified
  • @kimmonismus's leaked benchmarks (self-described as unverified) show it beating Opus 5 and "Sol 5.6" on Terminal Bench 2.1 and HLE; @flavioAd expressed amazement and shared analysis; @Able-Line2683's repeated benchmark screenshots cannot be verified
  • @ExpensiveSyrup6529 shared Artificial Analysis third-party benchmark results (image gallery), figures not officially confirmed
  • @OfficialLoganK posted a 73.7% DeepSWE 1.1 score with eval link; @vedantmisra and @haider1 relayed claims of topping DeepSWE v1.1 at 74% (vs 65% for 3.7 Flash), at the cost of high output token usage; @ChrisGPT cites 89.4% on Terminal-Bench 2.1 while noting that leaderboard is outdated (now at version 4). All third-party, unconfirmed by Google
  • Reports relayed by @Hesamation speculate Google may skip 3.5 Pro and go straight to Gemini 4 after training completes — media speculation
  • @markk predicted a same-day launch; Andrew Curran (via @TheZachMueller) said "within hours" (the post also mentions Commerce Secretary Lutnick confirming semiconductor tariffs to CNBC) — take WSJ's original report as authoritative

Why it matters

  • If a Flash model already beats Anthropic's flagship in internal tests, it reshapes the mid-size model landscape; the leaked Terminal Bench 2.1, HLE, and DeepSWE 73.7%–74% scores would further support this if genuine, though high token usage could raise real-world costs
  • Dropping 3.5 Pro suggests a two-tier "lightweight Flash + Gemini 4 flagship" roadmap, skipping intermediate versions
  • Rumored pricing identical to 3.7 Flash would mean stronger coding at the same price, pressuring competitors; full availability on Antigravity signals Google is making agentic coding a core distribution scenario

25 more related posts →

Episode 2 · Google Launches Gemini 3.8 Flash and Cyber Variant (2026-09-02, 110 posts)

Google officially launched Gemini 3.8 Flash and the security-focused 3.8 Flash Cyber on September 2, announced by CEO Sundar Pichai, Google DeepMind, Google AI, GeminiApp and executive Logan Kilpatrick. This is the third Flash update in six weeks and the fourth Flash model in four months. 3.8 Flash is live in the Gemini app, Google AI Studio, the API and Antigravity, available to Pro and Ultra subscribers, and positioned as a workhorse model for long-horizon coding and autonomous agents.

Confirmed

  • 3.8 Flash: billed as the smartest Flash yet, with significant gains over 3.7 Flash in software engineering, agentic tasks and multi-step reasoning; DeepSWE 1.1 score of 71% approaches Opus 5 at far lower cost (user-side assessment). Artificial Analysis measured a high-reasoning intelligence index of 59 (vs 56 for 3.7 Flash) and $0.58 per task, placing it on the price-performance Pareto frontier.
  • Pricing and speed: $0.75 input / $0.75 output pricing unchanged — input $0.75 and output $3.75, matching 3.7 Flash's launch price. Per Reddit user NewVeterinarian5384's compilation of the official blog, output reaches 305 tokens/sec, nearly double second-place Muse Spark 1.2 (154 t/s) and far ahead of GPT-5.6.
  • Capabilities: Google AI highlights complex agents and multi-step tasks spanning coding, video understanding and everyday assistance; effort controls retained, with a demo autonomously building a 3D game and self-correcting.
  • 3.8 Flash Cyber: announced by Pichai, DeepMind and Google AI as Google's strongest cybersecurity model, featuring expert-level vulnerability detection, autonomous patching and large-scale patch generation; 2.6x improvement in Chrome patch correctness over the 3.5 generation and Pareto-frontier results on CWE-Bench, while keeping Flash-tier speed and pricing.

Why it matters

  • Same-price performance and speed gains squeeze competitors' price-performance room: YouTube channel Prompt Engineering calls it Google's strongest agentic coding model, and blogger SabooShubham argues the rapid cadence particularly benefits AI agent workflows.
  • Cyber brings frontier vulnerability discovery and autonomous patching to defensive security teams at Flash-level pricing, potentially reshaping security operations automation.

90 more related posts →

Episode 3 · HF Engineer Praises Gemini 3.8 as Well-Rounded Agentic Model (2026-09-03, 3 posts)

Hugging Face engineer Omar Sanseviero shared that he has been building with Gemini 3.8 for days, calling it a well-balanced agentic model with strong performance across finance, legal, and multimodal tasks—though Google has not officially confirmed the model.

Episode 4 · Gemini Flash ships 3.6-3.8 in six weeks, testers complain of churn (2026-09-03, 2 posts)

Google has rolled out Gemini Flash 3.6, 3.7 and 3.8 within six weeks. Commentators complain the rapid cadence yields limited gains, with new checkpoints landing before evals can even finish.

Episode 5 · Developer claims Gemini 3.8 Flash is overfit to public benchmarks (2026-09-03, 2 posts)

Developer Bindu Reddy claims Gemini 3.8 Flash is overfit to public benchmarks, as it underperforms 3.7 Flash on her private benchmarks, with regression in data analysis and hidden problem sets.

Episode 6 · Gemini 3.8 Flash Tops Image Reasoning Benchmarks, 30% Faster (2026-09-04, 2 posts)

Roboflow's Skalski published a deep review of Gemini 3.8 Flash, calling it possibly his favorite Gemini yet: it ranked first in image reasoning and data extraction, second in object detection, and ran about 30% faster than Gemini 3.7 Flash.