Google Launches Gemini Flash Models, Delays Pro Version Amid Coding Concerns

Google quietly shipped three new Gemini “Flash” models today — but the promised Pro upgrade is still nowhere to be seen, and investors noticed. What landed - Gemini 3.6 Flash — the headline release geared for fast, cost‑sensitive agent workloads. - Gemini 3.5 Flash‑Lite — a throughput‑optimized, ultra‑cheap model for high‑volume pipelines. - Gemini 3.5 Flash Cyber — a specialized model locked down for governments and vetted partners for vulnerability discovery and remediation (no broad public release). What didn’t: Gemini 3.5 Pro Google had teased a Pro variant of Gemini 3.5 at I/O in May with a promised follow‑up within a month. That didn’t happen. Bloomberg reports Google delayed 3.5 Pro after internal evaluations found it missed targets — especially on coding tasks. A late‑June retrain using new data didn’t fix the gaps, and the Pro SKU was held back. The market reacted: Alphabet shares fell about 4.4% on the news, roughly wiping out an estimated $200 billion in market value in a single session. (The last Pro‑tier release from Google was Gemini 3.1 Pro in February.) Why the Flash line matters Flash models are Google’s speed‑and‑cost-optimized offerings for agentic use cases — think automated bots that handle document workflows, data pipelines, web browsing and other semi‑autonomous tasks. Pro models, by contrast, are slower and more expensive but aimed at heavy reasoning and complex engineering work. Key technical and cost details - Gemini 3.6 Flash: - Produces ~17% fewer output tokens than 3.5 Flash (Artificial Analysis Index). - Pricing: $1.50 per million input tokens and $7.50 per million output tokens (output down from $9 on 3.5 Flash). - Benchmarks: DeepSWE v1.1 (long‑horizon software engineering) 49% (vs 37% for 3.5 Flash); MLE‑Bench 63.9% (vs 49.7%). OSWorld‑Verified (real‑screen task control) 83.0%, ahead of Claude Sonnet 5 (81.2%) and GPT‑5.6 Luna (72.6). - Weak spots remain: GPT‑5.6 Luna still leads on DeepSWE (67%) and Terminal‑Bench 2.1 (84.7%). Claude Sonnet 5 tops GDPval‑AA v2 (Elo) at 1607 vs 3.6 Flash’s 1421. - Gemini 3.5 Flash‑Lite: - Throughput: 350 output tokens/sec. - Pricing: $0.30 per million input and $2.50 per million output. - Designed for very high‑volume jobs (mass document processing, agentic search/indexing). - Coding performance improved versus older 3 Flash — Terminal‑Bench 2.1: 54% vs 31% — despite much lower cost. - Gemini 3.5 Flash Cyber: - Not publicly available. Restricted for government and vetted partner use for vulnerability finding/patching due to dual‑use risks. Hands‑on coding note The outlet’s quick coding test with 3.6 Flash was underwhelming: the model produced malformed HTML that didn’t render. A third‑party tool (Deepseek) identified 11 bugs and applied 8 fixes to turn the output into a playable game — suggesting the model’s high‑level reasoning was on track but micro‑detail and reliability were weak. In practice, that translates to cheaper but potentially lengthy “vibe coding” sessions — iterative fixes required to reach usable output. Why crypto builders should care - Lower per‑token costs and higher throughput can materially lower the bill for agentic crypto tools: automated trading bots, on‑chain monitors, bulk smart‑contract indexing and audit pipelines, and session compaction for agents that need to summarize long conversations or transaction histories. - But the coding reliability caveat matters: models that get the high‑level logic right but botch the details are risky for generating deployable smart contracts or automating critical ops without strong human review. - The Cyber model’s restricted availability underscores ongoing concerns about dual‑use AI tools for vulnerability discovery — a capability with obvious appeal (and risks) for security teams in crypto. What’s next Google says it has begun “its most ambitious pre‑training run yet” for Gemini 4 — full pre‑training is underway — and is hyping the next major iteration. Meanwhile, Gemini 3.6 Flash and 3.5 Flash‑Lite are available now in the Gemini app, Google AI Studio, and via API. Google promises Gemini 3.5 Pro will ship “as soon as it’s ready.” Bottom line Google doubled down on speed and cost with today’s Flash releases, which will be attractive for high‑volume crypto use cases. But the delay of a Pro‑class model — and the coding reliability issues reported — are a reminder that cheaper and faster doesn’t yet replace the need for careful engineering, auditing, and human oversight in production crypto systems.